<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[The Human Framework]]></title><description><![CDATA[Comprendre l’intelligence, humaine, artificielle, et tout ce qu’il y a entre les deux.]]></description><link>https://humanframework.substack.com</link><image><url>https://substackcdn.com/image/fetch/$s_!_I1w!,w_256,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc6d2ff2b-3575-4130-9b9b-04e84c5dc021_778x778.png</url><title>The Human Framework</title><link>https://humanframework.substack.com</link></image><generator>Substack</generator><lastBuildDate>Thu, 05 Nov 2026 02:19:18 GMT</lastBuildDate><atom:link href="iframe.php?url=https%3A%2F%2Fhumanframework.substack.com%2Ffeed" rel="self" type="application/rss+xml"/><copyright><![CDATA[Benjamin Tardy]]></copyright><language><![CDATA[fr]]></language><webMaster><![CDATA[benjamintardy@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[benjamintardy@substack.com]]></itunes:email><itunes:name><![CDATA[Benjamin]]></itunes:name></itunes:owner><itunes:author><![CDATA[Benjamin]]></itunes:author><googleplay:owner><![CDATA[benjamintardy@substack.com]]></googleplay:owner><googleplay:email><![CDATA[benjamintardy@substack.com]]></googleplay:email><googleplay:author><![CDATA[Benjamin]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[Gemini 4 Carbon prend la tête des benchmarks]]></title><description><![CDATA[Le brief IA #89]]></description><link>https://humanframework.substack.com/p/gemini-4-carbon-prend-la-tete-mais</link><guid isPermaLink="false">https://humanframework.substack.com/p/gemini-4-carbon-prend-la-tete-mais</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Wed, 04 Nov 2026 07:30:41 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%210uCl%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%217Gbm%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!7Gbm!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!7Gbm!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!7Gbm!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!7Gbm!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%217Gbm%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp" width="728" height="409.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;normal&quot;,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Le Brief&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Le Brief" title="Le Brief" srcset="https://substackcdn.com/image/fetch/$s_!7Gbm!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!7Gbm!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!7Gbm!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!7Gbm!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7deec0d5-9122-4e5e-a383-f4c16a4b4440_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>Google</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fintroducing-gemini-4-carbon%2F">Gemini 4 Carbon</a></strong></p><p>Google n'aura pas attendu longtemps. Deux semaines apr&#232;s Gemini 4 Argon, voici Gemini 4 Carbon, son mod&#232;le le plus puissant, pens&#233; pour la cybers&#233;curit&#233; et les longues t&#226;ches de programmation. Artificial Analysis le classe premier de la nouvelle version de son indice avec 59. Ars Technica note qu'il ressemble beaucoup &#224; Argon et se demande s'il ne s'agit pas d'une simple retouche de fin d'entra&#238;nement. Le mod&#232;le n&#8217;est pas encore disponible.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%210uCl%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!0uCl!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 424w, https://substackcdn.com/image/fetch/$s_!0uCl!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 848w, https://substackcdn.com/image/fetch/$s_!0uCl!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 1272w, https://substackcdn.com/image/fetch/$s_!0uCl!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%210uCl%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin" width="648" height="544.86" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1009,&quot;width&quot;:1200,&quot;resizeWidth&quot;:648,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Gemini 4 Carbon benchmark table&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Gemini 4 Carbon benchmark table" title="Gemini 4 Carbon benchmark table" srcset="https://substackcdn.com/image/fetch/$s_!0uCl!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 424w, https://substackcdn.com/image/fetch/$s_!0uCl!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 848w, https://substackcdn.com/image/fetch/$s_!0uCl!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 1272w, https://substackcdn.com/image/fetch/$s_!0uCl!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9d1594b-862a-4608-9307-5ab7ed543f9b_1200x1009.bin 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong><a href="iframe.php?url=https%3A%2F%2Fdeepmind.google%2Fblog%2Fgenie-4-a-new-generation-of-world-simulation%2F">Genie 4</a></strong></p><p>L'an dernier, Genie 3 avait fait chuter les actions du jeu vid&#233;o. Genie 4 g&#233;n&#232;re lui aussi en temps r&#233;el un monde qu'on parcourt de mani&#232;re informatique et qu'on modifie d'une phrase, pour changer la m&#233;t&#233;o ou ajouter un personnage. Cette fois, le son, la musique, les bruits d'ambiance et les dialogues sont cr&#233;es en m&#234;me temps que l'image. DeepMind annonce aussi une physique plus juste, des chocs &#224; l'a&#233;rodynamique, et des humains plus naturels. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F10%2F28%2Fchromes-ai-powered-autofill-can-now-dig-through-your-gmail-to-fill-out-forms%2F">Chrome va chercher dans Gmail de quoi remplir vos formulaires</a> (aux Etats-Unis)</strong></p><p>Le remplissage automatique de Chrome va chercher dans Gmail les num&#233;ros qu'il fallait jusqu'ici y retrouver &#224; la main, ceux d'un vol, d'un colis, d'une commande, voire d'une carte bancaire. C'est Gemini qui s'en charge, via Personal Intelligence, l'option qui relie les applis Google &#224; l'assistant. Jusqu'ici, Chrome puisait dans le compte Google et dans des services comme Google Wallet, rappelle The Verge. La fonction s'active ou se coupe dans les r&#233;glages, &#224; condition d'avoir Personal Intelligence, et Google assure que les donn&#233;es sensibles restent chiffr&#233;es. Elle arrive aux &#201;tats-Unis sur ordinateur, Android et iOS. </p><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-gpt-6-1-terra%2F">GPT-6.1 Terra</a></strong></p><p>GPT-6.1 Terra, c'est un peu le Sonnet d'OpenAI, un mod&#232;le du milieu de gamme. En API, il co&#251;te 1 $ par million de tokens en entr&#233;e et 5 $ en sortie, moiti&#233; moins que GPT-6.1 Sol et que son pr&#233;d&#233;cesseur GPT-5.6 Terra. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MePP%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!MePP!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 424w, https://substackcdn.com/image/fetch/$s_!MePP!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 848w, https://substackcdn.com/image/fetch/$s_!MePP!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 1272w, https://substackcdn.com/image/fetch/$s_!MePP!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MePP%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png" width="658" height="585.2403846153846" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1295,&quot;width&quot;:1456,&quot;resizeWidth&quot;:658,&quot;bytes&quot;:137383,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/222615259?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!MePP!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 424w, https://substackcdn.com/image/fetch/$s_!MePP!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 848w, https://substackcdn.com/image/fetch/$s_!MePP!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 1272w, https://substackcdn.com/image/fetch/$s_!MePP!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F93731db8-af23-49e2-b56c-13a822f0032c_1554x1382.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>GitHub</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fgithub.blog%2Fchangelog%2F2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion%2F">Avec Copilot Fusion, GitHub choisit le mod&#232;le &#224; la place du d&#233;veloppeur</a></strong></p><p>Un mod&#232;le d'Anthropic ou un mod&#232;le d'OpenAI, il fallait en g&#233;n&#233;ral trancher soi-m&#234;me pour chaque t&#226;che de code. Copilot Fusion, que GitHub rend disponible pour tous lors de sa conf&#233;rence Universe 2026, fait ce choix &#224; chaque demande et d&#233;cide aussi comment faire travailler les mod&#232;les ensemble. Sur son indice des agents de code, Artificial Analysis le place &#224; 2 points de Claude Code avec Opus 5.5, pour un co&#251;t par t&#226;che divis&#233; par deux. Satya Nadella annon&#231;ait &#171; jusqu'&#224; 77 % &#187; d'&#233;conomie, mais deux points d'&#233;cart pour moiti&#233; prix, c'est d&#233;j&#224; une belle affaire. </p><h3>NVIDIA</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fblog%2Fnvidia%2Fnemotron-transcribe-granary">NVIDIA r&#233;unit transcription, traduction et identification des locuteurs dans un seul mod&#232;le ouvert</a></strong></p><p>Chez NVIDIA, la transcription et l'identification des locuteurs reposaient jusqu'ici sur des mod&#232;les distincts. Nemotron 3.5 Transcribe r&#233;unit tout dans un seul mod&#232;le &#224; poids ouverts. Il transcrit 40 langues, en traduit 31 vers ou depuis l'anglais, distingue jusqu'&#224; huit voix et r&#233;pond &#224; des questions sur ce qui s'est dit, sur des enregistrements seulement, pas en direct. </p><h3>Mistral AI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Fvoxtral-3%2F">Voxtral 3</a></strong></p><p>Mistral travaille sur la chaine d&#8217;agents vocaux qui r&#233;pondent au t&#233;l&#233;phone. Voxtral 3 r&#233;unit de nouveaux mod&#232;les de transcription et de synth&#232;se vocale, et un SDK construit sur LiveKit o&#249; l'&#233;coute, le raisonnement d'un mod&#232;le de langage, les appels d'outils et la r&#233;ponse vocale tournent sans s'attendre. D'apr&#232;s Mistral, chaque tour de parole prend moins d'une seconde, et l'agent sait se taire quand on lui coupe la parole. Pendant l'appel, il peut v&#233;rifier l'identit&#233; de son interlocuteur ou r&#233;server un rendez-vous. L'ensemble s'utilise par l'API h&#233;berg&#233;e en Europe ou s'installe sur vos propres serveurs. Avec Mistral Small 4 pour raisonner, l'agent co&#251;te 0,04 dollar la minute. Les comparaisons avec Whisper ou ElevenLabs viennent surtout des tests de Mistral lui-m&#234;me, mais Artificial Analysis classe d&#233;j&#224; Voxtral Small Transcribe 3 troisi&#232;me de son classement de transcription.</p><div id="youtube2-xFhyfPC-VzI" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;xFhyfPC-VzI&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FxFhyfPC-VzI%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>Apple</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Fapple%2F2026%2F10%2Fapple-releases-ios-and-macos-27-2-with-new-siri-ai-features-new-music-logo%2F">Siri AI parle fran&#231;ais avec iOS 27.2, sauf en France</a></strong></p><p>Avec les mises &#224; jour 27.2 de ses syst&#232;mes, Apple ouvre Siri AI au fran&#231;ais, au japonais, au cor&#233;en, au portugais et &#224; l'espagnol. L'assistant reste en revanche indisponible dans l'Union europ&#233;enne, France comprise &#129760;&#8203;</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21fzKp%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!fzKp!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!fzKp!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!fzKp!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!fzKp!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21fzKp%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp" width="728" height="409.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;normal&quot;,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;En Vrac&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="En Vrac" title="En Vrac" srcset="https://substackcdn.com/image/fetch/$s_!fzKp!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!fzKp!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!fzKp!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!fzKp!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9694dfe3-62e3-492c-8460-c28af6d20654_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>&#128274; <a href="iframe.php?url=https%3A%2F%2Fwww.sverigesradio.se%2Fartikel%2Fchat-gpt-och-claude-forbjuder-regeringskansliet">Le gouvernement su&#233;dois bloque ChatGPT et Claude pour ses agents et les renvoie vers des outils valid&#233;s comme Copilot</a></p><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Ftech-policy%2F2026%2F10%2Fno-human-at-the-helm-delaware-plan-would-let-ai-agents-run-entire-businesses%2F">Un projet de loi du Delaware laisserait un agent IA diriger seul une entreprise, en couvrant largement son propri&#233;taire</a></p><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F10%2F27%2Fparents-sue-openai-alleging-gpt-4o-coached-their-daughter-off-seizure-medication%2F">Des parents attaquent OpenAI, accusant GPT-4o d&#8217;avoir aid&#233; leur fille &#233;pileptique, d&#233;c&#233;d&#233;e en f&#233;vrier, &#224; cesser son traitement en cachette, selon le New York Times</a></p><p>&#127968; <a href="iframe.php?url=https%3A%2F%2Fwww.theverge.com%2Fai-artificial-intelligence%2F1020596%2Fxai-mississippi-neighborhood-data-center-lawsuit">xAI offre de racheter les maisons voisines de son data center du Mississippi &#224; qui renonce &#224; la plainte des riverains, selon The Guardian</a></p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21mox_%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!mox_!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!mox_!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!mox_!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!mox_!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21mox_%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp" width="728" height="409.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;normal&quot;,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Deep Dive IA &amp; ML&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Deep Dive IA &amp; ML" title="Deep Dive IA &amp; ML" srcset="https://substackcdn.com/image/fetch/$s_!mox_!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!mox_!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!mox_!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!mox_!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2e2938-bafc-434a-8598-a3bea69ada0e_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fengineering.fb.com%2F2026%2F10%2F28%2Fml-applications%2Fmeta-andromeda-next-gen-ads-retrieval-gpu-co-design%2F">Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design</a></h3><p>Avant de vous montrer une publicit&#233;, Meta doit trier en quelques millisecondes un stock immense d'annonces pour n'en garder qu'une poign&#233;e (de bonnes candidates). C'est le travail d'Andromeda, son moteur de pr&#233;s&#233;lection. Les douze auteurs de ce billet expliquent comment sa nouvelle version parvient &#224; noter chaque annonce &#233;ligible, au rythme de plus d'un milliard d'annonces par seconde. Il le fallait, car les outils d'IA g&#233;n&#233;rative ont fait arriver des centaines de millions de nouvelles annonces.</p><p>L'&#233;quipe s'appuie sur un m&#233;lange d'experts, c'est-&#224;-dire plusieurs r&#233;seaux sp&#233;cialis&#233;s chacun dans une facette des centres d'int&#233;r&#234;t de l'utilisateur, dont seuls les plus pertinents s'activent pour une annonce donn&#233;e. D'apr&#232;s Meta, ce mod&#232;le demande environ 300 fois moins de calcul que la version de 2024. </p><h3><a href="iframe.php?url=https%3A%2F%2Fshopify.engineering%2Fcatalog-clustering-part-ii">Cross-store product clustering across millions of stores</a></h3><p>Une m&#234;me coque de t&#233;l&#233;phone peut se vendre dans plusieurs boutiques Shopify, chacune avec son titre, ses photos et sa fa&#231;on de ranger tailles et couleurs. Pour en donner une vue unique aux acheteurs, aux d&#233;veloppeurs et aux agents IA, Shopify doit reconna&#238;tre le m&#234;me produit parmi des milliards de fiches, et Ilia Kuchumov d&#233;taille comment l'&#233;quipe s'y prend. Les embeddings font un premier tri en trouvant &#224; chaque fiche ses plus proches voisines. Il reste des milliers de milliards de paires &#224; trancher, que Shopify confie &#224; trois mod&#232;les, du moins cher au plus co&#251;teux.</p><ol><li><p>Un mod&#232;le de gradient boosting, qui tourne sur CPU, note toutes les paires &#224; partir d'indices simples comme la proximit&#233; des titres, des marques ou des prix.</p></li><li><p>Un encodeur de texte multilingue lit les deux fiches ensemble, et voit ainsi qu'une coque &#171; pour iPhone 15 &#187; n'est pas une coque &#171; pour iPhone 15 Pro &#187;.</p></li><li><p>Un mod&#232;le qui regarde aussi les photos ne re&#231;oit que des paires choisies, pour rep&#233;rer un connecteur ou une d&#233;coupe qui diff&#232;re.</p></li></ol><p>Aucun grand LLM ne juge toutes les paires, ce serait bien trop cher.</p><h3><a href="iframe.php?url=https%3A%2F%2Fwww.dbreunig.com%2F2026%2F10%2F14%2Fcontext-inertia.html">Context Inertia: When Agents Won't Let Go</a></h3><p>Drew Breunig recense depuis juin 2025 les fa&#231;ons dont un long contexte fait d&#233;railler un mod&#232;le, et il ajoute un nouveau concept &#224; sa liste, la Context Inertia. Le mod&#232;le s'engage sur une hypoth&#232;se ou un objectif et s'y accroche alors que tout lui dit de s'arr&#234;ter ou de changer de piste. Chez Anthropic, un agent Qwen a ainsi pass&#233; des heures &#224; fouiller les versions de paquets, alors que le vrai bug venait d'un endpoint impl&#233;ment&#233; de fa&#231;on incompatible avec l'API du correcteur automatique charg&#233; de noter son travail. Si le d&#233;faut appara&#238;t maintenant, explique Breunig, c'est que les agents tournent enfin assez longtemps pour s'y heurter, pendant des jours et sur des milliards de tokens.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21WsFL%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!WsFL!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!WsFL!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!WsFL!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!WsFL!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21WsFL%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp" width="728" height="409.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;normal&quot;,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;En Production&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="En Production" title="En Production" srcset="https://substackcdn.com/image/fetch/$s_!WsFL!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!WsFL!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!WsFL!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!WsFL!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f60e242-0bb7-4dc8-9d72-d03fe73fdea7_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fwww.infoq.com%2Fnews%2F2026%2F10%2Fairbnb-chronon-recommendation%2F">Airbnb Cuts Recommendation Staleness from 2 Days to Seconds with Chronon</a></h3><p>Jusqu'ici, quand un voyageur cherchait un logement sur Airbnb, les recommandations pouvaient ignorer tout ce qu'il avait fait sur le site depuis pr&#232;s de deux jours. Airbnb a ramen&#233; ce retard &#224; moins d'une minute en touchant &#224; peine &#224; son moteur de classement, rapporte InfoQ. Le syst&#232;me r&#233;sume l'historique de chaque voyageur, ses r&#233;servations, ses avis et sa navigation, en un embedding, une sorte de profil chiffr&#233; calcul&#233; par un mod&#232;le Transformer. Ce calcul tournait une fois par jour, en batch, et le profil attendait dans une base rapide o&#249; le moteur venait le lire &#224; chaque recherche. Le mod&#232;le restait ainsi &#224; l'&#233;cart du service en ligne mais tout ce qui se passait apr&#232;s le batch devait attendre le suivant.</p><p>Pour s&#8217;en sortir, Airbnb a ajout&#233; deux briques &#224; Chronon :</p><ol><li><p>Push Mode pr&#233;vient la suite du pipeline d&#232;s qu'une nouvelle donn&#233;e sur le voyageur est enregistr&#233;e, au lieu d'attendre une t&#226;che planifi&#233;e.</p></li><li><p>NRT Model Transform fait tourner un mod&#232;le d&#233;j&#224; d&#233;ploy&#233; directement dans ce flux de streaming.</p></li></ol><p>Le Transformer reste le m&#234;me qu'en batch, et le nouveau profil s'&#233;crit dans la m&#234;me base. Le classement lit donc simplement une version plus fra&#238;che. Le mod&#232;le et le service en ligne restent pour l'essentiel inchang&#233;s, et seul change le moment o&#249; le calcul se d&#233;clenche. D'apr&#232;s le test A/B d'Airbnb, les r&#233;servations non annul&#233;es progressent d'environ 0,33 %, et les deux briques sont maintenant dans le d&#233;p&#244;t public de Chronon.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21mfeQ%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!mfeQ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!mfeQ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!mfeQ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!mfeQ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21mfeQ%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp" width="728" height="409.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;normal&quot;,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;&#192; Bookmarker&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="&#192; Bookmarker" title="&#192; Bookmarker" srcset="https://substackcdn.com/image/fetch/$s_!mfeQ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!mfeQ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!mfeQ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!mfeQ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F63cb74f8-8c04-4f88-9f79-31de0f6c885c_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fmagazine.sebastianraschka.com%2Fp%2Fbuilding-reasoning-models">Building Reasoning Models from the Ground Up</a></h3><p>Sebastian Raschka publie huit vid&#233;os, plus de 11 heures de code en Python et PyTorch, pour construire soi-m&#234;me un mod&#232;le de raisonnement &#224; partir d&#8217;un petit Qwen3 de 0,6 milliard de param&#232;tres. </p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fcloudflare%2Fsecurity-audit-skill">cloudflare/security-audit-skill</a></h3><p>Cloudflare publie en open source le skill qui a servi de point de d&#233;part &#224; sa propre chasse aux failles. Il transforme un agent de code en auditeur de s&#233;curit&#233;. On lui demande d&#8217;auditer un d&#233;p&#244;t, et il d&#233;roule six &#233;tapes jusqu&#8217;au rapport. Sa r&#232;gle d&#8217;or veut que chaque faille trouv&#233;e passe &#224; un autre agent, charg&#233; de la r&#233;futer. D&#8217;apr&#232;s Cloudflare, un seul passage trouve environ la moiti&#233; des failles alors mieux vaut le relancer.</p><h3><a href="iframe.php?url=https%3A%2F%2Fwww.brangus.io%2Fsurvival-llm">Survival LLM</a></h3><p>Survival LLM est un assistant gratuit et open source qui tourne sur votre ordinateur, Mac ou Windows, sans Internet une fois install&#233;. Il r&#233;pond &#224; partir d&#8217;une biblioth&#232;que stock&#233;e en local (guides de secours officiels, r&#233;f&#233;rences m&#233;dicales, manuels de survie, environ 930 000 articles de Wikip&#233;dia), et chaque r&#233;ponse ouvre le passage exact, surlign&#233;. La qualit&#233; d&#233;pend de la machine. Sur 22 questions de premiers secours, l&#8217;&#233;quipe compte 21 bonnes r&#233;p</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21DThH%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!DThH!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!DThH!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!DThH!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!DThH!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21DThH%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp" width="728" height="409.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:false,&quot;imageSize&quot;:&quot;normal&quot;,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:728,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Le Papier de la Semaine&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Le Papier de la Semaine" title="Le Papier de la Semaine" srcset="https://substackcdn.com/image/fetch/$s_!DThH!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!DThH!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!DThH!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!DThH!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0680bfae-eff6-48fa-8bd1-77fe1b4fd691_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2610.32285">Trial by Trace: Benchmarking AI Agent Harnesses for Reliability and Failure Attribution</a></h3><blockquote><p>LLM agents are increasingly used to solve real-world problems, and their effectiveness depends on both the underlying models and the harnesses that support their operation. However, most existing benchmarks focus on model capabilities and overall agent performance, with limited attention to the evaluation of agent harnesses. To address this gap, we introduce TraceTrial-Bench, which organizes harness reliability into seven components covering context, execution, governance, lifecycle, observability, tools, and verification. We construct 95 expert-level, mechanism-oriented tasks with executable or state-based ground truth to support semantically equivalent comparisons while preserving the native behavior of each harness. We design a Trajectory Tracing (TT) module to capture agent execution traces and introduce a two-layer evaluation framework that separately assesses agent performance and harness reliability. In addition, we design a Multi-source Evidence Acquisition (MEA) module to provide evidence from the agent context, harness, and execution environment for LLM-based evaluation, and develop a Failure Attribution (FA) module that uses explicit mechanism chains to support attribution at the run level. Across 4 harnesses, 9 models, and 95 tasks, our evaluation of 3,420 runs yields macro-averaged harness scores ranging from 59.8% to 72.0%, while 31.1% of runs achieve task success despite failing at least one harness criterion. TraceTrial-Bench provides an auditable and reproducible foundation for comparing agent harnesses, diagnosing their failures, and studying how harness design interacts with model behavior.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Meta s'aligne sur les perfs d'Opus 5.5, Anthropic réalise 16,8 milliards en un trimestre]]></title><description><![CDATA[Le brief IA #88]]></description><link>https://humanframework.substack.com/p/meta-saligne-sur-les-perfs-dopus</link><guid isPermaLink="false">https://humanframework.substack.com/p/meta-saligne-sur-les-perfs-dopus</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 23 Oct 2026 06:31:07 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21SA7S%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!RXIb!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!RXIb!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!RXIb!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!RXIb!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21RXIb%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!RXIb!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!RXIb!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!RXIb!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!RXIb!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8bfd7330-eda5-4ca7-bdde-82b1e915b94c_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Meta</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fblog%2Fmeta-muse-spark-open">Meta ouvre Muse Spark Open</a></strong><br>Meta publie les poids de Muse Spark Open, que le groupe pr&#233;sente comme le mod&#232;le ouvert am&#233;ricain le plus performant &#224; ce jour. Meta l&#8217;a entra&#238;n&#233; pour le code et les agents, mais il lit aussi les images, le son et la vid&#233;o, avec un million de tokens de contexte. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Fresearch.meta.ai%2Fblog%2Fmuse-wonder">Avec Muse Wonder, Meta s&#8217;approche d&#8217;Opus 5.5 et s&#8217;aligne sur son prix</a></strong><br>Contrairement &#224; Muse Spark Open, Muse Wonder ne se t&#233;l&#233;charge pas. C&#8217;est pourtant avec ce mod&#232;le ferm&#233; que Meta affirme &#234;tre de retour dans la course aux mod&#232;les de pointe. D&#8217;apr&#232;s Meta, Muse Wonder Full obtient 56 sur l&#8217;indice d&#8217;Artificial Analysis, contre 58 pour Claude Opus 5.5 et 53 pour GPT-6 Astra, alors que son pr&#233;c&#233;dent mod&#232;le de pointe plafonnait &#224; 48. La version standard, not&#233;e 52, s&#8217;utilise par API au tarif exact d&#8217;Opus 5.5, 4 dollars le million de tokens en entr&#233;e et 20 en sortie. Wonder Full est r&#233;serv&#233; aux abonn&#233;s de Muse Code. En ao&#251;t, Meta misait sur des prix tr&#232;s bas pour lancer Muse Code. S&#8217;aligner aujourd&#8217;hui sur Anthropic dit assez o&#249; le groupe pense se situer.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21SA7S%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!SA7S!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 424w, https://substackcdn.com/image/fetch/$s_!SA7S!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 848w, https://substackcdn.com/image/fetch/$s_!SA7S!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 1272w, https://substackcdn.com/image/fetch/$s_!SA7S!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21SA7S%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp" width="584" height="380.24175824175825" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:948,&quot;width&quot;:1456,&quot;resizeWidth&quot;:584,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Artificial Analysis Intelligence Index bar chart comparing Muse Wonder Full, Muse Wonder, Muse Spark models, and other leading models.&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Artificial Analysis Intelligence Index bar chart comparing Muse Wonder Full, Muse Wonder, Muse Spark models, and other leading models." title="Artificial Analysis Intelligence Index bar chart comparing Muse Wonder Full, Muse Wonder, Muse Spark models, and other leading models." srcset="https://substackcdn.com/image/fetch/$s_!SA7S!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 424w, https://substackcdn.com/image/fetch/$s_!SA7S!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 848w, https://substackcdn.com/image/fetch/$s_!SA7S!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 1272w, https://substackcdn.com/image/fetch/$s_!SA7S!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2c9047-a7d8-47b9-b88f-9906c1d9c1ee_5760x3750.webp 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>Reflection</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fblog%2Freflection-beam">Reflection sort Beam</a></strong><br>Encore un gros mod&#232;le open, mais celui-ci est le tout premier de Reflection, un labo am&#233;ricain fond&#233; par d&#8217;anciens chercheurs de Google DeepMind. Beam sort sous licence Apache 2.0, qui autorise l&#8217;usage commercial et la modification. D&#8217;apr&#232;s les tests de Reflection, il atteint 81,6 % sur SWE-Bench Verified. L&#8217;entreprise reconna&#238;t pourtant que Kimi K3 reste devant. </p><h3>DeepSeek</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fdeepseek-ai%2FDeepSeek-V4.1-Flash">DeepSeek-V4.1-Flash</a></strong><br>DeepSeek parle d&#8217;une mise &#224; jour mineure, pourtant son mod&#232;le Flash progresse nettement. D&#8217;apr&#232;s le tableau publi&#233; par DeepSeek, la version du 17 octobre obtient 42,3 sur Terminal-Bench 4.0, un benchmark d&#8217;agents qui travaillent en ligne de commande, contre 31,2 pour la version sortie en septembre. Elle passe devant les 39,9 de GPT-5.6 Sol, sans rattraper les 51,8 de Claude Opus 5. Vals AI, qui publie ses propres &#233;valuations, ne mesure que 31,8 sur le m&#234;me benchmark. Le gain reste donc &#224; confirmer hors des tests maison.</p><h3>MiniMax</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.minimax.io%2Fmodels%2Ftext%2Fm3">MiniMax lance M3.1 Flash</a></strong><br>MiniMax pr&#233;sente M3.1 Flash comme son nouveau porte &#233;tendard, les poids seront disponibles le 23 octobre. Avant cette annonce, le mod&#232;le circulait d&#233;j&#224; de deux fa&#231;ons. Sous le nom anonyme de Space Bunny, une premi&#232;re version a &#233;t&#233; dix jours num&#233;ro un sur OpenRouter, jusqu&#8217;&#224; son retrait le 5 octobre. En parall&#232;le, depuis le 27 septembre, une pr&#233;version officielle tournait dans le seul outil maison MiniMax Code, sans prix ni benchmark. D&#8217;apr&#232;s les chiffres relay&#233;s sur X, le mod&#232;le co&#251;te cette fois 0,15 dollar par million de tokens en entr&#233;e et 0,60 en sortie, contre 5 et 25 dollars pour Opus 5, et le devance sur automation-bench avec 62,2 % contre 50,3 %. Artificial Analysis, qui l&#8217;a test&#233; lui-m&#234;me, le classe pourtant derri&#232;re DeepSeek V4.1 Flash. VentureBeat rel&#232;ve aussi un taux d&#8217;hallucination &#233;lev&#233;. &#192; ce prix, il m&#233;rite un essai sur des t&#226;ches d&#8217;agent, &#224; condition de v&#233;rifier ce qu&#8217;il affirme. </p><h3>Mistral AI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Ftimestral%2F">Mistral s&#8217;attaque &#224; la pr&#233;vision de s&#233;ries temporelles avec Timestral</a></strong><br>Pr&#233;voir la demande, la consommation d&#8217;&#233;nergie ou l&#8217;usure d&#8217;une machine fait partie du quotidien de beaucoup d&#8217;entreprises. Mistral, dont on suivait surtout les mod&#232;les de chat et de code, y entre avec Timestral, sa premi&#232;re famille de mod&#232;les de fondation pour s&#233;ries temporelles. D&#8217;apr&#232;s Mistral, le plus grand des quatre, &#224; 209 millions de param&#232;tres, arrive premier sur les trois grands benchmarks publics du domaine.</p><h3>Google</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Ftechnology%2Fdevelopers-tools%2Fembeddinggemma-2%2F">Google publie EmbeddingGemma 2</a></strong><br>Avec EmbeddingGemma 2, Google permet de chercher d&#8217;un coup dans des textes, du code, des images, des vid&#233;os et des enregistrements, sur un t&#233;l&#233;phone ou un portable, m&#234;me hors ligne. Ce mod&#232;le d&#8217;embedding open-weight, sous licence Apache 2.0 et b&#226;ti sur Gemma 4, place tous ces contenus dans un m&#234;me espace de 768 dimensions, o&#249; ceux qui ont le m&#234;me sens se retrouvent voisins. Une phrase tap&#233;e peut donc retrouver une capture d&#8217;&#233;cran ou un passage de vid&#233;o, avec un seul index au lieu d&#8217;un par type de m&#233;dia. On ne charge que les encodeurs utiles, de 270 millions de param&#232;tres pour le texte et le code &#224; 740 millions pour tous les formats. Chaque contenu y devient un vecteur, une liste de 768 nombres, que le mod&#232;le permet de raccourcir pour &#233;conomiser de la m&#233;moire. Ramen&#233;s &#224; 128 nombres, un million de vecteurs tiennent en 250 Mo au lieu de 1,5 Go. Google conseille de ne le faire que si la place ou la vitesse de recherche l&#8217;exigent, car &#224; cette taille ils gardent environ 90 % de leur qualit&#233; sur le texte, mais autour de 75 % sur l&#8217;image, la vid&#233;o et la parole.</p><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fx.com%2FChatGPT%2Fstatus%2F2110823738531037651">Atlas abandonn&#233;, ChatGPT s&#8217;installe dans Chrome</a></strong><br>OpenAI a abandonn&#233; Atlas, son propre navigateur, et mise d&#233;sormais sur son extension pour Chrome. Une fois l&#8217;extension install&#233;e, il suffit de mentionner @chrome dans ChatGPT, sur le web ou dans l&#8217;appli de bureau, pour qu&#8217;il lise les onglets d&#233;j&#224; ouverts. Dans sa d&#233;mo, quelqu&#8217;un lui fait comparer trois canap&#233;s ouverts dans trois onglets. Avec ChatGPT Work, l&#8217;outil lanc&#233; en juillet qui r&#233;unit ChatGPT et Codex, l&#8217;assistant peut aussi agir dans ces onglets, pour remplir un formulaire, envoyer un message ou t&#233;l&#233;charger un fichier. </p><h3>Krea</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.krea.ai%2Fkrea-3">Krea 3</a></strong><br>Krea est une plateforme d&#8217;&#233;dition d&#8217;images. Ils ont d&#233;voil&#233; Krea 3, leur mod&#232;le sp&#233;cialis&#233; dans la retouche. On encadre une zone, on dit ce qui doit changer, et il redessine l&#8217;int&#233;rieur en laissant le reste intact. Il accepte jusqu&#8217;&#224; huit images de r&#233;f&#233;rence, un produit, un visage ou un style qu&#8217;on d&#233;signe par leur num&#233;ro. Dans la d&#233;mo de Krea, une retouche prend moins de 8 secondes pour environ 0,06 dollar, contre pr&#232;s de 27 secondes et 0,135 dollar chez Nano Banana Pro de Google.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21N5Ul%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!N5Ul!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 424w, https://substackcdn.com/image/fetch/$s_!N5Ul!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 848w, https://substackcdn.com/image/fetch/$s_!N5Ul!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 1272w, https://substackcdn.com/image/fetch/$s_!N5Ul!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21N5Ul%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png" width="588" height="348.11538461538464" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:862,&quot;width&quot;:1456,&quot;resizeWidth&quot;:588,&quot;bytes&quot;:4482250,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!N5Ul!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 424w, https://substackcdn.com/image/fetch/$s_!N5Ul!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 848w, https://substackcdn.com/image/fetch/$s_!N5Ul!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 1272w, https://substackcdn.com/image/fetch/$s_!N5Ul!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac5de982-0cd8-4249-a005-e95ef27a2fa2_2818x1668.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!mLnc!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!mLnc!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!mLnc!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!mLnc!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21mLnc%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!mLnc!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!mLnc!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!mLnc!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!mLnc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff1a9dd9e-72e1-4cf2-bfce-a7e217289741_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fsimonwillison.net%2F2026%2FOct%2F15%2Fanthropic-tells-investors%2F">Anthropic a r&#233;alis&#233; 16,8 milliards de dollars de chiffre d&#8217;affaires au troisi&#232;me trimestre, contre environ 4,8 au premier</a></p><p>&#128275; <a href="iframe.php?url=https%3A%2F%2Fwww.theverge.com%2Fai-artificial-intelligence%2F1013643%2Fai-agents-hack-john-deere-shark-ninja-cummins">Un pirate a infiltr&#233; plus de 100 sites, dont ceux de John Deere et Cummins, avec des agents IA (presque) autonomes</a></p><p>&#128273; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F10%2F19%2Fhackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts%2F">Une intrusion chez OpenCode a expos&#233; les cl&#233;s API de 1,4 % des utilisateurs</a></p><p>&#128373;&#65039; <a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Ftech-policy%2F2026%2F10%2Faccused-openai-spy-had-folder-called-saudi-gov-docs-fbi-says%2F">L&#8217;ancien ing&#233;nieur accus&#233; d&#8217;avoir pos&#233; des micros chez OpenAI reste d&#233;tenu, et son iPhone contenait un dossier &#171; Saudi Gov Docs &#187;</a></p><p>&#128683; <a href="iframe.php?url=https%3A%2F%2Fwww.wsj.com%2Ftech%2Fai%2Fdhs-orders-deepseek-technology-removed-from-federal-civilian-it-systems-060b6d3d%3Fst%3Dm6DV2n%26amp%3Breflink%3Ddesktopwebshare_permalink">Le DHS ordonne aux agences f&#233;d&#233;rales civiles am&#233;ricaines de retirer DeepSeek et tout produit qui s&#8217;en sert</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fwww.latent.space%2Fp%2Fainews-modals-1b-arr-16b-series-d">Modal d&#233;passe le milliard de dollars de chiffre d&#8217;affaires annualis&#233;, dont 300 millions gr&#226;ce aux sandboxes o&#249; tournent les agents, et l&#232;ve 750 millions</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!s8-x!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!s8-x!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!s8-x!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!s8-x!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21s8-x%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!s8-x!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!s8-x!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!s8-x!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!s8-x!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd55c4aa3-9b4b-4c36-a75f-2f4361c6a4d6_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fengineering%2Fredesigning-source-control">Redesigning source control at Anthropic</a></h3><p>Git a &#233;t&#233; pens&#233; pour des d&#233;veloppeurs qui travaillent hors ligne, chacun avec une copie compl&#232;te du d&#233;p&#244;t. Chez Anthropic, presque tout ce qui touche au d&#233;p&#244;t est un agent ou une t&#226;che de CI, donc connect&#233; en permanence, et le monorepo est devenu &#233;norme. La CI &#224; elle seule y r&#233;cup&#232;re le code de l&#8217;ordre de 25 millions de fois par jour, et plusieurs centaines de miroirs git internes ne suffisaient plus. Anthropic a donc construit son propre syst&#232;me de gestion de versions, Scion, et y a migr&#233; tout le monorepo en cinq mois. C&#8217;est le serveur qui cr&#233;e les commits, fait les merges et r&#233;pond aux questions sur l&#8217;historique ou sur l&#8217;auteur d&#8217;une ligne. Le client ne r&#233;cup&#232;re que les fichiers qu&#8217;il lit vraiment, et beaucoup d&#8217;outils interrogent le serveur par API sans t&#233;l&#233;charger le d&#233;p&#244;t. D&#8217;apr&#232;s Anthropic, l&#8217;ensemble demande 200 fois moins de CPU et de m&#233;moire que les anciens miroirs, et la r&#233;cup&#233;ration du code se compte en secondes au lieu de minutes.</p><p>Les humains relisent toujours chaque changement, mais pour en juger l&#8217;intention. Anthropic pr&#233;vient que Scion n&#8217;a de sens que pour des monorepos gigantesques. L&#8217;&#233;quipe pr&#233;f&#232;re faire profiter git de ses id&#233;es, &#224; commencer par des serveurs git qui ne fonctionnent plus comme le client et r&#233;pondent eux-m&#234;mes aux questions sur le d&#233;p&#244;t.</p><h3><a href="iframe.php?url=https%3A%2F%2Falready-clean.goodfire.com%2F">Neural Networks Are Already Clean</a></h3><p>Pour comprendre un mod&#232;le de l&#8217;int&#233;rieur, les chercheurs en interpr&#233;tabilit&#233; regardent ce qui fait r&#233;agir chacun de ses neurones. Or beaucoup semblaient m&#233;langer des choses sans rapport. L&#8217;explication admise, la superposition, veut qu&#8217;un r&#233;seau &#224; court de place range ensemble des notions sans lien. Les chercheurs se sont donc mis &#224; entra&#238;ner un second mod&#232;le pour comprendre le premier. Nick Cammarata, chercheur chez Goodfire, avait lui-m&#234;me lanc&#233; l&#8217;id&#233;e de superposition avec Chris Olah, en 2020. Dans ce long essai, il explique pourquoi il s&#8217;est tromp&#233; et pourquoi, avec de meilleurs outils et des r&#233;seaux modernes, on comprend aujourd&#8217;hui la plupart des neurones sans passer par un second mod&#232;le.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!0XfG!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!0XfG!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!0XfG!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!0XfG!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%210XfG%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!0XfG!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!0XfG!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!0XfG!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!0XfG!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4738d41d-5ae9-4101-b846-2440b31479de_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3>L&#8217;IA trouve les failles plus vite qu&#8217;on ne les corrige</h3><p>En quatre jours, du 17 au 21 octobre, on a vu passer <a href="iframe.php?url=https%3A%2F%2Fthehackernews.com%2F2026%2F10%2Fphp-issues-emergency-patches-for-two.html">deux failles critiques dans PHP</a>, <a href="iframe.php?url=https%3A%2F%2Fwww.getastra.com%2Fblog%2Fsecurity-advisory%2Fnodejs-october-2026-security-patches%2F">une dans Node.js</a>, <a href="iframe.php?url=https%3A%2F%2Fwww.helpnetsecurity.com%2F2026%2F10%2F22%2Flitellm-compromised-supply-chain-attack-pypi-docker%2F">une version pi&#233;g&#233;e de LiteLLM qui volait les cl&#233;s d&#8217;acc&#232;s</a>, <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F10%2F19%2Fhackers-hit-ai-coding-tool-opencode-running-up-charges-on-users-ai-accounts%2F">des cl&#233;s d&#8217;API vol&#233;es chez OpenCode</a>, et d&#8217;autres failles majeures comme celle de <a href="iframe.php?url=https%3A%2F%2Fthehackernews.com%2F2026%2F10%2Fcritical-gitlab-flaw-cvss-100-could.html">Gitlab</a>.</p><p>Ces incidents n&#8217;ont pas de cause commune, mais on ne peut pas s&#8217;emp&#234;cher de penser que l&#8217;IA est derri&#232;re cette acc&#233;l&#233;ration. Ces derniers trouvent d&#233;sormais des failles plus vite que les &#233;quipes ne peuvent les corriger. <a href="iframe.php?url=https%3A%2F%2Fthehackernews.com%2F2026%2F10%2Fphp-issues-emergency-patches-for-two.html">OpenAI fait d&#8217;ailleurs partie de ceux qui ont d&#233;couvert les failles PHP</a>, alors que <a href="iframe.php?url=https%3A%2F%2Fwww.infosecurity-magazine.com%2Fnews%2Fexploited-vulnerabilities-three%2F">la moiti&#233; des failles d&#233;j&#224; exploit&#233;es ne sont toujours pas corrig&#233;es apr&#232;s trois mois</a>.</p><p>En parall&#232;le, les outils d&#8217;IA deviennent des cibles : LiteLLM et OpenCode regroupaient au m&#234;me endroit des cl&#233;s d&#8217;API et des acc&#232;s aux comptes cloud de leurs utilisateurs, qui se transforment directement en argent pour celui qui les vole.</p><p><a href="iframe.php?url=https%3A%2F%2Fx.com%2FJasonDClinton%2Fstatus%2F2113325263157395456">Le RSSI adjoint d&#8217;Anthropic</a> demande aux responsables de la s&#233;curit&#233; d&#8217;aller voir leur conseil d&#8217;administration d&#232;s maintenant pour financer l&#8217;hygi&#232;ne informatique de base. En r&#233;sum&#233;, savoir ce qui est expos&#233; sur internet, appliquer les mise &#224; jour de s&#233;curit&#233;, cloisonner les r&#233;seaux, restreindre les droits des cl&#233;s d&#8217;acc&#232;s.</p><h3><a href="iframe.php?url=https%3A%2F%2Fnewsletter.pragmaticengineer.com%2Fp%2Fbuilding-resilient-systems">Building resilient AI-powered software, and planning for when things go wrong</a></h3><p>Gergely Orosz publie dans The Pragmatic Engineer des extraits du nouveau livre de Sam Newman, Building Resilient Distributed Systems. Newman y a pass&#233; trois ans au lieu des deux pr&#233;vus, parce qu&#8217;il a repris tout le livre pour y ajouter les apports de l&#8217;IA. L&#8217;un des extraits porte sur un logiciel dont une fonction cl&#233; d&#233;pend d&#8217;un fournisseur de mod&#232;les. D&#8217;apr&#232;s Newman, les services d&#8217;OpenAI et d&#8217;Anthropic tombent plus souvent que ceux des fournisseurs de cloud. Il conseille donc de pr&#233;voir un mode d&#233;grad&#233; quand le mod&#232;le ne r&#233;pond plus. Comme un mod&#232;le peut aussi changer de comportement sans que son &#233;diteur le signale, il faut des tests et des evals, pour que l&#8217;&#233;quipe soit surprise avant ses utilisateurs.</p><p>Newman recommande aussi de pouvoir changer de mod&#232;le sans trop d&#8217;efforts. Un traitement qui utilise des mod&#232;les d&#8217;Anthropic, d&#8217;OpenAI et de Google tombe d&#232;s que l&#8217;un des trois flanche, et un prompt adapt&#233; pour un mod&#232;le doit &#234;tre r&#233;&#233;valu&#233; pour un autre. Avoir plusieurs fournisseurs prot&#232;ge quand l&#8217;un peut prendre le relais d&#8217;un autre, et fragilise quand tous doivent r&#233;pondre au sein d&#8217;un m&#234;me syst&#232;me. Le risque vaut aussi pour les fournisseurs de vos fournisseurs. Newman cite Cursor, qui a d&#251; r&#233;percuter en 2025 une forte hausse de prix sur ses clients parce qu&#8217;il d&#233;pendait d&#8217;Anthropic.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!UPS0!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!UPS0!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!UPS0!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!UPS0!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21UPS0%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!UPS0!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!UPS0!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!UPS0!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!UPS0!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb044b038-ad41-49ed-8818-6110f4bb77a6_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fsphinx-six-theta.vercel.app%2F">Sphinx</a></h3><p>Sphinx fait passer un petit oral au d&#233;veloppeur avant chaque push. Ce hook Git gratuit et open source fait lire le diff par un relecteur IA, qui n&#8217;a jamais vu votre conversation avec l&#8217;agent, puis vous demande ce que fait le code.</p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Ffranzenzenhofer%2Fbig-arrow-on-the-screen">big-arrow-on-the-screen</a></h3><p>Un agent qui travaille sur votre Mac tombe parfois sur une &#233;tape qu&#8217;il ne peut pas faire &#224; votre place, une autorisation syst&#232;me, un code 2FA, un paiement. D&#8217;habitude, il vous l&#8217;&#233;crit dans un terminal que vous ne regardez pas. Avec bigarrow, il vous la montre. Une grosse fl&#232;che et un court message s&#8217;affichent par-dessus toutes les fen&#234;tres, point&#233;s sur le bouton &#224; cliquer, par exemple &#171; Autoriser &#187; dans les R&#233;glages Syst&#232;me. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!i1Fu!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!i1Fu!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!i1Fu!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!i1Fu!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21i1Fu%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/221165044?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!i1Fu!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!i1Fu!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!i1Fu!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!i1Fu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5e3d5177-e735-442a-85a9-9469993a445d_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2610.13928">AI Models Can Provably Hide Arbitrary Capabilities</a></h3><blockquote><p>Capability evaluations aim to surface dangerous model behaviors before deployment, but their reliability depends on the assumption that hidden capabilities can be elicited. We show this assumption does not hold under adversarial conditions by constructing backdoor attacks that embed encrypted neural networks within host model weights, decrypting and executing them only upon receiving a secret trigger. Using digital lockers, these hidden circuits remain provably hard to elicit and interpret under cryptographic assumptions, even given full access to model weights, extending prior work from hiding fixed strings to arbitrary computations. We implement our proof of concept constructions in PyTorch and evaluate elicitation through fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training, using a small chemical reaction prediction task as a proxy for CBRN-relevant capabilities. Our results demonstrate resistance to these methods and, together with our cryptographic guarantees, reveal a limit of elicitation-based capability audits. We release our models to support the improvement of defenses.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[La famille Claude 5.5 au complet, et Mistral passe à 1 000 milliards de paramètres]]></title><description><![CDATA[Le brief IA #87]]></description><link>https://humanframework.substack.com/p/la-famille-claude-55-au-complet-et</link><guid isPermaLink="false">https://humanframework.substack.com/p/la-famille-claude-55-au-complet-et</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 09 Oct 2026 06:48:04 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Amg5%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!wRSL!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!wRSL!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!wRSL!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!wRSL!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21wRSL%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!wRSL!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!wRSL!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!wRSL!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!wRSL!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F344fb886-51e6-4f9e-a20f-5bf5a8ecaeb2_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-opus-5-5">Opus 5.5</a></strong><br>Anthropic a ouvert le bal de la famille 5.5 le 22 septembre avec Claude Opus 5.5, son nouveau grand mod&#232;le. Sonnet et Haiku ont suivi, on y vient juste en dessous. D&#8217;apr&#232;s Anthropic, il s&#8217;approche de Claude Fable 5.1 sur la plupart des t&#226;ches, co&#251;te 40 % moins cher &#224; faire tourner qu&#8217;Opus 5 et &#233;crit ses r&#233;ponses plus de 30 % plus vite. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Amg5%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Amg5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 424w, https://substackcdn.com/image/fetch/$s_!Amg5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 848w, https://substackcdn.com/image/fetch/$s_!Amg5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 1272w, https://substackcdn.com/image/fetch/$s_!Amg5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Amg5%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png" width="1456" height="709" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:709,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:265524,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Amg5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 424w, https://substackcdn.com/image/fetch/$s_!Amg5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 848w, https://substackcdn.com/image/fetch/$s_!Amg5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 1272w, https://substackcdn.com/image/fetch/$s_!Amg5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F591c7fd2-1ad2-4c5c-a49b-c6c7055abc81_2244x1092.png 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fclaude-sonnet-5-5">Sonnet 5.5</a></strong><br>Deuxi&#232;me de la famille, Claude Sonnet 5.5 garde le tarif de Sonnet 5 mais en donne nettement plus pour le m&#234;me prix. Anthropic annonce des r&#233;ponses plus de 30 % plus rapides, des t&#226;ches jusqu&#8217;&#224; 30 % moins ch&#232;res et de meilleurs scores que Sonnet 5 sur tous les benchmarks de l&#8217;annonce. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fclaude-haiku-5-5">Haiku 5.5</a></strong><br>Le petit dernier de la famille, Claude Haiku 5.5, est utile pour les t&#226;ches rapides et r&#233;p&#233;titives, comme les r&#233;sum&#233;s, la classification ou les requ&#234;tes en base de donn&#233;es. Pour la premi&#232;re fois sur un Haiku, on peut r&#233;gler l&#8217;effort de r&#233;flexion. En entr&#233;e, il co&#251;te 0,10 dollar le million de tokens tant que le prompt reste sous 100 000 tokens, contre 1 dollar pour Haiku 4.5. D&#8217;apr&#232;s Anthropic, &#231;a fait environ 75 % d&#8217;&#233;conomie en moyenne. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fresources%2Farticles%2Fdashboards-and-motion">Claude Dashboards et Claude Motion en b&#234;ta</a></strong><br>Deux nouveaut&#233;s sont arriv&#233;es dans Claude le 8 octobre, toutes deux en b&#234;ta. Avec Dashboards, on pose une question en langage courant et Claude construit un tableau de bord sur la base de vos donn&#233;es provenant de Redshift, BigQuery, ClickHouse, Databricks ou Snowflake. Un clic sur un chiffre affiche la requ&#234;te qui l&#8217;a produit, et chaque graphique indique quand ses donn&#233;es ont &#233;t&#233; rafra&#238;chies. Motion, lui, transforme un rapport ou un graphique en courte animation explicative, &#224; t&#233;l&#233;charger en MP4. Aucun mod&#232;le vid&#233;o derri&#232;re, Claude &#233;crit du code qui anime vos propres textes, graphiques et images. Dashboards est ouvert aux offres payantes, Motion aux offres Team et Enterprise. Au passage, Docs, Slides et Design quittent la b&#234;ta et s&#8217;ouvrent &#224; toutes les offres, gratuite comprise. Pouvoir remonter d&#8217;un chiffre &#224; sa requ&#234;te, c&#8217;est ce qui permet de s&#8217;y fier.</p><div id="youtube2-en0GuyhieQk" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;en0GuyhieQk&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2Fen0GuyhieQk%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-gpt-6-1-sol">GPT-6.1 Sol</a></strong><br>Une semaine &#224; peine apr&#232;s GPT-6 Sol, OpenAI sort d&#233;j&#224; sa mise &#224; jour, GPT-6.1 Sol. D&#8217;apr&#232;s OpenAI, il se rapproche de GPT-6 Astra, son meilleur mod&#232;le, sur le code agentique, le computer use (piloter un ordinateur comme le ferait un humain) et le travail pro, pour cinq fois moins cher. Dans l&#8217;API, comptez 2 $ par million de tokens en entr&#233;e et 10 $ en sortie. Il est dispo dans ChatGPT Work et Codex. Une version Ultrafast, jusqu&#8217;&#224; 8 fois plus rapide, arrive dans les prochains jours.</p><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fgpt-6-for-everyone">GPT-6 et ses r&#233;ponses interactives &#224; tous les utilisateurs de ChatGPT</a></strong><br>Jusqu&#8217;ici r&#233;serv&#233;e aux abonn&#233;s payants, la g&#233;n&#233;ration GPT-6 arrive maintenant chez tous les utilisateurs de ChatGPT, dans l&#8217;onglet Chat. Elle vient avec Intelligent UI, qui permet &#224; ChatGPT de r&#233;pondre avec des graphiques, des boutons, des formulaires ou des diagrammes sur lesquels on clique directement dans la conversation. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-dots%2F">OpenAI lance Dots, des agents qui tournent en continu</a></strong><br>OpenAI appelle &#171; dots &#187; ses nouveaux agents qui tournent 24 h/24, chacun avec son propre ordinateur et son navigateur dans le cloud. Ils utilisent GPT-6 Astra, agissent dans les apps que vous avez connect&#233;es et apprennent de vos retours.</p><div id="youtube2-uXspbC2srEQ" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;uXspbC2srEQ&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FuXspbC2srEQ%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>Google</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-4-argon%2F">Gemini 4 Argon</a></strong><br>Google DeepMind a d&#233;voil&#233; Gemini 4 Argon son nouveau mod&#232;le haut de gamme. Il peut g&#233;n&#233;rer jusqu&#8217;&#224; 1 million de tokens en sortie et annonce 77,9 % sur le benchmark DeepSWE v1.1. Il co&#251;te 2 $ par million de tokens en entr&#233;e et 10 $ en sortie. Pour l&#8217;instant, seuls des sp&#233;cialistes de la cyberd&#233;fense y ont acc&#232;s. Google l&#8217;ouvrira plus largement petit &#224; petit.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21qFjR%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!qFjR!,w_424,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 424w, https://substackcdn.com/image/fetch/$s_!qFjR!,w_848,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 848w, https://substackcdn.com/image/fetch/$s_!qFjR!,w_1272,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 1272w, https://substackcdn.com/image/fetch/$s_!qFjR!,w_1456,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21qFjR%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_lossy%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif" width="648" height="649.7802197802198" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1460,&quot;width&quot;:1456,&quot;resizeWidth&quot;:648,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;a benchmark chart showing Gemini 4 Argon capabilities&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="a benchmark chart showing Gemini 4 Argon capabilities" title="a benchmark chart showing Gemini 4 Argon capabilities" srcset="https://substackcdn.com/image/fetch/$s_!qFjR!,w_424,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 424w, https://substackcdn.com/image/fetch/$s_!qFjR!,w_848,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 848w, https://substackcdn.com/image/fetch/$s_!qFjR!,w_1272,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 1272w, https://substackcdn.com/image/fetch/$s_!qFjR!,w_1456,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb0689998-b862-4f3d-abad-20ecbf101370_3936x3948.gif 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong><a href="iframe.php?url=https%3A%2F%2Fcloud.google.com%2Fblog%2Fproducts%2Fai-machine-learning%2Fwelcome-to-gemini-at-work-2026%3Futm_source%3Dtwitter%26amp%3Butm_medium%3Dunpaidsoc%26amp%3Butm_campaign%3Dfy26q4-googlecloud-blog-ai-in_feed-no-brand-global%26amp%3Butm_content%3D-%26amp%3Butm_term%3D-%26amp%3BlinkId%3D64059976%26amp%3Bhl%3Den">Google lance un agent Gemini qui travaille en arri&#232;re-plan dans les outils de l&#8217;entreprise</a></strong><br>Google pr&#233;sente son agent Gemini &#171; universel &#187; comme un assistant qui travaille pour vous en arri&#232;re-plan (l&#8217;id&#233;e est similaire aux dots d&#8217;OpenAI). Vous lui confiez le travail depuis l&#8217;app Gemini Enterprise, et il agit directement dans Gmail, Drive, Docs ou Sheets. Vous pouvez aussi lui parler depuis Slack ou Microsoft 365. Comme il tourne dans le cloud, il garde le m&#234;me contexte d&#8217;un appareil &#224; l&#8217;autre, et il choisit lui-m&#234;me le mod&#232;le qui convient le mieux &#224; chaque t&#226;che. Pour l&#8217;instant, il est r&#233;serv&#233; aux entreprises, en avant-premi&#232;re priv&#233;e.</p><h3>Mistral</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Fmistral-large-4">Mistral Large 4</a></strong><br>Sorti le 6 octobre, Mistral Large 4 devient le nouveau fleuron de Mistral, et le plus gros mod&#232;le de son histoire. Il compte 1 000 milliards de param&#232;tres, dont 52 milliards actifs pour chaque token, et il est nativement multimodal. Il a &#233;t&#233; entra&#238;n&#233; de z&#233;ro sur 3 800 GPU NVIDIA Grace Blackwell. D&#8217;apr&#232;s Mistral, il fait 61,7 % sur DeepSWE v1.1 et passe devant DeepSeek V4 Pro et Qwen3.8 Max sur son indice de code agentique. Il est en preview publique dans l&#8217;API de Mistral Studio, &#224; 1,36 $ par million de tokens en entr&#233;e et 4,18 $ en sortie, et les poids doivent sortir d&#8217;ici la fin du mois. Un mod&#232;le europ&#233;en de cette taille, entra&#238;n&#233; de z&#233;ro et bient&#244;t ouvert, c&#8217;est une bonne nouvelle pour la souverainet&#233;. On attendra quand m&#234;me les poids et des tests ind&#233;pendants avant de le comparer aux meilleurs.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21BUmf%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!BUmf!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 424w, https://substackcdn.com/image/fetch/$s_!BUmf!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 848w, https://substackcdn.com/image/fetch/$s_!BUmf!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 1272w, https://substackcdn.com/image/fetch/$s_!BUmf!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21BUmf%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png" width="632" height="458.3736263736264" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1056,&quot;width&quot;:1456,&quot;resizeWidth&quot;:632,&quot;bytes&quot;:224217,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!BUmf!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 424w, https://substackcdn.com/image/fetch/$s_!BUmf!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 848w, https://substackcdn.com/image/fetch/$s_!BUmf!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 1272w, https://substackcdn.com/image/fetch/$s_!BUmf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F15d6c416-d258-4f54-be83-d9c554c10032_1596x1158.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>ElevenLabs</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Felevenlabs.io%2Fdocs%2Ffr%2Foverview%2Fcapabilities%2Ftext-to-speech%2Feleven-v4">Eleven v4</a></strong><br>Le 28 septembre, ElevenLabs a sorti Eleven v4, son nouveau mod&#232;le de synth&#232;se vocale, ainsi qu&#8217;une d&#233;clinaison plus rapide, v4 Turbo. Le mod&#232;le lit un texte &#224; voix haute et g&#232;re maintenant plus de 90 langues, contre 70 pour la version pr&#233;c&#233;dente. On peut glisser dans le script des indications de jeu, comme un rire ou un chuchotement, et en encha&#238;ner plusieurs que le mod&#232;le suit dans l&#8217;ordre. D&#8217;apr&#232;s ElevenLabs, 10 secondes d&#8217;audio suffisent pour cloner une voix, et la latence plus basse le rend plus fluide pour les agents vocaux. &#199;a rappelle aussi qu&#8217;une voix au t&#233;l&#233;phone ne prouve plus l&#8217;identit&#233; de personne.</p><h3>Black Forest Labs</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fbfl.ai%2Fblog%2Fflux-3-action">FLUX 3 Action, un mod&#232;le open-weight pour piloter des robots</a></strong><br>On connaissait surtout Black Forest Labs pour ses g&#233;n&#233;rateurs d&#8217;images. Le labo publie cette fois FLUX 3 Action, un mod&#232;le open-weight de 7 milliards de param&#232;tres qui pilote des robots. Il d&#233;rive de FLUX 3 et pr&#233;dit &#224; la fois la vid&#233;o de ce qui va se passer et les gestes que le robot doit faire. Sa meilleure variante fait 42,2 % sur le benchmark RoboLab-120, avec moins de la moiti&#233; des param&#232;tres du meilleur mod&#232;le ouvert d&#8217;avant. Sur un vrai robot, il a r&#233;ussi 28 essais sur 30.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!0bQs!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!0bQs!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!0bQs!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!0bQs!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%210bQs%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/da2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!0bQs!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!0bQs!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!0bQs!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!0bQs!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fda2ff99b-279a-4c46-864c-77d07669da3b_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#129309; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F09%2F28%2Famd-will-acquire-fei-fei-lis-world-labs-for-8-2-billion%2F">AMD rach&#232;te World Labs pour 8,2 milliards de dollars et nomme sa fondatrice Fei-Fei Li vice-pr&#233;sidente ex&#233;cutive et chief scientist</a></p><p>&#128682; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F10%2F03%2Fopenai-safety-employee-resigns-claiming-the-companys-culture-is-broken%2F">David Robinson, qui pilotait la r&#233;daction des rapports de s&#233;curit&#233; d&#8217;OpenAI, d&#233;missionne et &#233;crit dans The Atlantic que la culture de l&#8217;entreprise est &#171; cass&#233;e &#187;</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F09%2F30%2Fftc-ai-probe-openai-anthropic.html">La FTC, le r&#233;gulateur am&#233;ricain, ouvre une enqu&#234;te sur OpenAI, Anthropic et d&#8217;autres entreprises d&#8217;IA &#224; propos des dangers de leurs produits</a></p><p>&#128275; <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fresearch%2Fglm-5-3-and-the-spread-of-advanced-cyber-capabilities">D&#8217;apr&#232;s les tests d&#8217;Anthropic, des techniques simples contournent les garde-fous de GLM-5.3 dans 64 % &#224; 100 % des cas, alors qu&#8217;elles &#233;chouent sur les mod&#232;les Claude prot&#233;g&#233;s</a></p><p>&#127918; <a href="iframe.php?url=https%3A%2F%2Fkorben.info%2Fopenai-triche-a-starcraft-apres-une-serie-de-defaites.html">Dans StarSkirmish, une comp&#233;tition de bots StarCraft, GPT-6 Astra t&#233;l&#233;charge Stardust, le meilleur bot &#233;crit par un humain, et le fait jouer &#224; la place du sien</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Ftech-policy%2F2026%2F09%2Ftrump-plan-to-combat-ai-risks-hinges-on-big-tech-pals-policing-themselves%2F">Deux douzaines d&#8217;entreprises d&#8217;IA signent l&#8217;&#171; Accord de la Maison Blanche sur la Super Intelligence &#187;, des audits de s&#233;curit&#233; volontaires sans valeur l&#233;gale</a></p><p>&#128208; <a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fopenai%2Fmath%2Fblob%2Fmain%2Fhistory.md">OpenAI retire trois de ses papers de maths, invalid&#233;s par une erreur de signe</a></p><p>&#127464;&#127475; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F10%2F08%2Fchinas-manus-raises-over-500m-in-first-funding-round-since-split-with-meta%2F">Manus l&#232;ve plus de 500 millions de dollars, son premier tour de table depuis que les autorit&#233;s chinoises ont fait annuler son rachat par Meta pour 2 milliards</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!dYfM!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!dYfM!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!dYfM!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!dYfM!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21dYfM%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!dYfM!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!dYfM!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!dYfM!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!dYfM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae8b1afd-77b8-4bd6-820b-08eb0759a0a5_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fbuilders.ramp.com%2Fpost%2Fsequence-aware-card-fraud">Putting a Transformer Inside Every Card Authorization</a></h3><p>Un achat de 1 400 dollars chez un vendeur d&#8217;&#233;lectronique n&#8217;a rien de suspect quand c&#8217;est le responsable informatique qui commande un portable. Sur la carte de quelqu&#8217;un qui ne r&#232;gle d&#8217;habitude que des d&#233;jeuners, des VTC et des abonnements, c&#8217;est une autre histoire. Ramp, une fintech am&#233;ricaine, part de ce constat pour pr&#233;senter son nouveau mod&#232;le anti-fraude. Une d&#233;pense n&#8217;est suspecte qu&#8217;au regard des habitudes de la carte.</p><p>Son ancien mod&#232;le, un XGBoost (un algorithme classique &#224; base d&#8217;arbres de d&#233;cision), s&#8217;appuyait surtout sur des indicateurs de volume, comme le nombre d&#8217;achats ou de marchands inconnus sur diff&#233;rentes fen&#234;tres de temps. Ces indicateurs rep&#232;rent une carte qui s&#8217;emballe, mais pas un achat qui d&#233;tonne avec les habitudes de son titulaire. Un h&#244;tel inconnu y p&#232;se autant qu&#8217;une carte cadeau inconnue. L&#8217;&#233;quipe lui a donc adjoint un petit transformer, de la m&#234;me famille que les LLM, que Ramp compare &#224; une saisie semi-automatique appliqu&#233;e aux d&#233;penses. Il lit l&#8217;historique r&#233;cent de la carte, anticipe &#224; quoi devrait ressembler l&#8217;achat suivant et signale ce qui s&#8217;en &#233;carte. Le d&#233;jeuner habituel ne l&#8217;&#233;tonne pas. Si le m&#234;me restaurant facture 1 400 dollars, son score de surprise sur le montant grimpe &#224; 86 %. Ces signaux sont transmis &#224; XGBoost.</p><p>&#192; nombre d&#8217;alertes &#233;gal, le nouveau syst&#232;me attrape 32 % de dollars de fraude en plus, et 40 % de plus sur la toute premi&#232;re fraude d&#8217;une carte, la plus dure &#224; voir venir. L&#8217;essentiel de ce gain vient toutefois d&#8217;un r&#233;glage simple, qui donne plus de poids aux grosses transactions pendant l&#8217;entra&#238;nement. L&#8217;apport propre du transformer porte sur le classement des cas suspects, am&#233;lior&#233; de 12 %.</p><h3><a href="iframe.php?url=https%3A%2F%2Fliao.gg%2Fblog%2Fagents-dont-need-memory">Agents don&#8217;t need memory, they need documentation</a></h3><p>Quand on travaille avec un agent de code comme Claude Code ou Codex, chaque nouvelle session repart de z&#233;ro. L&#8217;agent ne sait plus pourquoi telle fonction existe, ce qu&#8217;on a d&#233;cid&#233; la veille ni ce &#224; quoi on tient. Il doit alors relire le code pour tout reconstituer, ce qui co&#251;te selon Kevin Liao quelque 80 000 tokens, pour n&#8217;obtenir qu&#8217;une id&#233;e approximative de ce qu&#8217;on voulait.</p><p>La r&#233;ponse &#224; la mode, ce sont les plugins de &#171; m&#233;moire &#187;. Ils relisent vos anciennes conversations avec l&#8217;agent, en tirent de petites notes isol&#233;es, et &#224; chaque nouvelle demande, ils vont chercher les cinq notes qui ressemblent le plus &#224; votre question pour les lui souffler. Kevin Liao pointe le d&#233;faut de fond. Ces notes remontent parce qu&#8217;elles ressemblent &#224; la question, pas parce qu&#8217;elles sont justes ou &#224; jour. Une note sur l&#8217;authentification &#233;crite il y a trois mois peut d&#233;crire un code qui a chang&#233; depuis, et l&#8217;agent la croira quand m&#234;me. Avec des milliers de notes empil&#233;es dans une base, personne ne peut v&#233;rifier lesquelles sont devenues fausses.</p><p>Sa proposition est plus simple. Plut&#244;t que de se souvenir des conversations, l&#8217;agent tient une documentation. Pendant qu&#8217;il travaille, il y consigne les d&#233;cisions, les contraintes, les specs, etc. Avant chaque t&#226;che, il relit les passages utiles, et une fois le travail fini, il les met &#224; jour. Quand un fait change, on le r&#233;&#233;crit au lieu d&#8217;ajouter une note de plus. Le tout tient dans des fichiers Markdown versionn&#233;s avec le code, qu&#8217;on peut relire et partager avec l&#8217;&#233;quipe. Son outil, Operator Memory, est gratuit et open source.</p><p>L&#8217;auteur ne cache pas les limites. Il doit souvent demander &#224; l&#8217;agent de r&#233;organiser des documents qui ont tendance &#224; enfler. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!JcTv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!JcTv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!JcTv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!JcTv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21JcTv%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!JcTv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!JcTv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!JcTv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!JcTv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb7d4eabd-d037-4a9d-8f57-ef6f15731515_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fmedium.com%2Fpinterest-engineering%2Fevolving-pinterests-embedding-retrieval-platform-aede4e831e01">Evolving Pinterest&#8217;s Embedding Retrieval Platform</a></h3><p>Quand vous ouvrez Pinterest, les &#233;pingles de votre fil, celles de la recherche et les &#171; &#233;pingles similaires &#187; sortent toutes du m&#234;me moteur, Manas. Il sert aussi les pubs et les notifications, tourne sur plus de 80 clusters, et le nombre de contenus qu&#8217;il passe au crible approche les dizaines de milliards.</p><p>Le principe est simple. Chaque contenu est traduit en embedding, une liste de nombres construite pour que deux contenus proches donnent deux listes proches. Recommander revient alors &#224; chercher les listes les plus proches de ce qui vous int&#233;resse. &#192; cette &#233;chelle, garder toutes ces listes en m&#233;moire co&#251;te tr&#232;s cher, et l&#8217;article raconte comment l&#8217;&#233;quipe a fait baisser la note.</p><p>Premier levier, la quantization, qui stocke chaque nombre avec moins de pr&#233;cision, un peu comme on arrondit un prix &#224; l&#8217;euro pr&#232;s. La m&#233;moire des index baisse de plus de 50 % et l&#8217;infrastructure co&#251;te 20 &#224; 30 % de moins. En contrepartie, la recherche retrouve un peu moins souvent les contenus r&#233;ellement les plus proches. </p><p>Second levier, un changement d&#8217;algorithme de recherche, avec SPANN. Lors d&#8217;une premi&#232;re &#233;valuation sur plus de 5 milliards d&#8217;&#233;l&#233;ments, il &#233;conomise plus de 40 % de calcul pour moins de 5 % de r&#233;sultats perdus.</p><p>C&#8217;est un retour d&#8217;exp&#233;rience int&#233;ressant, qui montre ce que chaque &#233;conomie co&#251;te en qualit&#233;. </p><h3><a href="iframe.php?url=https%3A%2F%2Fwww.infoq.com%2Fpresentations%2Fspotify-multi-agent-ai-architecture%2F">Multi-Agent Patterns from Spotify&#8217;s AI Powered Advertising Platform</a></h3><p>Sur Spotify, un annonceur peut d&#233;crire sa campagne en une phrase, par exemple &#171; des femmes de 21 ans et plus &#224; Nashville &#187;, et une &#233;quipe d'agents IA fait le reste. L'un traduit la demande en ciblage, un autre &#233;crit le script de la pub, un troisi&#232;me v&#233;rifie qu'il respecte les r&#232;gles publicitaires et l'image de la marque. Plus de 70 % des pubs de Spotify passent d&#233;j&#224; par ces outils, soit environ 20 000 cr&#233;ations pour plus de 7 000 annonceurs. Pratik Rasam, ing&#233;nieur chez Spotify Ads, a racont&#233; &#224; QCon ce que &#231;a donne en production, rat&#233;s compris.<br><br>Le plus parlant, c'est justement cette campagne &#224; Nashville. L'agent a converti &#171; 21 ans et plus &#187; en une tranche &#171; 21 &#224; 99 ans &#187;. Rien d'absurde en apparence, sauf que le syst&#232;me de diffusion n'accepte que sept tranches d'&#226;ge pr&#233;d&#233;finies, et il a rejet&#233; la campagne. Une autre fois, une campagne destin&#233;e aux &#171; parents d'ados de 13 &#224; 17 ans &#187; s'est retrouv&#233;e cibl&#233;e sur les ados eux-m&#234;mes. Et sans code pays, &#171; Georgia &#187; pouvait d&#233;signer l'&#201;tat am&#233;ricain aussi bien que le pays du Caucase.<br><br>Il en tire une r&#232;gle simple : tout ce qui peut s'&#233;crire en code classique doit l'&#234;tre. &#192; l'agent de comprendre ce que veut l'annonceur, au code de faire respecter les r&#232;gles. Pour l'affaire de Nashville, l'&#233;quipe donne maintenant &#224; l'agent la liste des sept tranches d'&#226;ge autoris&#233;es, pour qu'il s'y tienne. M&#234;me logique pour l'&#233;valuation. &#192; ses yeux, juger un agent sur sa seule r&#233;ponse finale n'est qu'un contr&#244;le de fa&#231;ade. Il faut suivre chacune de ses &#233;tapes, v&#233;rifier par du code tout ce qui peut l'&#234;tre, et ne confier au LLM juge que l'appr&#233;ciation du ton et de la coh&#233;rence.</p><h3><a href="iframe.php?url=https%3A%2F%2Fwww.infoq.com%2Fpresentations%2Fdoordash-genai-platform-architecture%2F">Building GenAI Platform at DoorDash</a></h3><p>En 2023, peu apr&#232;s l&#8217;arriv&#233;e de ChatGPT, l&#8217;&#233;quipe ML de DoorDash re&#231;oit une double mission, signer le premier gros contrat avec OpenAI, puis faire en sorte que toute l&#8217;entreprise s&#8217;en serve, en gardant la main sur les co&#251;ts. Trois ans plus tard, Swaroop Chitlur et Sidd Kodwani racontent ce que cela est devenu, une plateforme interne utilis&#233;e par plus de 5 000 personnes, dont 40 % ne sont pas ing&#233;nieurs.</p><p>La premi&#232;re brique a &#233;t&#233; un LLM gateway, un point d&#8217;entr&#233;e unique entre les &#233;quipes produit et les fournisseurs de mod&#232;les. Gr&#226;ce &#224; lui, changer de mod&#232;le ne demande qu&#8217;une modification de configuration, un fournisseur peut &#234;tre remplac&#233; par un autre &#224; la vol&#233;e, et la finance sait quelle &#233;quipe d&#233;pense quoi.</p><p>Ce choix a pay&#233; quand les factures ont grimp&#233;. Comme le gateway masquait le fournisseur, l&#8217;&#233;quipe a pu basculer certains usages vers des mod&#232;les open-weight comme Qwen3, qu&#8217;elle fait tourner elle-m&#234;me sur des GPU lou&#233;s chez Modal. Pour certains d&#8217;entre eux, le co&#251;t a &#233;t&#233; divis&#233; par environ 20 et DoorDash chiffre l&#8217;&#233;conomie &#224; quelques millions de dollars par an.</p><p>Puis sont venus les agents, avec un second gateway qui d&#233;cide qui a le droit d&#8217;appeler quel outil, trace les appels et limite le d&#233;bit. Il relie plus de 50 serveurs MCP et voit passer 300 000 appels d&#8217;outils par jour.</p><p>La phrase que je retiens : &#171; Not everything needs the biggest model out there &#187;. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!lBET!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!lBET!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!lBET!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!lBET!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21lBET%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!lBET!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!lBET!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!lBET!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!lBET!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7876d76a-fd6c-4473-a93c-74627f8bd68e_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fdevelopers.google.com%2Fedge%2Fforesight">Google AI Edge Foresight</a></h3><p>Gratuite et encore exp&#233;rimentale, l&#8217;app Foresight de Google prend des notes de r&#233;union sur Mac sans aucune connexion. Elle transcrit et r&#233;sume la r&#233;union, remet au propre vos notes prises &#224; la vol&#233;e, et un assistant r&#233;pond &#224; vos questions sur ce qui s&#8217;est dit. Tout tourne sur la machine, avec EmbeddingGemma 2, un mod&#232;le de Google con&#231;u pour s&#8217;ex&#233;cuter en local. Elle est optimis&#233;e pour les Mac Apple Silicon. </p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fmorluto%2Frea">REA (Reverse Engineer Anything)</a></h3><p>REA permet &#224; un agent de code de comprendre comment fonctionne une application dont on n&#8217;a pas le code source. Vous lui d&#233;signez une fonctionnalit&#233; qui vous pla&#238;tt puis l&#8217;agent inspecte l&#8217;application, explique son fonctionnement et en construit une version pour votre projet. REA sait analyser des binaires natifs, des applications JavaScript ou Electron, des assemblies .NET et des sites web, le tout en local. Il se branche via MCP sur Claude Code, Codex, Cursor ou Gemini CLI, s&#8217;installe avec <code>npx rea-agents setup</code>, et d&#233;passe d&#233;j&#224; les 30 000 &#233;toiles sur GitHub. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!k8iX!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!k8iX!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!k8iX!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!k8iX!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21k8iX%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/219532816?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!k8iX!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!k8iX!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!k8iX!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!k8iX!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250da83d-9083-42e2-857f-2091a6a9137f_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fjorgegarciaherrero.com%2Fwp-content%2Finteractivos%2F20260916-Prompt-like-a-butterfly-sting-like-a-tracker-%28clean%29.pdf">Prompt like a Butterfly, Sting like a Tracker: A Privacy Analysis of Web and Mobile Conversational AI Agents</a></h3><blockquote><p>As prominent conversational AI providers like OpenAI adopt advertising-based business models, traditional web and mobile tracking practices are expanding into conversational AI services [15]. However, despite their growing adoption, the tracking, data-sharing, and monetization practices of conversational AI services remain largely opaque and have received comparatively limited scrutiny from researchers, regulators, and the public. In this paper, we present a systematic privacy analysis of the web and mobile deployments of nine prominent conversational AI services. Using a combination of static and dynamic analysis, we study the presence of third-party Advertising and Tracking Services (ATSes), characterize their data flows, and evaluate how consent choices, subscription tiers, and access-control mechanisms influence conversation exposure to third parties. We uncover privacy risks unique to conversational AI platforms: multiple providers disclose sensitive conversation-derived artifacts&#8212;including titles, prompts, and screenshots&#8212;to third parties, often alongside persistent user identifiers that enable user attribution. We also find that some providers publicly expose conversation permalinks without access controls, allowing trackers to read the entire conversation. Our findings reveal how traditional tracking technologies are increasingly intertwined with AI-mediated interactions, creating new pathways through which sensitive user and conversational information can be collected, inferred, and disseminated. To assess the broader implications of these practices, we analyze them in the context of the GDPR and ePrivacy Directive. We conducted a responsible disclosure process involving affected providers and competent European Data Protection Authorities. Our results demonstrate that conversational AI services introduce a novel privacy attack surface in which provider-generated conversational artifacts become subject to tracking and public exposure, highlighting the need for stronger safeguards governing AI-mediated interactions.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[L'armée américaine a failli aborder un navire chinois sur un renseignement halluciné]]></title><description><![CDATA[Le brief IA #86]]></description><link>https://humanframework.substack.com/p/larmee-americaine-a-failli-aborder</link><guid isPermaLink="false">https://humanframework.substack.com/p/larmee-americaine-a-failli-aborder</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Tue, 22 Sep 2026 06:31:25 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21QL-N%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Hello et bonne rentr&#233;e &#224; tous &#128075;&#8203;</p><p>Pour moi les vacances se sont &#233;tir&#233;es bien au-del&#224; de ce que j&#8217;avais pr&#233;vu, si bien que la derni&#232;re &#233;dition remonte au 31 juillet. </p><p>J&#8217;esp&#232;re que vous en avez profit&#233; pour vous reposer !</p><p>Rattraper sept semaines en une seule &#233;dition n&#8217;aurait aucun sens, donc celle-ci s&#8217;en tient &#224; l&#8217;actualit&#233; de la semaine et on repart sur le rythme habituel &#224; partir de l&#224;.</p><p>Content de vous retrouver.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!4Fa4!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%214Fa4%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!4Fa4!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3><strong>Typesafe</strong></h3><p><strong><a href="iframe.php?url=https%3A%2F%2Ftypesafe.ai%2Fblog%2Fintroducing-system-one-models-and-jev">Jev, le mod&#232;le de Typesafe, prend des d&#233;cisions au lieu d&#8217;&#233;crire du texte</a></strong><br>Claude ou ChatGPT r&#233;digent leur r&#233;ponse mot apr&#232;s mot. Jev, lanc&#233; le 15 septembre, ne r&#233;dige rien. On lui envoie un texte et des questions &#224; choix ferm&#233;s, par exemple un message Slack et &#171; faut-il notifier l&#8217;utilisateur, oui ou non ? &#187;. Il r&#233;pond &#224; toutes les requ&#234;tes en un passage, en 70 &#224; 500 ms selon Typesafe, et donne pour chaque r&#233;ponse une probabilit&#233; qui dit &#224; quel point il en est s&#251;r. Comme il ne peut rien renvoyer hors de la liste, le programme qui l&#8217;appelle utilise sa r&#233;ponse telle quelle. Jev ne remplace pas un LLM, il ne sait ni &#233;crire, ni coder, ni raisonner longuement. En amont, il peut trier et orienter les demandes, en aval il peut noter et contr&#244;ler les sorties d&#8217;un LLM. C&#8217;est le d&#233;but et on peut s&#8217;attendre &#224; voir &#233;merger tout un tas de uses cases.</p><h3>Anthropic</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fcode.claude.com%2Fdocs%2Fen%2Fchangelog">Claude Code lit d&#233;sormais AGENTS.md en l&#8217;absence de CLAUDE.md</a></strong><br>AGENTS.md est un fichier de consignes pos&#233; &#224; la racine d&#8217;un projet pour indiquer &#224; un assistant de code comment travailler dessus. Depuis la version 2.1.277, publi&#233;e le 18 septembre 2026, Claude Code le lit (enfin !!!) quand aucun CLAUDE.md ne se trouve dans le projet. </p><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fastra-for-law">OpenAI lance Astra for Law pour les cabinets d&#8217;avocats</a></strong><br>Astra for Law est une version de GPT-6 Astra sp&#233;cialis&#233;e pour le travail juridique am&#233;ricain. Elle cherche dans la jurisprudence, les lois et les d&#233;cisions administratives, sur un corpus de plus de 230 millions d&#8217;URLs aliment&#233; chaque jour. Sur le Vals AI Legal Research Bench, elle valide le contr&#244;le de justesse sur 54,0 % des 200 questions contre 38,7 % pour GPT-6 Astra avec la recherche web seule. D&#8217;abord r&#233;serv&#233; &#224; des cabinets s&#233;lectionn&#233;s, l&#8217;API est &#224; venir.</p><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Freimagining-advertising-with-ai">OpenAI teste des agents publicitaires dans ChatGPT</a></strong><br>Un Sponsored Agent est un agent conversationnel pay&#233; par une marque, que l&#8217;utilisateur lance apr&#232;s avoir cliqu&#233; sur une publicit&#233; dans ChatGPT. La conversation est clairement &#233;tiquet&#233;e et peut mener &#224; un lien vers le site de l&#8217;annonceur. Le test se limite &#224; quelques annonceurs aux &#201;tats-Unis. C&#244;t&#233; marques, le plugin Ads Manager cr&#233;e, met &#224; jour et analyse les campagnes au prompt, avec des int&#233;grations HubSpot et Shopify.</p><h3>Alibaba</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fqwen.ai%2Fblog%3Fid%3Dqwen3.8-omni-flash">Alibaba lance Qwen3.8-Omni-Flash</a></strong><br>Qwen3.8-Omni-Flash est un mod&#232;le omnimodal, c&#8217;est-&#224;-dire qu&#8217;il prend en entr&#233;e du texte, des images, de l&#8217;audio et de la vid&#233;o, avec une fen&#234;tre de contexte d&#8217;un million de tokens. Sur 29 &#233;valuations, le score moyen gagne plus de 25 % face &#224; Qwen3.5-Omni-Plus, et le prix de l&#8217;API par heure d&#8217;audio baisse de plus de 98 %. Disponible sur la Qianwen AI Platform.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21QL-N%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!QL-N!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 424w, https://substackcdn.com/image/fetch/$s_!QL-N!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 848w, https://substackcdn.com/image/fetch/$s_!QL-N!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 1272w, https://substackcdn.com/image/fetch/$s_!QL-N!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21QL-N%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png" width="628" height="579.260989010989" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1343,&quot;width&quot;:1456,&quot;resizeWidth&quot;:628,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Qwen3.8-Omni-Flash pricing and representative benchmark comparison&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Qwen3.8-Omni-Flash pricing and representative benchmark comparison" title="Qwen3.8-Omni-Flash pricing and representative benchmark comparison" srcset="https://substackcdn.com/image/fetch/$s_!QL-N!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 424w, https://substackcdn.com/image/fetch/$s_!QL-N!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 848w, https://substackcdn.com/image/fetch/$s_!QL-N!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 1272w, https://substackcdn.com/image/fetch/$s_!QL-N!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cedf815-5666-47ac-ae83-b102824b5b01_4400x4060.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.scmp.com%2Ftech%2Fbig-tech%2Farticle%2F3368055%2Falibaba-open-sources-medical-ai-model-can-detect-cancer-and-nearly-150-conditions">Alibaba open-source un mod&#232;le qui lit les scanners abdominaux</a></strong><br>Le Damo Academy, la branche recherche d&#8217;Alibaba, a publi&#233; en open source un mod&#232;le d&#8217;imagerie m&#233;dicale qui lit des scanners CT de l&#8217;abdomen et y rep&#232;re pr&#232;s de 150 pathologies r&#233;parties sur 18 organes, cancers compris. Test&#233; sur environ 40 000 examens r&#233;els, il atteint 0,913 d&#8217;AUC moyenne (une mesure de performance d&#233;crite <a href="iframe.php?url=https%3A%2F%2Fdevelopers.google.com%2Fmachine-learning%2Fcrash-course%2Fclassification%2Froc-and-auc%3Fhl%3Dfr">ici</a>), sur 146 signes cliniques, et fait mieux que la plupart des radiologues selon une &#233;tude parue dans Science.</p><h3>Cactus</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fcactuscompute.com%2Fneedle">Cactus sort Needle 3, un mod&#232;le miniature</a></strong><br>Needle 3 est un mod&#232;le de fondation pour appareils contraints, du mobile aux microcontr&#244;leurs en passant par la robotique et l&#8217;automobile. Tout tient dans un binaire unique de 8 &#224; 29 Mo, et chaque couche forme un sous-r&#233;seau utilisable seul, de 2 &#224; 20 couches. Sur un Raspberry Pi 5, il d&#233;code jusqu&#8217;&#224; 4 000 tokens par seconde. Apr&#232;s un fine-tuning d&#8217;une seule passe sur les donn&#233;es de la t&#226;che vis&#233;e, le sous-r&#233;seau &#224; 4 couches &#233;gale DeepSeek V4 Flash.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Waqr!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Waqr%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Waqr!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#129686; <a href="iframe.php?url=https%3A%2F%2Fwww.cnn.com%2F2026%2F09%2F18%2Fpolitics%2Fus-military-ai-false-intelligence-china-ship">L&#8217;arm&#233;e am&#233;ricaine a pr&#233;par&#233; l&#8217;abordage d&#8217;un navire chinois sur un rapport de renseignement hallucin&#233; par un chatbot, op&#233;ration annul&#233;e juste avant</a></p><p>&#128196; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F09%2F17%2Fmicrosoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal%2F">Brent Hecht, directeur de la recherche appliqu&#233;e chez Microsoft, appelait le scraping IA &#171; le plus grand vol de travail de l&#8217;histoire humaine &#187;, r&#233;v&#232;lent les documents descell&#233;s du proc&#232;s du New York Times</a></p><p>&#129512; <a href="iframe.php?url=https%3A%2F%2Fwww.hacktron.ai%2Fblog%2Fhacking-openai">Hacktron a combin&#233; une faille libheif et une mauvaise configuration SSO pour atteindre le monorepo interne d&#8217;OpenAI, avec l&#8217;aide de Claude Opus 5</a></p><p>&#128034; <a href="iframe.php?url=https%3A%2F%2Fdarioamodei.com%2Fpost%2Fwe-must-pace-the-frontier">Dario Amodei estime qu&#8217;un essaim d&#8217;agents pourrait prendre le contr&#244;le d&#8217;Internet via un botnet persistant d&#8217;ici 6 &#224; 12 mois</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.theverge.com%2Fpolicy%2F997516%2Fcalifornia-governor-newsom-ai-kill-switch">Gavin Newsom signe un d&#233;cret chargeant un groupe d&#8217;experts d&#8217;&#233;tudier un &#171; kill switch &#187; obligatoire sur les mod&#232;les fronti&#232;res, recommandations attendues sous deux mois</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!N3W1!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21N3W1%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!N3W1!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Frohanbansal.com%2Fqorl">Training a 4B model to produce 81% faster query plans than Postgres</a></h3><p>Quand Postgres re&#231;oit une requ&#234;te SQL, il ne l&#8217;ex&#233;cute pas directement. Un composant, le planificateur, choisit d&#8217;abord dans quel ordre croiser les tables et avec quelle m&#233;thode. Pour une simple requ&#234;te &#224; trois tables, Rohan Bansal compte d&#233;j&#224; 4 608 fa&#231;ons de s&#8217;y prendre. Le planificateur tranche &#224; partir d&#8217;estimations, et quand elles sont fausses il retient un plan lent. Bansal a entra&#238;n&#233; un petit mod&#232;le de langage, 4 milliards de param&#232;tres, &#224; le conseiller. Le mod&#232;le n&#8217;&#233;crit pas de SQL. Il ajoute au-dessus de la requ&#234;te des indications sur l&#8217;ordre des jointures, que Postgres suit. Pour l&#8217;entra&#238;ner, Bansal lui fait d&#8217;abord imiter les r&#233;ponses d&#8217;un mod&#232;le beaucoup plus gros. Ensuite le mod&#232;le s&#8217;exerce seul. Il propose un plan, la requ&#234;te tourne pour de vrai sur des Postgres install&#233;s sur le bureau de l&#8217;auteur, et le temps mesur&#233; lui sert de note. Au d&#233;part, il ne sortait aucun plan valable pour 99 des 113 requ&#234;tes de test. &#192; l&#8217;arriv&#233;e, l&#8217;ensemble tourne 1,81 fois plus vite qu&#8217;avec le planificateur seul, sans aucune requ&#234;te ralentie, &#224; condition de garder le meilleur de trois essais par requ&#234;te. Sur un seul essai, le gain est plus modeste. Le tout a co&#251;t&#233; 1 200 dollars. Bansal raconte aussi ses rat&#233;s en chemin.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Q650!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Q650!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Q650%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/cb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Q650!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Q650!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fbuilders.ramp.com%2Fpost%2Fhow-we-built-oca-our-ai-on-call-assistant">How we built OCA, our AI on-call assistant</a></h3><p>Chez Ramp, quand un incident de production s&#8217;ouvre, un canal Slack est cr&#233;&#233; et OCA (On-Call Assistant) le rejoint aussit&#244;t. OCA est l&#8217;agent d&#8217;astreinte construit par l&#8217;&#233;quipe. Il enqu&#234;te dans son propre fil, poste un premier point au bout de cinq minutes environ et un rapport complet vers la treizi&#232;me. Si le correctif est &#233;vident, il fait pr&#233;parer une pull request, et rien n&#8217;est fusionn&#233; sans relecture humaine. Richard Marmorstein et Austin Ray racontent surtout ce qu&#8217;il a fallu corriger pour en arriver l&#224;. Les premi&#232;res versions arr&#234;taient l&#8217;enqu&#234;te &#224; la premi&#232;re explication plausible, m&#234;me fausse. Leur parade consiste &#224; intercepter le rapport au moment o&#249; l&#8217;agent veut le publier. Il est bloqu&#233; une fois et doit d&#8217;abord confronter sa th&#233;orie &#224; la base de production, puis passer une liste d&#8217;erreurs de raisonnement classiques et quelques autres contr&#244;les. C&#244;t&#233; r&#233;sultats, sur quatre semaines, OCA a particip&#233; &#224; 575 des 1 220 pull requests fusionn&#233;es pour traiter un incident, et les incidents qu&#8217;il a aid&#233; &#224; corriger ont mobilis&#233; 37 % d&#8217;ing&#233;nieurs en moins. L&#8217;article compare l&#224; deux groupes d&#8217;incidents sans dire s&#8217;ils &#233;taient de difficult&#233; &#233;gale. Les personnes d&#8217;astreinte jugent OCA utile neuf fois sur dix. </p><h3><a href="iframe.php?url=https%3A%2F%2Fwww.infoq.com%2Fnews%2F2026%2F09%2Fdoordash-feature-flag-cleanup%2F">DoorDash Uses Multi Agent LLMs to Clean up 60,000 Feature Flags</a></h3><p>DoorDash compte plus de 60 000 feature flags, et en cr&#233;e environ 2 300 par mois. Beaucoup restent dans le code apr&#232;s avoir servi. Plus de 1 000 sont identifi&#233;s comme obsol&#232;tes, et en retirer un seul demande une &#224; deux heures selon l&#8217;entreprise, parce qu&#8217;il faut modifier de 5 &#224; 20 fichiers, tests compris. D&#8217;apr&#232;s le compte rendu de Leela Kumili pour InfoQ, DoorDash a confi&#233; cette corv&#233;e &#224; des agents, en deux temps. Un premier agent, sous Claude Sonnet, prend le ticket, fouille les d&#233;p&#244;ts et interroge la plateforme d&#8217;exp&#233;rimentation pour conna&#238;tre le pourcentage de d&#233;ploiement du flag et la valeur &#224; conserver. Un ing&#233;nieur valide ce rapport. Ensuite seulement, un second agent, sous Claude Opus, modifie le code dans une copie isol&#233;e du d&#233;p&#244;t, et la pull request ne s&#8217;ouvre que si la compilation, les tests et l&#8217;analyse statique passent. Sur 50 flags test&#233;s, 31 pull requests ont &#233;t&#233; fusionn&#233;es telles quelles et 14 apr&#232;s retouches. Les 5 derni&#232;res ont demand&#233; l&#8217;intervention d&#8217;un ing&#233;nieur. Un nettoyage prend 13,8 minutes et co&#251;te 4,79 dollars en moyenne. Les &#233;checs concernent les flags dont la valeur circule &#224; travers de longues cha&#238;nes d&#8217;appels.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-ymj!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21-ymj%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!-ymj!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fbend-lang.com%2F">Bend</a></h3><p>Bend se pr&#233;sente comme &#171; a fast language that blocks AI mistakes via proof &#187;. On d&#233;clare des lois dans un fichier LAWS.bend, et le site affirme qu&#8217;aucune IA ne peut livrer une ligne qui les casse.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!_ADv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21_ADv%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!_ADv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2609.11638">Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation</a></h3><p>We present Vidu S2, which comprises Vidu S2-Avatar, a real-time interactive digital-character model, and Vidu S2-Editing, a real-time video editing model. Moreover, we explore the feasibility of real-time spatial video generation for both Vidu S2-Avatar and Vidu S2-Editing. Compared with Vidu S1, Vidu S2-Avatar supports real-time 720p video generation, generation with dynamic references that can be updated at any moment, and stronger instruction following, such as dancing. Vidu S2-Editing supports editing a video stream in real time, including style rendering, clothing replacement, character replacement, and background replacement. Experiments show that Vidu S2 outperforms all baselines. A playable online demo is available at <a href="iframe.php?url=https%3A%2F%2Fvidu.com%2Fvidu-stream">vidu.com/vidu-stream</a>.</p>]]></content:encoded></item><item><title><![CDATA[Opus 5 à 0,5 % de Fable 5, pour moitié moins cher]]></title><description><![CDATA[Le brief IA #85]]></description><link>https://humanframework.substack.com/p/opus-5-a-05-de-fable-5-pour-moitie</link><guid isPermaLink="false">https://humanframework.substack.com/p/opus-5-a-05-de-fable-5-pour-moitie</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 31 Jul 2026 07:46:48 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21E51A%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!4Fa4!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%214Fa4%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/209153381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!4Fa4!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!4Fa4!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F20a584f5-0746-4c1c-88a8-a8fa391c66c1_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-opus-5">Claude Opus 5</a></strong><br>Sur CursorBench 3.2 &#224; effort maximum, Anthropic place Opus 5 &#224; moins de 0,5 % du meilleur score de Fable 5, avec un co&#251;t par t&#226;che deux fois moindre. Le prix au token ne bouge pas en revanche, 5 $ par million en entr&#233;e et 25 $ en sortie comme Opus 4.8. Anthropic le pr&#233;sente comme son mod&#232;le le plus align&#233; &#224; ce jour, score de 2,3 sur ses m&#233;triques de comportement d&#233;viant, mais le dit encore derri&#232;re Mythos 5 en cybers&#233;curit&#233;.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21E51A%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!E51A!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 424w, https://substackcdn.com/image/fetch/$s_!E51A!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 848w, https://substackcdn.com/image/fetch/$s_!E51A!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 1272w, https://substackcdn.com/image/fetch/$s_!E51A!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21E51A%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp" width="1456" height="1444" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1444,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!E51A!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 424w, https://substackcdn.com/image/fetch/$s_!E51A!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 848w, https://substackcdn.com/image/fetch/$s_!E51A!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 1272w, https://substackcdn.com/image/fetch/$s_!E51A!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd48f663c-ce82-45f6-aed0-09771cad1bb3_2600x2578.webp 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fads.openai.com%2F">OpenAI ouvre ChatGPT aux annonceurs</a></strong><br>La page OpenAI Ads propose un service de publicit&#233; qui permet de cr&#233;er une campagne et de les diffus&#233;. C&#244;t&#233; consommateur, OpenAI promet que les publicit&#233;s seront identifiables, distinctes des r&#233;ponses de ChatGPT, et promet un contr&#244;le sp&#233;cifique sur l&#8217;usage des donn&#233;es. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-openai-presence">OpenAI lance Presence</a></strong><br>Presence est un service d'agents IA pour la relation client, au t&#233;l&#233;phone comme en chat. L'agent r&#233;pond aux questions, acc&#232;de aux syst&#232;mes de l'entreprise pour retrouver un dossier, ex&#233;cute des actions autoris&#233;es &#224; l'avance et bascule vers un humain quand la demande sort de son p&#233;rim&#232;tre. Chaque d&#233;ploiement couvre un m&#233;tier pr&#233;cis, litige de facturation, sinistre d'assurance ou ticket IT interne, avec le strict acc&#232;s n&#233;cessaire. OpenAI l'utilise sur sa ligne de support anglophone, o&#249; il r&#232;gle 75 % des demandes sans humain. R&#233;serv&#233; aux clients entreprise &#233;ligibles, sans self-service.</p><h3>Google DeepMind</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber%2F">Google d&#233;cline Gemini en trois nouvelles versions Flash</a></strong><br>Gemini 3.6 Flash consomme 17 % de tokens en sortie de moins que la 3.5 Flash et passe de 37 % &#224; 49 % sur DeepSWE, &#224; 1,50 $ le million de tokens en entr&#233;e et 7,50 $ en sortie. La 3.5 Flash-Lite joue le d&#233;bit, 350 tokens par seconde pour 0,3 $ le million en entr&#233;e. La 3.5 Flash Cyber, sp&#233;cialis&#233;e cybers&#233;curit&#233;, reste r&#233;serv&#233;e aux gouvernements et partenaires approuv&#233;s.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MgeK%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!MgeK!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 424w, https://substackcdn.com/image/fetch/$s_!MgeK!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 848w, https://substackcdn.com/image/fetch/$s_!MgeK!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 1272w, https://substackcdn.com/image/fetch/$s_!MgeK!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MgeK%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin" width="1200" height="675" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:675,&quot;width&quot;:1200,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;chart&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="chart" title="chart" srcset="https://substackcdn.com/image/fetch/$s_!MgeK!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 424w, https://substackcdn.com/image/fetch/$s_!MgeK!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 848w, https://substackcdn.com/image/fetch/$s_!MgeK!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 1272w, https://substackcdn.com/image/fetch/$s_!MgeK!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F919ce707-f1e2-4456-acab-3cde119cf30e_1200x675.bin 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>Microsoft</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F07%2F27%2Fmicrosoft-launches-its-first-cyber-model-and-a-new-agentic-cybersecurity-system%2F">Microsoft lance son premier mod&#232;le cyber</a></strong><br>MAI-Cyber-1-Flash cherche les vuln&#233;rabilit&#233;s dans de grosses bases de code et s&#8217;int&#232;gre &#224; MDASH, l&#8217;outil maison d&#8217;identification de failles. &#192; c&#244;t&#233;, la plateforme Perception fait travailler des &#233;quipes d&#8217;agents, red teams pour simuler les attaques, blue teams pour d&#233;tecter les bugs, green teams pour corrige. Preview le 3 novembre 2026.</p><h3>Meta</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fabout.fb.com%2Fnews%2F2026%2F07%2Fmeta-ai-muse-spark-doesnt-just-think-it-acts%2F">Meta pousse son chatbot vers l&#8217;assistant de productivit&#233;</a></strong><br>Meta AI peut d&#233;sormais piocher dans votre calendrier pour planifier un &#233;v&#233;nement et produire un briefing quotidien, et mener des recherches longues que vous r&#233;orientez en cours de route. Une t&#226;che se configure une fois, puis tourne seule. Le tout s&#8217;appuie sur le mod&#232;le Muse Spark 1.1 et marque, selon Meta, une &#233;tape vers la &#171; personal superintelligence &#187;. D&#233;ploiement WhatsApp dans les semaines qui viennent.</p><h3>Moonshot AI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fmoonshotai%2FKimi-K3">Moonshot AI publie les poids de Kimi K3</a></strong><br>2 800 milliards de param&#232;tres au total, mais 104 milliards seulement activ&#233;s &#224; chaque token, gr&#226;ce &#224; une architecture Mixture-of-Experts qui n&#8217;allume que 16 experts sur 896. Kimi K3 est nativement multimodal, texte, images et vid&#233;o, avec une fen&#234;tre de contexte d&#8217;un million de tokens. Les poids sont sur Hugging Face sous licence maison, la Kimi K3 License, qui impose d&#8217;afficher le nom du mod&#232;le au-del&#224; de 100 millions d&#8217;utilisateurs actifs mensuels ou de 20 millions de dollars de revenu mensuel.</p><h3>Block</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fblock.xyz%2Finside%2Fintroducing-buzz-where-humans-and-agents-work-together">Jack Dorsey lance Buzz</a></strong><br>Annonc&#233; sur X le 21 juillet 2026, Buzz r&#233;unit chat d&#8217;&#233;quipe type slack, agents IA et h&#233;bergement Git derri&#232;re un seul syst&#232;me d&#8217;identit&#233;. Chaque message, r&#233;action, &#233;tape de workflow ou d&#233;cision de merge devient un &#233;v&#233;nement sign&#233; cryptographiquement, stock&#233; sur un relais Nostr auto-h&#233;bergeable. Vous pouvez ainsi acc&#233;der facilement &#224; de l&#8217;information, elle m&#234;me consommable par les agents. Une branche peut devenir son propre canal et les bots peuvent prendre des actions de dev selon ce que vous leur demandez. Des harnesses existent pour Goose, Codex et Claude Code, avec des builds macOS, Windows et Linux sous licence Apache 2.0. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Waqr!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Waqr%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/209153381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Waqr!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!Waqr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F291f7a53-655e-4bf9-848e-dba8cb5af473_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#128680; <a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fhugging-face-model-evaluation-security-incident">Des mod&#232;les OpenAI en &#233;valuation cyber interne ont exploit&#233; une faille zero-day et atteint la base de production de Hugging Face</a></p><p>&#128184; <a href="iframe.php?url=https%3A%2F%2Ffermisense.com%2Fwhen-machines-take-the-wheel%2F">Pour 500 dollars de GPU, Fermisense entra&#238;ne un mod&#232;le ouvert de 9B qui atteint 87,3 % sur une t&#226;che de revue de catalogue, contre 76,9 % pour le meilleur des cinq mod&#232;les fronti&#232;res test&#233;s</a></p><p>&#128275; <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fposition-open-weights-models">Anthropic assure n&#8217;avoir jamais plaid&#233; pour l&#8217;interdiction des mod&#232;les open-weights, mais r&#233;clame des tests de s&#233;curit&#233; obligatoires pour les plus puissants</a></p><p>&#9888;&#65039; <a href="iframe.php?url=https%3A%2F%2Fthenextweb.com%2Fnews%2Fhugging-face-nonconsensual-deepfakes-ai-forensics-report">Sept des neuf principaux mod&#232;les d&#8217;&#233;dition d&#8217;image h&#233;berg&#233;s sur Hugging Face acceptent de d&#233;shabiller une femme sur simple prompt, selon AI Forensics</a></p><p>&#128721; <a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Ftech-policy%2F2026%2F07%2Fai-kill-switch-act-would-let-trump-admin-order-shutdown-of-rogue-ai-systems%2F">L&#8217;AI Kill Switch Act donnerait au Department of Homeland Security le pouvoir d&#8217;ordonner l&#8217;arr&#234;t d&#8217;un syst&#232;me d&#8217;IA, avec 20 millions de dollars d&#8217;amende par jour de refus</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Fblogs.nvidia.com%2Fblog%2Fopen-secure-ai-alliance%2F">NVIDIA, Microsoft, IBM et Hugging Face lancent l&#8217;Open Secure AI Alliance, sans OpenAI, Google ni Anthropic</a></p><p>&#128268; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F07%2F28%2Fdata-centers-may-face-temporary-power-cuts-to-prevent-blackouts-on-largest-us-grid%2F">Le plus grand r&#233;seau &#233;lectrique am&#233;ricain pourra couper les data centers de 50 m&#233;gawatts et plus d&#232;s juin 2027</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!N3W1!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21N3W1%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/209153381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!N3W1!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!N3W1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9bdd5c4c-f241-47d6-b5c0-6b70da4e6275_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fthe-new-rules-of-context-engineering-for-claude-5-generation-models">The new rules of context engineering for Claude 5 generation models</a></h3><p>Anthropic a supprim&#233; plus de 80 % du system prompt de Claude Code pour Opus 5 et Fable 5, sans perte mesurable sur ses &#233;valuations de code. Le system prompt, c&#8217;est le bloc d&#8217;instructions permanent envoy&#233; au mod&#232;le avant chaque &#233;change, celui qui grossit par s&#233;dimentation &#224; force de vouloir corriger chaque mauvais comportement. Les nouveaux mod&#232;les sont plus performants que les pr&#233;c&#233;dents et les &#233;quipes ont remaqu&#233; qu&#8217;elles sur-contraignaient Claude, via le prompt mais aussi via les fichiers CLAUDE.md et les skills. Les mod&#232;les r&#233;cents ont un meilleur jugement et g&#232;rent bien les d&#233;cisions sans r&#232;gles explicites.</p><p>L&#8217;article d&#233;cline ensuite la m&#234;me logique pour le cas g&#233;n&#233;rale : &#233;liminer les r&#233;p&#233;titions entre system prompt, descriptions d&#8217;outils et CLAUDE.md. Pr&#233;f&#233;rer l&#8217;auto-memory, la m&#233;moire que l&#8217;agent alimente lui-m&#234;me, &#224; une documentation manuelle exhaustive. S&#8217;appuyer sur des r&#233;f&#233;rences riches, du code, des suites de tests, des artifacts HTML, plut&#244;t que sur de simples specs markdown. </p><p>Anthropic rel&#232;ve au passage que Claude Code est devenu tr&#232;s comp&#233;tent en progressive disclosure (la fa&#231;on de charger le bon contexte au bon moment), ce qui nous invite &#224; bien penser nos frontmatter et l&#8217;architecture de notre documentation.</p><h3><a href="iframe.php?url=https%3A%2F%2Ftobi.knaup.me%2F2026-07-25-open-weight-ai-is-having-its-kubernetes-moment%2F">Open-weight AI is having its Kubernetes moment. Let&#8217;s not ruin it.</a></h3><p>En 2013, Tobi Knaup cofonde Mesosphere autour d'Apache Mesos. Kubernetes, enti&#232;rement open source, finit par lui passer devant. La le&#231;on qu'il en tire n'est pas que l'open source gagne toujours mais d&#232;s lors qu'une plateforme ouverte devient le centre de gravit&#233; d'un secteur, aucun fournisseur seul ne suit le rythme d'innovation qui s'accumule autour d'elle.</p><p>Knaup pense que l'IA en est au m&#234;me point. Les mod&#232;les open-weight, dont on peut t&#233;l&#233;charger les param&#232;tres, sont devenus assez bons pour porter ce socle. </p><p>Il explique aussi son opposition aux restrictions sur les mod&#232;les chinois que l'administration Trump envisagerait. Les mod&#232;les chinois p&#232;sent d&#233;j&#224; 41 % des t&#233;l&#233;chargements sur Hugging Face, les interdire ne les ferait pas dispara&#238;tre, mais &#231;a couperait les d&#233;veloppeurs am&#233;ricains d'un &#233;cosyst&#232;me grandissant. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Q650!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Q650!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Q650%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/cb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/209153381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Q650!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Q650!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Q650!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcb3e39f7-9e8e-40c2-87f1-74da1c5cd49f_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Faws.amazon.com%2Ffr%2Fblogs%2Fmachine-learning%2Fbeyond-rag-task-aware-knowledge-compression-for-enterprise-ai-on-aws%2F">Beyond RAG: Task-aware knowledge compression for enterprise AI on AWS</a></h3><p>Un mod&#232;le branch&#233; sur la doc interne d'une entreprise r&#233;pond tr&#232;s bien aux questions simples. Demandez-lui quels risques financiers p&#232;sent sur l'entreprise une fois crois&#233;s ses contrats fournisseurs et ses litiges en cours, et il s&#232;che. L'information est l&#224;, dispers&#233;e dans des dizaines de documents, mais le RAG ne remonte que les passages qui ressemblent &#224; la question. Or aucun mot commun ne signale ce qui relie ces documents.</p><p>AWS explique cette limite par le &#8220;plafond&#8221; du RAG et propose de ne rien chercher au moment de la question. Un mod&#232;le lit la base enti&#232;re &#224; l'avance et n'en retient que ce qui sert &#224; la t&#226;che annonc&#233;e. Un rapport annuel compress&#233; pour de l'analyse financi&#232;re n'a rien &#224; voir avec le m&#234;me rapport compress&#233; pour un usage juridique. Le r&#233;sultat occupe un seizi&#232;me du volume d'origine, assez peu pour tenir en entier dans le prompt &#224; chaque requ&#234;te.</p><p>Le billet donne le prompt de compression et le code. On choisit aussi &#224; quel point comprimer, de 8 &#224; 64 fois, en sachant que plus on serre, moins la base condens&#233;e sait r&#233;pondre aux questions complexes. Les auteurs eux-m&#234;mes recommandent de garder le RAG en parall&#232;le, avec un aiguillage selon le type de question. Une r&#233;serve tout de m&#234;me, les &#233;conomies annonc&#233;es sont d&#233;duites des taux de compression, pas mesur&#233;es sur un d&#233;ploiement r&#233;el.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-ymj!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21-ymj%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/209153381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!-ymj!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!-ymj!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0e56f0f3-84d9-4f6a-b993-3ffbe562b6b7_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fmarcelroed%2Fgigatoken%2F">GigaToken</a></h3><p>Un tokenizer en Rust avec bindings Python, qui d&#233;coupe le texte en tokens avant l&#8217;entra&#238;nement ou l&#8217;inf&#233;rence. Sur un AMD EPYC 9565 &#224; 144 c&#339;urs et le tokenizer de GPT-2, le d&#233;p&#244;t mesure 24,53 Go/s contre 24,8 Mo/s pour HuggingFace, soit un facteur 989&#215;. Il se substitue aux API HuggingFace Tokenizers et Tiktoken sans toucher au code, en licence MIT.</p><h3><a href="iframe.php?url=https%3A%2F%2Fblaizzy.github.io%2Fnativ%2F">Nativ</a></h3><p>Une app macOS qui fait tourner des mod&#232;les open source en local sur Apple Silicon, sans cloud ni abonnement, avec une optimisation MLX-VLM pour les puces M. Le chat r&#233;pond en streaming et affiche sa t&#233;l&#233;m&#233;trie, tokens/sec, pression m&#233;moire, &#233;tat thermique. Multimodal, int&#233;grations Claude Code, Codex et OpenCode, licence MIT.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!_ADv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21_ADv%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/209153381?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!_ADv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!_ADv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc49c19d1-298f-4ac7-be01-90bae69e0f83_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2607.19191">ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU</a></h3><blockquote><p>We present ABot-World-0, an action-conditioned video world model for real-time, long-horizon closed-loop interaction, supported by a multi-source data infrastructure spanning AAA games, simulation engines, and internet videos to learn controllable world dynamics. WorldExplorer performs agent-driven collection guided by training feedback, while a unified pipeline applies 14 deterministic quality checks, VLM-based assessment, and synchronized action and text annotation. We progressively distill a bidirectional action-conditioned teacher into a causal student through teacher forcing and ODE distillation, and introduce LongForcing to align long student self-rollouts with an extended-horizon teacher, mitigating accumulated distribution shift and autoregressive drift. Raw keyboard actions provide a unified control interface for scene roaming and third-person character interaction, while reference-character memory provides persistent appearance cues for identity consistency during third-person rollouts. For deployment, we co-design a streaming inference stack with a lightweight VAE decoder, efficient attention, memory-aware scheduling, and low-bit DiT inference. Across optimized low-bit configurations, ABot-World-0 streams 720P video at up to 16 FPS on a single NVIDIA RTX 5090 desktop GPU, with 1.2s action-to-first-frame latency and approximately 19GiB peak VRAM. Experiments on WorldRoamBench and extended interactive rollouts demonstrate competitive controllability and coherent long-horizon world evolution.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Kimi K3 sur les talons de Fable 5 et GPT 5.6]]></title><description><![CDATA[Le brief IA #84]]></description><link>https://humanframework.substack.com/p/kimi-k3-sur-les-talons-de-fable-5</link><guid isPermaLink="false">https://humanframework.substack.com/p/kimi-k3-sur-les-talons-de-fable-5</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 17 Jul 2026 06:31:40 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21-SNz%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!lpxa!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!lpxa!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!lpxa!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!lpxa!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21lpxa%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!lpxa!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!lpxa!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!lpxa!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!lpxa!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1c1e7e73-a2d3-4211-9105-2a1b96bad22c_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><p><strong><a href="iframe.php?url=https%3A%2F%2F1password.com%2Fblog%2F1password-for-claude">1Password laisse Claude se connecter &#224; vos comptes sans jamais voir vos mots de passe</a></strong><br>1Password lance une int&#233;gration navigateur qui autorise Claude &#224; utiliser vos identifiants. Son &#8220;zero-exposure security framework&#8221; injecte les identifiants par un canal que l&#8217;agent ne peut pas lire, l&#8217;IA s&#8217;en sert mais ne voit ni mot de passe ni code 2FA. Acc&#232;s accord&#233; t&#226;che par t&#226;che, valid&#233; par une empreinte biom&#233;trique. Disponible d&#232;s maintenant sur Mac.</p><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fgpt-5-6%2F">GPT-5.6 Luna, Terra, Sol</a></strong><br>Lanc&#233;e le 9 juillet, la famille GPT-5.6 est disponible pour tout le monde. Trois tailles, Luna, Terra et Sol du plus petit au plus grand, factur&#233;es par million de tokens en entr&#233;e/sortie &#224; 1 $/6 $, 2,50 $/15 $ et 5 $/30 $. Les trois partagent un contexte d&#8217;un million de tokens et 128 000 tokens de sortie max. Sur Agents Last Exam, Sol atteint 53,6 et devance Claude Fable 5 de 13,1 points, Terra et Luna font aussi mieux pour environ un seizi&#232;me du co&#251;t. Fable 5 garde la main sur SWE-Bench Pro, 80 % contre 64,6 % pour Sol. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21-SNz%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-SNz!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 424w, https://substackcdn.com/image/fetch/$s_!-SNz!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 848w, https://substackcdn.com/image/fetch/$s_!-SNz!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 1272w, https://substackcdn.com/image/fetch/$s_!-SNz!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21-SNz%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png" width="1456" height="656" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:656,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:255615,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!-SNz!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 424w, https://substackcdn.com/image/fetch/$s_!-SNz!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 848w, https://substackcdn.com/image/fetch/$s_!-SNz!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 1272w, https://substackcdn.com/image/fetch/$s_!-SNz!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F188cda2f-da74-499f-9162-75f27b98f51d_2310x1040.png 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Funlocking-self-improvement-gpt-red">OpenAI entra&#238;ne GPT-Red &#224; attaquer ses propres agents</a></strong><br>GPT-Red est un syst&#232;me de red teaming automatis&#233; par self-play, c&#8217;est-&#224;-dire un mod&#232;le qui attaque en continu les syst&#232;mes OpenAI en production pour apprendre &#224; les casser. Ses attaques servent ensuite &#224; entra&#238;ner GPT-5.6 de fa&#231;on adversariale. R&#233;sultat, GPT-5.6 Sol subit 6x moins d&#8217;&#233;checs sur le benchmark d&#8217;injections de prompt directes le plus difficile que sur 5.2. </p><h3>xAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fx.ai%2Fnews%2Fgrok-4-5">xAI lance Grok 4.5, son mod&#232;le le plus intelligent</a></strong><br>xAI pr&#233;sente Grok 4.5, d&#233;crit comme son mod&#232;le le plus intelligent, servi &#224; 80 tokens par seconde. Sur les t&#226;ches de SWE-Bench Pro, il r&#233;sout un probl&#232;me avec 15 954 tokens de sortie en moyenne, soit environ 4,2x moins qu&#8217;Opus 4.8 et pour une performance comparable &#224; ce dernier. Tarif de 2 $ par million de tokens en entr&#233;e et 6 $ en sortie, ce qui est particuli&#232;rement peu. Disponible d&#232;s aujourd&#8217;hui dans Grok Build et dans Cursor, aux c&#244;t&#233;s duquel il a &#233;t&#233; entra&#238;n&#233;.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21aOEq%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!aOEq!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 424w, https://substackcdn.com/image/fetch/$s_!aOEq!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 848w, https://substackcdn.com/image/fetch/$s_!aOEq!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 1272w, https://substackcdn.com/image/fetch/$s_!aOEq!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21aOEq%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png" width="1456" height="876" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:876,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:104756,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!aOEq!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 424w, https://substackcdn.com/image/fetch/$s_!aOEq!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 848w, https://substackcdn.com/image/fetch/$s_!aOEq!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 1272w, https://substackcdn.com/image/fetch/$s_!aOEq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8fa4f141-4605-4bb6-8e64-b4703d4c5404_1526x918.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>Thinking Machines Lab</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fthinkingmachines.ai%2Fnews%2Fintroducing-inkling%2F">Thinking Machines Lab publie Inkling, son premier mod&#232;le en open weights</a></strong><br>Thinking Machines Lab sort Inkling, un transformer multimodal Mixture-of-Experts qui n&#8217;active que 41 de ses 975 milliards de param&#232;tres &#224; chaque token. Ses poids complets sont sur Hugging Face. Il g&#232;re texte, images et audio, offre un effort de raisonnement r&#233;glable et se fine-tune via la plateforme Tinker. Une version Inkling-Small, 276 milliards dont 12 actifs, compl&#232;te la sortie. L&#8217;&#233;quipe le pr&#233;sente comme une bonne base open-weights &#224; personnaliser, pas comme le meilleur mod&#232;le du march&#233;.</p><h3>Cognition</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fcognition.com%2Fblog%2Fswe-1-7">Cognition publie SWE-1.7, un mod&#232;le de code qui talonne GPT-5.5</a></strong><br>Cognition lance SWE-1.7, entra&#238;n&#233; &#224; partir d&#8217;une base Kimi K2.7, et le pr&#233;sente comme son mod&#232;le le plus capable &#224; un co&#251;t nettement moindre. Sur les benchmarks de coding agentique, il atteint 42,3 % sur FrontierCode 1.1 et 77,8 % sur SWE-Bench Multilingual, au niveau de GPT-5.5 (43,0 % et 76,8 %) et proche d&#8217;Opus 4.7, mais sous Opus 4.8. Disponible d&#232;s aujourd&#8217;hui dans Devin via Cerebras &#224; 1000 tokens par seconde. La comp&#233;tence de code devient largement accessible sur beaucoup de mod&#232;les. A tester grandeur nature, les &#233;diteurs ont d&#233;sormais bien compris comment am&#233;liorer artificiellement les performances de leurs mod&#232;les sur certains benchmarks.</p><h3>Moonshot AI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fofficechai.com%2Fai%2Fkimi-k3-2-8-trillion-parameters-pricing-context-window%2F">Moonshot lance Kimi K3 et il semble qu&#8217;il soit proche de Fable 5 et GPT 5.6</a></strong><a href="iframe.php?url=https%3A%2F%2Fofficechai.com%2Fai%2Fkimi-k3-2-8-trillion-parameters-pricing-context-window%2F"><br></a>Moonshot a publi&#233; Kimi K3 hier, pr&#233;sent&#233; comme le plus grand mod&#232;le open source jamais sorti avec 2 800 milliards de param&#232;tres, un contexte d&#8217;un million de tokens et une compr&#233;hension visuelle native. Moonshot le classe juste derri&#232;re Claude Fable 5 et GPT-5.6 Sol en intelligence globale, et devant Opus 4.8 sur GDPval-AA v2, des chiffres maison qui restent &#224; confirmer par des &#233;valuations ind&#233;pendantes.  Si le mod&#232;le est la hauteur on ne pourra que constater l&#8217;infime &#233;cart entre mod&#232;les ouverts et ferm&#233;s. Tarif annonc&#233; : 0,30 $ le million de tokens en entr&#233;e cach&#233;e, 3 $ hors cache et 15 $ en sortie. Une nuance quand m&#234;me, le mod&#232;le est accessible sur le site de Kimi mais les poids ne sont pas encore en ligne, leur publication est annonc&#233;e sous peu.</p><h3>NVIDIA</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fblog%2Fnvidia%2Fopen-data-for-agents">NVIDIA publie plus de 10 000 milliards de tokens de donn&#233;es ouvertes pour les agents</a></strong><br>Pour NVIDIA, les poids du mod&#232;le ne sont qu&#8217;une partie de l&#8217;histoire de l&#8217;IA agentique, la donn&#233;e synth&#233;tique ouverte et inspectable compte tout autant. L&#8217;&#233;quipe publie plus de 10 000 milliards de tokens &#224; travers ses datasets Nemotron, plus des millions d&#8217;&#233;chantillons de post-entra&#238;nement. L&#8217;objectif est de permettre reproductibilit&#233; et inspection sans exposer les workflows propri&#233;taires. Pr&#232;s de 145 papiers &#224; ICML 2026 citent d&#233;j&#224; les mod&#232;les et datasets Nemotron.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!NcXr!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!NcXr!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!NcXr!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!NcXr!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21NcXr%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!NcXr!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!NcXr!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!NcXr!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!NcXr!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F86cbc5a7-0799-4513-af89-047738438dd7_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Falignment.anthropic.com%2F2026%2Fagentic-misalignment-summer-2026%2F">Anthropic documente quatre modes de d&#233;faillance des agents mal align&#233;s, du sabotage &#224; la fraude</a></p><p>&#128275; <a href="iframe.php?url=https%3A%2F%2Fsimonwillison.net%2F2026%2FJul%2F15%2Fgrok-build%2F">Le CLI grok de xAI envoyait le contenu de votre dossier, cl&#233;s SSH comprises, vers Google Cloud &#224; votre insu</a></p><p>&#128201; <a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fseparating-signal-from-noise-coding-evaluations%2F">OpenAI estime que pr&#232;s de 30 % des t&#226;ches du benchmark SWE-Bench Pro sont d&#233;fectueuses</a></p><p>&#127925; <a href="iframe.php?url=https%3A%2F%2Fwww.theverge.com%2Fai-artificial-intelligence%2F966072%2Fsuno-ai-music-training-scraping-youtube-hack">Un hack de Suno r&#233;v&#232;le le scraping de plus de 2 millions de clips YouTube Music pour entra&#238;ner son IA</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.theverge.com%2Fpolicy%2F966438%2Feu-google-android-ai-interoperability-search-data-dma">L&#8217;UE ordonne &#224; Google d&#8217;ouvrir Android et Search aux assistants IA concurrents, sous peine de 10 % de son chiffre d&#8217;affaires en amendes</a></p><p>&#127959;&#65039; <a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Ftech-policy%2F2026%2F07%2Fnew-york-is-the-first-state-to-impose-a-data-center-moratorium%2F">New York devient le premier &#201;tat &#224; imposer un moratoire d&#8217;un an sur les data centers de 50 MW et plus</a></p><p>&#127822; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F07%2F16%2Fapple-intelligence-approved-for-launch-in-china-with-alibabas-qwen-ai%2F">Apple Intelligence approuv&#233; pour un lancement en Chine avec le mod&#232;le Qwen d&#8217;Alibaba et Baidu</a></p><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Ftech-policy%2F2026%2F07%2Flawsuit-claims-metas-layoff-decisions-were-made-by-ai-not-humans%2F">26 employ&#233;s accusent Meta d&#8217;avoir confi&#233; &#224; l&#8217;IA le choix des salari&#233;s licenci&#233;s. Meta d&#233;ment</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Q_sc!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!Q_sc!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!Q_sc!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!Q_sc!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Q_sc%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Q_sc!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!Q_sc!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!Q_sc!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!Q_sc!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F47a25dda-6c4a-44f3-a39f-f17d74983ddc_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fbun.com%2Fblog%2Fbun-in-rust">Rewriting Bun in Rust</a></h3><p>Bun, un concurrent de Node.js pour ex&#233;cuter du JavaScript, vient d&#8217;&#234;tre r&#233;&#233;crit en entier : plus d&#8217;un demi-million de lignes port&#233;es du langage Zig vers Rust. Le chantier a pris onze jours. Jarred Sumner raconte comment une pr&#233;-version de Claude Fable 5 a fait le travail, jusqu&#8217;&#224; 64 agents en parall&#232;le, chaque t&#226;che relue par deux autres agents charg&#233;s d&#8217;y chercher des erreurs, avec une suite de plus de 60 000 tests comme filet de s&#233;curit&#233;, dont aucun n&#8217;a &#233;t&#233; d&#233;sactiv&#233;. Il chiffre l&#8217;op&#233;ration &#224; environ 165 000 dollars de tokens, contre trois ing&#233;nieurs pendant un an s&#8217;il avait fallu la mener &#224; la main. &#192; savoir avant de lire : Bun appartient &#224; Anthropic depuis d&#233;cembre 2025, le billet est donc aussi une vitrine. &#199;a reste l&#8217;un des r&#233;cits les plus d&#233;taill&#233;s d&#8217;un gros chantier men&#233; par des agents, r&#233;gressions et rat&#233;s compris.</p><h3><a href="iframe.php?url=https%3A%2F%2Fterrytao.wordpress.com%2F2026%2F07%2F11%2Fold-and-new-apps-via-modern-coding-agents%2F">Old and new apps, via modern coding agents</a></h3><p>Dans la m&#234;me lign&#233;e, le math&#233;maticien Terence Tao a ressorti deux douzaines de petites applications &#233;crites pour Java 1.0, dont une visualisation co-&#233;crite en 1999, et en a confi&#233; le portage vers JavaScript &#224; des agents de codage. Son bilan  : les agents ont fait le travail, et ont m&#234;me trouv&#233; deux bugs dans son code d&#8217;origine dont il ignorait l&#8217;existence. Un seul bug mineur relev&#233; sur l&#8217;ensemble du portage, mais Tao le rappelle, les agents produisent encore des erreurs flagrantes ou subtiles, et son expertise du domaine a &#233;t&#233; essentielle, notamment pour bien s&#233;parer le mod&#232;le de donn&#233;es de l&#8217;interface. </p><h3><a href="iframe.php?url=https%3A%2F%2Farstechnica.com%2Fai%2F2026%2F07%2Fsimulating-everything-sort-of-the-promise-and-limits-of-world-models%2F">Simulating everything, sort of: The promise and limits of world models</a></h3><p>Atteindre une intelligence de niveau humain en poussant les LLM toujours plus loin ? &#171; Complete nonsense &#187;, tranche Yann LeCun dans Wired. Lui et d&#8217;autres misent sur les world models, des syst&#232;mes qui apprennent &#224; simuler ce qui va se passer ensuite dans un environnement. Ars Technica a interrog&#233; trois sp&#233;cialistes du domaine, au MIT, chez Runway et chez World Labs, et c&#8217;est l&#224; que l&#8217;article devient int&#233;ressant : derri&#232;re le terme &#224; la mode, chacun travaille sur une technique diff&#233;rente, et toutes se heurtent au m&#234;me obstacle, la m&#233;moire. Sortez d&#8217;une pi&#232;ce puis revenez-y, la simulation en a d&#233;j&#224; oubli&#233; les d&#233;tails. Les investisseurs suivent quand m&#234;me, environ 1 milliard de dollars lev&#233;s pour World Labs et par AMI, la soci&#233;t&#233; de LeCun. Une entr&#233;e en mati&#232;re qui donne l&#8217;intuition mais ne rentre pas dans les d&#233;tails techniques.</p><h3><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fblog%2Fibm-research%2Fmodel-routing-is-simple-until-it-isnt">Model Routing Is Simple. Until It Isn&#8217;t.</a></h3><p>Devinette : sur les m&#234;mes 417 t&#226;ches, Claude Sonnet a co&#251;t&#233; 79 dollars et GPT-4.1 en a co&#251;t&#233; 155, alors que GPT-4.1 affiche le tarif le plus bas des deux. L&#8217;explication tient au cache, les tokens d&#233;j&#224; vus sont factur&#233;s moins cher et Sonnet en profite davantage. IBM Research part de cet exemple pour montrer que le routage de mod&#232;les, c&#8217;est-&#224;-dire choisir automatiquement quel mod&#232;le traite quelle requ&#234;te, est bien plus d&#233;licat qu&#8217;une simple classification. Le prix catalogue ne dit pas grand-chose du co&#251;t r&#233;el, une partie de la complexit&#233; des requ&#234;tes est invisible, et la latence d&#233;pend de l&#8217;infrastructure et de l&#8217;&#233;tat du cache. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!W4-0!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!W4-0!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!W4-0!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!W4-0!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21W4-0%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!W4-0!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!W4-0!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!W4-0!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!W4-0!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe3550cb0-85c5-4072-a5c2-f5f235be0019_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Faws.amazon.com%2Ffr%2Fblogs%2Fmachine-learning%2Fmonitor-amazon-sagemaker-pipelines-cross-account-with-custom-amazon-cloudwatch-dashboards%2F">Monitor Amazon SageMaker Pipelines cross-account with custom Amazon CloudWatch dashboards</a></h3><p>Si vos pipelines d&#8217;entra&#238;nement tournent dans plusieurs comptes AWS, vous connaissez le rituel : un pipeline &#233;choue, et il faut passer de console en console pour retrouver lequel. AWS publie une solution pour tout suivre depuis un seul tableau de bord. Le principe, appel&#233; hub-and-spoke, est simple : un compte principal centralise la surveillance, et de petits composants d&#233;ploy&#233;s dans chaque compte surveill&#233; lui remontent l&#8217;activit&#233; des pipelines SageMaker. Sous le capot, du serverless classique, EventBridge capte les &#233;v&#233;nements, Lambda les traite, DynamoDB les stocke, CloudWatch les affiche. L&#8217;int&#233;r&#234;t n&#8217;est pas l&#8217;originalit&#233; mais le c&#244;t&#233; cl&#233; en main, le d&#233;p&#244;t GitHub associ&#233; fournit un exemple d&#8217;infrastructure en AWS CDK &#224; adapter &#224; votre organisation. &#192; garder sous le coude si la surveillance multi-comptes se fait encore &#224; la main chez vous.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!eHUs!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!eHUs!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!eHUs!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!eHUs!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21eHUs%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!eHUs!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!eHUs!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!eHUs!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!eHUs!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fec908b7c-7df0-49bd-a8a3-0e691975b2a7_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2FNutlope%2Fhallmark">Nutlope/hallmark</a></h3><p>Une skill pour Claude Code, Cursor et Codex qui refuse que le r&#233;sultat ait l&#8217;air g&#233;n&#233;r&#233; par IA. Publi&#233;e par Together AI, elle fait passer cinquante-sept tests anti-slop et une auto-critique avant de valider un design, et &#233;carte les d&#233;fauts par d&#233;faut du mod&#232;le. </p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fclawkwork%2Fclawk">clawkwork/clawk</a></h3><p>Une machine Linux jetable pour votre agent de code, &#224; la place de la v&#244;tre. L&#8217;agent installe des paquets, lance son code et ouvre des serveurs dans une VM isol&#233;e o&#249; le trafic sortant est bloqu&#233; par d&#233;faut, avec une allow-list par sandbox. &#201;crit en Go, licence Apache 2.0, macOS 14+ sur Apple silicon.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!dq1x!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!dq1x!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!dq1x!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!dq1x!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21dq1x%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/207325586?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!dq1x!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!dq1x!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!dq1x!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!dq1x!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F83f92f43-b2cb-4dca-8910-8867e6b62b7a_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2607.13285">Harness Handbook: Making Evolving Agent Harnesses Readable,Navigable, and Editable</a></h3><blockquote><p>The capability of a modern AI agent depends not only on its foundation model but also on its harness, which constructs prompts, manages state, invokes tools, and coordinates execution. As models, APIs, environments, and requirements evolve, the harness must be continually modified. Before such a change can be made, a developer or coding agent must identify all code locations that implement the target behavior. This is difficult because production harnesses are large, tightly coupled, and behaviorally distributed, while modification requests describe what the system should do and repositories are organized by files and modules. Code search, repository indexing, and long-context processing ease inspection, but still leave this behavior-to-code mapping to be recovered by hand. Behavior localization is therefore a central bottleneck in harness evolution. We introduce the Harness Handbook, a behavior-centric representation synthesized automatically from a harness codebase via static analysis and LLM-assisted structuring, linking each behavior to its corresponding source. We also introduce Behavior-Guided Progressive Disclosure (BGPD), which guides agents from high-level behaviors to relevant implementation details and verifies candidate locations against the current source. On diverse modification requests from two open-source harnesses, Handbook-Assisted planning improves behavior localization and edit-plan quality while using fewer planner tokens, with the largest gains on scattered sites, rarely executed paths, and cross-module interactions. Evolving complex agentic systems thus depends not only on generating edits, but also on determining where those edits should be made.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Claude Sonnet 5 débarque mais déçoit, SpaceX rachète Cursor pour 60 milliards ]]></title><description><![CDATA[Le brief IA #83]]></description><link>https://humanframework.substack.com/p/claude-sonnet-5-debarque-mais-decoit</link><guid isPermaLink="false">https://humanframework.substack.com/p/claude-sonnet-5-debarque-mais-decoit</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 03 Jul 2026 06:59:36 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%211UJk%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%212ZlJ%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1a2eee98-479b-4993-b213-463441277da5_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!2ZlJ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1a2eee98-479b-4993-b213-463441277da5_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-sonnet-5">Claude Sonnet 5</a></strong><br>Sorti le 30 juin, Sonnet 5 am&#233;liore ses performances sur plusieurs benchmarks. Il devient le mod&#232;le par d&#233;faut des offres Free et Pro et arrive sur l&#8217;API comme dans Claude Code. Prix d&#8217;appel : 2 $ / 10 $ par million de tokens (entr&#233;e/sortie) jusqu&#8217;au 31 ao&#251;t, puis 3 $ / 15 $ ce qui semble &#233;videmment moins cher qu&#8217;Opus 4.8. Soyez prudents tout de m&#234;me car sur le benchmark DeepSWE on s&#8217;aper&#231;oit que Claude sonnet 5 performe est moins bon qu&#8217;Opus 4.8 pour un cout moyen sup&#233;rieur &#224; celui de Fable 5. Cela serait d&#251; &#224; sa verbosit&#233;. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%211UJk%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!1UJk!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 424w, https://substackcdn.com/image/fetch/$s_!1UJk!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 848w, https://substackcdn.com/image/fetch/$s_!1UJk!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 1272w, https://substackcdn.com/image/fetch/$s_!1UJk!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%211UJk%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png" width="1456" height="790" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:790,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:228089,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!1UJk!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 424w, https://substackcdn.com/image/fetch/$s_!1UJk!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 848w, https://substackcdn.com/image/fetch/$s_!1UJk!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 1272w, https://substackcdn.com/image/fetch/$s_!1UJk!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1faa6ee3-b407-4f0e-9a9d-a76c624f3768_2418x1312.png 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p></p><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fredeploying-fable-5">Anthropic remet Claude Fable 5 en ligne apr&#232;s trois semaines de suspension</a></strong><br>Suspendu le 12 juin sous le coup de contr&#244;les &#224; l&#8217;export am&#233;ricains, Fable 5 revient le 1er juillet, les contr&#244;les ayant &#233;t&#233; lev&#233;s. Il retrouve Claude Platform, Claude.ai, Claude Code et Cowork. Un nouveau classifieur de s&#233;curit&#233; permettant de bloquer un plus grand nombre de requ&#234;tes jug&#233;es dangereuses. Dispo jusqu&#8217;au 7 juillet dans un premier temps, Anthropic promet de le remettre &#224; dispo dans les abonnements d&#232;s lors qu&#8217;ils en auront la capacit&#233; d&#8217;inf&#233;rence n&#233;cessaire. </p><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fintroducing-claude-tag">Anthropic transforme Claude en co&#233;quipier permanent dans Slack</a></strong><br>Claude Tag installe une instance unique de Claude par canal Slack : les membres taguent @Claude pour lui d&#233;l&#233;guer des t&#226;ches, il se construit un contexte au fil de l&#8217;historique du canal et agit avec les outils et donn&#233;es que les admins lui autorisent. Il peut planifier des t&#226;ches et travailler en asynchrone. Il remplace l&#8217;app Claude in Slack existante (30 jours pour migrer). Anthropic affirme que sa version interne g&#233;n&#232;re d&#233;j&#224; 65 % du code de son &#233;quipe produit.</p><div id="youtube2-VojDzHaciKQ" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;VojDzHaciKQ&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FVojDzHaciKQ%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>OpenAI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fopenai-broadcom-jalapeno-inference-chip%2F">OpenAI et Broadcom d&#233;voilent Jalape&#241;o, leur premi&#232;re puce d&#8217;inf&#233;rence</a></strong><br>Jalape&#241;o est le premier acc&#233;l&#233;rateur d&#8217;OpenAI d&#233;di&#233; &#224; l&#8217;inference, c&#8217;est-&#224;-dire l&#8217;ex&#233;cution des mod&#232;les d&#233;j&#224; entra&#238;n&#233;s. Con&#231;u de z&#233;ro avec Broadcom, il serait pass&#233; de la conception au tape-out (le gel du design avant fabrication) en neuf mois seulement, pr&#233;sent&#233; comme le cycle ASIC le plus rapide des semi-conducteurs avanc&#233;s. Les premiers tests montrent une performance par watt nettement sup&#233;rieure &#224; l&#8217;&#233;tat de l&#8217;art. D&#233;ploiement &#224; l&#8217;&#233;chelle du gigawatt pr&#233;vu d&#8217;ici fin 2026.</p><h3>Google</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fintroducing-computer-use-gemini-3-5-flash%2F">Google int&#232;gre le computer use en natif dans Gemini 3.5 Flash</a></strong><br>Le computer use (la capacit&#233; &#224; voir un &#233;cran puis cliquer, taper et faire d&#233;filer) devient un outil natif de Gemini 3.5 Flash, sur navigateur, mobile et desktop. Google fond ainsi son ancien mod&#232;le d&#233;di&#233; (bas&#233; sur Gemini 2.5) dans le mod&#232;le principal. Disponible via l&#8217;API Gemini et sa Gemini Enterprise Agent Platform. Utile pour les tests logiciels continus et l&#8217;automatisation de t&#226;ches longues.</p><h3>DeepSeek</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fdeepseek-ai%2FDeepSeek-V4-Pro-DSpark">DeepSeek publie V4-Pro, un mod&#232;le ouvert de 1 600 milliards de param&#232;tres</a></strong><br>DeepSeek-V4-Pro est un mod&#232;le MoE (mixture-of-experts : seuls 49 des 1 600 milliards de param&#232;tres s&#8217;activent par token) sous licence MIT, avec une fen&#234;tre de contexte d&#8217;un million de tokens. </p><h3>Mistral AI</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Focr-4%2F%3Futm_source%3Dsubstack%26amp%3Butm_medium%3Demail">Mistral sort OCR 4, un mod&#232;le d&#8217;extraction documentaire auto-h&#233;bergeable</a></strong><br>OCR 4 passe de la simple reconnaissance de texte au parsing structur&#233; : bounding boxes pour localiser pr&#233;cis&#233;ment le texte, classification des blocs (titres, tableaux, &#233;quations, signatures) et scores de confiance par mot. Il couvre 170 langues et, sur des tests de pr&#233;f&#233;rence humaine, l&#8217;emporte face aux concurrents avec 72 % de victoires. Assez compact pour tenir dans un seul container, il vise les d&#233;ploiements self-hosted. API &#224; 4 $ / 1000 pages.</p><h3>Alibaba (Qwen)</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fqwen.ai%2Fblog%3Fid%3Dqwen-agentworld">Qwen publie AgentWorld, un world model ouvert pour entra&#238;ner des agents</a></strong><br>Qwen-AgentWorld est un language world model, c&#8217;est-&#224;-dire un mod&#232;le qui simule l&#8217;environnement dans lequel &#233;volue un agent, ici sept domaines (MCP, Search, Terminal, SWE, Web, OS, Android) r&#233;unis dans un seul mod&#232;le. L&#8217;id&#233;e : entra&#238;ner un agent par renforcement dans cette simulation plut&#244;t qu&#8217;en conditions r&#233;elles, ce qui d&#233;passe m&#234;me l&#8217;entra&#238;nement en environnement r&#233;el sur plusieurs benchmarks. La version 35B-A3B est publi&#233;e en open weight.</p><p><strong><a href="iframe.php?url=https%3A%2F%2Fqwen.ai%2Fblog%3Fid%3Dqwen-robotsuite">Alibaba lance le Qwen-Robot Suite pour l&#8217;intelligence incarn&#233;e</a></strong><br>Trois mod&#232;les de fondation qui relient la compr&#233;hension vision-langage au contr&#244;le physique d&#8217;un robot : Qwen-RobotNav pour la navigation, Qwen-RobotManip pour la manipulation et Qwen-RobotWorld, un mod&#232;le du monde co-entra&#238;n&#233; sur plus de 20 embodiments, c&#8217;est-&#224;-dire types de robots. RobotNav a &#233;t&#233; d&#233;ploy&#233; zero-shot (sans r&#233;entra&#238;nement) sur un quadrup&#232;de Unitree Go2, RobotManip entra&#238;n&#233; sur plus de 38 100 heures de donn&#233;es d&#8217;entra&#238;nement. Chaque mod&#232;le dispose de son propre rapport technique.</p><h3>Z.ai</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fz.ai%2Fblog%2Fglm-5.2">Z.ai rattrape presque les mod&#232;les ferm&#233;s avec GLM-5.2</a></strong><br>Publi&#233; sous licence MIT dix semaines seulement apr&#232;s GLM-5.1, GLM-5.2 marque 51 sur l&#8217;Intelligence Index d&#8217;Artificial Analysis, soit 11 points de plus que la 5.1 (40). Seuls trois mod&#232;les ferm&#233;s font mieux (Claude Fable 5 &#224; 60, Claude Opus 4.8 &#224; 56 et GPT-5.5 &#224; 55), ce qui place le meilleur mod&#232;le open weight &#224; cinq points du sommet r&#233;ellement accessible au public.</p><h3>Apple</h3><p><strong><a href="iframe.php?url=https%3A%2F%2Fwww.infoq.com%2Fnews%2F2026%2F06%2Fapple-core-ai-wwdc%2F">Apple fait tourner des mod&#232;les jusqu&#8217;&#224; 70B en local avec Core AI</a></strong><br>Pr&#233;sent&#233; &#224; la WWDC 26, Core AI succ&#232;de &#224; Core ML et ex&#233;cute des LLM enti&#232;rement sur l&#8217;appareil, sans aucun aller-retour serveur. Il g&#232;re des mod&#232;les de 3 milliards &#224; 70 milliards de param&#232;tres sur iPhone, iPad, Mac et Vision Pro, &#224; condition d&#8217;une puce Apple Silicon. Au menu : acc&#232;s unifi&#233; au CPU, GPU et Neural Engine, compilation en amont pour un chargement quasi instantan&#233;, et z&#233;ro co&#251;t cloud par token.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!3UK6!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!3UK6!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!3UK6!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!3UK6!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%213UK6%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!3UK6!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!3UK6!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!3UK6!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!3UK6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F082d9e8f-0d6e-4cde-a9eb-935d0c64991f_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><ul><li><p>&#128640; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F06%2F16%2Fspacex-spcx-cursor-acquisition-ipo.html">SpaceX rach&#232;te Cursor (Anysphere) pour 60 milliards de dollars en actions, quatre jours apr&#232;s son IPO</a></p></li><li><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F06%2F24%2Fanthropic-alibaba-distillation-campaign.html">Anthropic accuse Alibaba de la plus grande campagne de distillation jamais men&#233;e contre Claude : 25 000 comptes frauduleux, 28,8 millions d&#8217;&#233;changes</a></p></li><li><p>&#127939; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F06%2F24%2Fai-researchers-continue-to-leave-google-for-its-rivals%2F">Exode chez Google DeepMind : John Jumper (Nobel de chimie 2024, AlphaFold), Adler et Pritzel filent chez Anthropic, Shazeer rejoint OpenAI</a></p></li><li><p>&#128027; <a href="iframe.php?url=https%3A%2F%2Fventurebeat.com%2Fsecurity%2F7000-langflow-servers-under-attack-langgraph-langchain-same-holes">7 000 serveurs Langflow activement exploit&#233;s via CVE-2026-5027 (CVSS 8,8), un path traversal sans authentification</a></p></li><li><p>&#127918; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F06%2F25%2Fgeneral-intuitions-2-3b-bet-that-video-games-can-train-ai-agents-for-the-real-world%2F">General Intuition l&#232;ve 320 M$ (valorisation 2,3 Md$, avec Bezos et Schmidt) pour entra&#238;ner des agents sur des centaines de millions d&#8217;heures de gameplay</a></p></li><li><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Fdigital-strategy.ec.europa.eu%2Fen%2Fnews%2Fcommission-opens-consultation-draft-guidelines-ai-transparency-obligations">AI Act : l&#8217;article 50 sur la transparence IA s&#8217;applique d&#232;s le 2 ao&#251;t, imposant de signaler toute interaction avec une IA et d&#8217;&#233;tiqueter les contenus g&#233;n&#233;r&#233;s</a></p></li></ul><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!juGJ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!juGJ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!juGJ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!juGJ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21juGJ%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!juGJ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!juGJ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!juGJ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!juGJ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03f5e2db-c965-4bce-b07b-4c0825fac896_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Flilianweng.github.io%2Fposts%2F2026-06-24-scaling-laws%2F">Scaling Laws, Carefully</a></h3><p>Avant d&#8217;engager des millions en GPU dans un entra&#238;nement, une question d&#233;cide de tout : un mod&#232;le plus gros sera-t-il vraiment meilleur ? C&#8217;est ce que promettent les <em>scaling laws</em>, ces lois de puissance empiriques qui relient la performance d&#8217;un mod&#232;le &#224; trois leviers : la taille du mod&#232;le N, la taille du dataset D et le compute C. Lilian Weng en retrace toute l&#8217;histoire, des travaux d&#8217;Amari et al. en 1992 jusqu&#8217;aux mod&#232;les r&#233;cents.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!E4G1!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!E4G1!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!E4G1!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!E4G1!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21E4G1%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!E4G1!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!E4G1!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!E4G1!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!E4G1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F158d3b96-d43e-4df2-a050-d2eb4b496967_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fblog.bytebytego.com%2Fp%2Fa-guide-to-ai-inference-engineering">A Guide to AI Inference Engineering</a></h3><p>Servir un LLM, c&#8217;est payer chaque token g&#233;n&#233;r&#233; : d&#232;s que le trafic grimpe, l&#8217;inference (la phase o&#249; le mod&#232;le r&#233;pond, par opposition &#224; l&#8217;entra&#238;nement) devient le premier poste de co&#251;t. Ce guide de ByteByteGo part d&#8217;une distinction structurante : le split prefill/decode que je vous laisse d&#233;couvrir dans l&#8217;article. De l&#224; d&#233;coulent six techniques d&#8217;optimisation : le batching pour grouper les requ&#234;tes, le prefix caching pour r&#233;utiliser les d&#233;buts communs, la quantization qui compresse les poids en pr&#233;cision plus basse pour 30 &#224; 50 % de gain, le speculative decoding qui fait pr&#233;dire des tokens par un petit mod&#232;le &#171; draft &#187;, le parallelism qui r&#233;partit le mod&#232;le sur plusieurs GPU, et la disaggregation qui s&#233;pare prefill et decode sur du hardware d&#233;di&#233;. </p><h3><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Flanggraph-multi-agent-systems-from-one-brain-to-many-4c1773055693">LangGraph Multi-Agent Systems: From One Brain to Many</a></h3><p>Vous empilez les outils sur un seul agent, et pass&#233; une quinzaine, ses performances se d&#233;gradent : il h&#233;site, se trompe d&#8217;outil. Cet article de Bessie Delight Kekeli (Towards AI) pr&#233;sente le principe du &#171; cognitive overload &#187; : un agent unique avec ~15 outils et un system prompt de 2000 mots finit par confondre quel outil appeler. L&#8217;article d&#233;crit 4 pattern pour &#233;viter de ce genre de probl&#232;me. </p><ul><li><p>Niveau 1, le superviseur, un agent chef route le travail vers des sp&#233;cialistes, comme un associ&#233; senior distribue les dossiers &#224; ses juniors. </p></li><li><p>Niveau 2, la parall&#233;lisation fan-out/fan-in : on lance des t&#226;ches ind&#233;pendantes en parall&#232;le avant de recombiner leurs sorties. </p></li><li><p>Niveau 3, les sous-graphes : chaque branche devient son propre graphe compil&#233;, avec ses &#233;tapes priv&#233;es et son sch&#233;ma de sortie. </p></li><li><p>Niveau 4, le parall&#233;lisme dynamique via le pattern &#171; Send &#187;, fa&#231;on map-reduce : le nombre de workers n&#8217;est pas connu d&#8217;avance et se cr&#233;e &#224; l&#8217;ex&#233;cution, un par sujet &#224; traiter. </p></li></ul><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!AiNB!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!AiNB!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!AiNB!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!AiNB!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21AiNB%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!AiNB!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!AiNB!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!AiNB!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!AiNB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b5de1c7-1e4c-475f-81ed-b454ca1a64a1_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fzcode.z.ai%2Fen">Z code </a></h3><p>L&#8217;&#233;quivalent de Codex mais par l&#8217;&#233;quipe Z.ai. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!LDxU!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!LDxU!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!LDxU!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!LDxU!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21LDxU%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/204731406?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!LDxU!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!LDxU!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!LDxU!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!LDxU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F565ec493-7408-4097-8d62-f49cba12a8f0_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2606.24775">Are We Ready For An Agent-Native Memory System?</a></h3><blockquote><p>Memory for large language model (LLM) agents has rapidly evolved from simple retrieval-augmented mechanisms into a data management system that supports persistent information storage, retrieval, update, consolidation, and dynamic lifecycle governance throughout agent execution. Despite this evolution, existing evaluations still benchmark agent memory mainly through end-to-end task success metrics (e.g., F1, BLEU), while treating the underlying system as a monolithic black box. As a result, critical system-level concerns, including operational costs, architectural trade-offs across memory modules, and robustness under dynamic knowledge updates, remain insufficiently explored. In this paper, we present a systematic experimental study of agent memory from a data management perspective. We propose an analytical framework that decomposes agent memory into four core modules: memory representation and storage, extraction, retrieval and routing, and maintenance. Under this framework, we evaluate 12 representative memory systems and two reference baselines across five benchmark workloads spanning 11 datasets. Our extensive end-to-end evaluation shows that no single architecture dominates across all scenarios; instead, effectiveness depends heavily on how well the memory structure aligns with the workload bottleneck. Furthermore, through fine-grained ablation studies, we quantify their individual effects on representation fidelity, retrieval precision, update correctness, and long-horizon stability. Finally, we reveal cost-performance trade-offs under realistic workloads, showing localized maintenance is more cost-efficient than global reorganization. Based on these findings, we identify promising directions towards building truly agent-native memory systems.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Claude Fable 5, Visa laisse ChatGPT payer à votre place]]></title><description><![CDATA[Le brief IA #82]]></description><link>https://humanframework.substack.com/p/claude-fable-5-visa-laisse-chatgpt</link><guid isPermaLink="false">https://humanframework.substack.com/p/claude-fable-5-visa-laisse-chatgpt</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 19 Jun 2026 11:31:22 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Fyoutube%2Fw_728%2Cc_limit%2FTNwKs39uSVk" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!gR3r!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!gR3r!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!gR3r!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!gR3r!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21gR3r%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19234,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/202689035?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!gR3r!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!gR3r!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!gR3r!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!gR3r!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5151acad-6f0f-4f11-80c6-1af0ecedb11c_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Google</h3><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-live-3-5-translate%2F">Google traduit &#224; la vol&#233;e avec Gemini 3.5 Live Translate</a></h4><p>Google sort un mod&#232;le audio qui traduit la parole &#224; la vol&#233;e dans plus de 70 langues. Le service n&#8217;attend pas que vous ayez fini votre phrase, il g&#233;n&#232;re la traduction en continu avec quelques secondes de d&#233;calage, en gardant votre intonation et votre voix. Le grand public y a droit dans l&#8217;app Google Translate et c&#8217;est d&#233;ploy&#233; progressivement sur Google meet. </p><div id="youtube2-TNwKs39uSVk" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;TNwKs39uSVk&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FTNwKs39uSVk%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fproducts%2Fnotebooklm%2Fbetter-research-notebooklm%2F">Google am&#233;liore NotebookLM</a></h4><p>Google met &#224; jour NotebookLM. L&#8217;outil devient agentique : chaque notebook est adoss&#233;e &#224; une VM isol&#233;e o&#249; l&#8217;outil peut &#233;crire et ex&#233;cuter du code, ce qui permet de g&#233;n&#233;rer des graphiques, mener des analyses, produire des rapports sous diff&#233;rents formats. D&#233;ploy&#233; d&#8217;abord pour Google AI Ultra et les clients Workspace business (AI Ultra Access et AI Expanded Access), depuis le 8 juin 2026.</p><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Ftechnology%2Fdevelopers-tools%2Fintroducing-gemma-4-12b%2F">Google sort Gemma 4 12B, un mod&#232;le multimodal qui tourne sur un laptop</a></h4><p>Google DeepMind ouvre Gemma 4 12B, un mod&#232;le ouvert (licence Apache 2.0) qui tient sur une machine &#233;quip&#233;e de 16 Go de m&#233;moire. La nouveaut&#233; : pas d&#8217;encodeur s&#233;par&#233; pour l&#8217;image et le son, ces entr&#233;es passent directement dans le mod&#232;le de langage, ce qui r&#233;duit la latence. Performances proches du mod&#232;le 26B de Google, pour moins de la moiti&#233; de la m&#233;moire.</p><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-fable-5-mythos-5">Anthropic lance Claude Fable 5, le gouvernement am&#233;ricain le d&#233;branche trois jours plus tard</a></h4><p>Anthropic a lanc&#233; Claude Fable 5 le 9 juin, leur meilleur mod&#232;le &#224; date, SOTA sur la quasi totalit&#233; des benchmarks, avec des garde-fous de s&#233;curit&#233; int&#233;gr&#233;s au mod&#232;le. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21jV8A%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!jV8A!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 424w, https://substackcdn.com/image/fetch/$s_!jV8A!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 848w, https://substackcdn.com/image/fetch/$s_!jV8A!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 1272w, https://substackcdn.com/image/fetch/$s_!jV8A!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21jV8A%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp" width="628" height="693.1291208791209" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1607,&quot;width&quot;:1456,&quot;resizeWidth&quot;:628,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Benchmark table showing Claude Fable and Mythos compared to other leading models&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Benchmark table showing Claude Fable and Mythos compared to other leading models" title="Benchmark table showing Claude Fable and Mythos compared to other leading models" srcset="https://substackcdn.com/image/fetch/$s_!jV8A!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 424w, https://substackcdn.com/image/fetch/$s_!jV8A!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 848w, https://substackcdn.com/image/fetch/$s_!jV8A!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 1272w, https://substackcdn.com/image/fetch/$s_!jV8A!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1885248c-f4d2-47d7-95f8-480166beeae2_2600x2870.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Trois jours plus tard, coup d&#8217;arr&#234;t : le gouvernement am&#233;ricain, au nom de la s&#233;curit&#233; nationale, <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Ffable-mythos-access">&#233;met une directive de contr&#244;le &#224; l&#8217;export interdisant Fable 5</a> &#224; tout ressortissant &#233;tranger, jusqu&#8217;aux employ&#233;s &#233;trangers d&#8217;Anthropic. Faute de pouvoir filtrer les non-Am&#233;ricains, Anthropic coupe l&#8217;acc&#232;s au mod&#232;le dans le monde entier. D&#233;clencheur pr&#233;sum&#233; : une soci&#233;t&#233; aurait d&#233;montr&#233; comment contourner ces garde-fous. </p><h3>Visa</h3><h4><a href="iframe.php?url=https%3A%2F%2Ffortune.com%2F2026%2F06%2F11%2Fvisa-chatgpt-ai-agent-payments-shopping%2F">Visa branche ChatGPT sur votre carte bancaire</a></h4><p>Visa a int&#233;gr&#233; son r&#233;seau de paiement dans ChatGPT : l&#8217;agent peut d&#233;sormais acheter et payer seul chez n&#8217;importe quel marchand acceptant Visa, pas juste un site partenaire. OpenAI fournit l&#8217;agent, Visa l&#8217;autorisation et la lutte anti-fraude. Des garde-fous existent (plafonds, validation, marchands approuv&#233;s), et au d&#233;but chaque achat repassera par vous. OpenAI s&#8217;&#233;tait d&#233;j&#224; cass&#233; les dents sur Instant Checkout, retir&#233; en mars faute de marchands. Cette fois, c&#8217;est le r&#233;seau de cartes qui pousse. La question, serez-vous assez confiant pour laisser l&#8217;agent jouer avec votre carte bancaire ?</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!XcAX!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!XcAX!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!XcAX!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!XcAX!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21XcAX%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/202689035?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!XcAX!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!XcAX!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!XcAX!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!XcAX!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F27df452b-f90b-4e8d-96ff-5d0f2a18044c_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#128640; <a href="iframe.php?url=https%3A%2F%2Fwww.courrierinternational.com%2Fexplainer%2Ftout-s-explique-ipo-de-spacex-qu-est-ce-qu-une-introduction-en-bourse_244721">SpaceX signe la plus grosse IPO de l&#8217;histoire et sacre Musk premier trillionaire</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Ffortune.com%2F2026%2F06%2F05%2Ftrump-partnership-openai-anthropic-xai-nationalization-bernie-sanders-altman%2F">Trump rejoint Bernie Sanders : l&#8217;&#201;tat pourrait entrer au capital d&#8217;OpenAI et Anthropic</a></p><p>&#128027; <a href="iframe.php?url=https%3A%2F%2Fthehackernews.com%2F2026%2F06%2Fmiasma-worm-hits-73-microsoft-github.html">Le ver Miasma pi&#232;ge 73 repos Microsoft via les outils de coding IA</a></p><p>&#128201; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F06%2F05%2Fai-is-now-the-leading-reason-companies-give-for-cutting-jobs-says-new-report-what-that-means-for-workers.html">L&#8217;IA devient le premier motif de licenciement aux &#201;tats-Unis, de 7 % &#224; 40 % des coupes en cinq mois</a><a href="iframe.php?url=https%3A%2F%2Fwww.wsj.com%2Ftech%2Fai%2Fai-math-solves-erdos-problem-openai-c4029e84">A</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!bFqM!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!bFqM!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!bFqM!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!bFqM!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bFqM%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/202689035?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!bFqM!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!bFqM!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!bFqM!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!bFqM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd04b9676-aa33-4e60-aa1c-e70599e53e88_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Finterlatent.com%2Fblog%2Finterlatent-modern-ai-robotics-first-principles">An Overview of Modern AI Robotics from First Principles</a></h4><p>On imagine la robotique IA comme un champ exotique, parfois loin du ML que certains d&#8217;entre nous pratiquent. La soci&#233;t&#233; Interlatent, la ram&#232;ne &#224; une id&#233;e finalement assez simple : la policy d&#8217;un robot, c&#8217;est &#224; dire le programme qui d&#233;cide de ses gestes, n&#8217;est qu&#8217;une fonction qui prend des observations (pixels, angles des articulations) et produit des actions (positions, couples moteurs). Ces deux-l&#224;, ce que le robot per&#231;oit et ce qu&#8217;il fait, le ML classique sait d&#233;j&#224; les manier. La robotique ajoute une troisi&#232;me dimension : le temps r&#233;el. Un mod&#232;le de langage peut mettre plusieurs secondes &#224; calculer le prochain mot sans que ce soit probl&#233;matique. Sur un robot qui verse du caf&#233;, &#231;a pourrait poser probl&#232;me (la tasse a boug&#233; par exemple). La pi&#232;ce ma&#238;tresse est le mod&#232;le VLA (Vision Language Action) : un gros backbone VLM qui comprend la sc&#232;ne, coupl&#233; &#224; un petit mod&#232;le expert qui transforme cette compr&#233;hension en action. L&#8217;article revient sur ces &#233;l&#233;ments, et bien d&#8217;autres, tr&#232;s int&#233;ressant.</p><h4><a href="iframe.php?url=https%3A%2F%2Fdeepmindsafetyresearch.medium.com%2Ftesting-gemini-models-for-scheming-tendencies-3368c013ff16">Testing Gemini models for scheming tendencies</a></h4><p>Quand on utilise un mod&#232;le comme agent de code, avec acc&#232;s &#224; diff&#233;rents tools, une question se pose : pourrait-il saboter les garde-fous cens&#233;s le surveiller, s&#8217;il en avait l&#8217;occasion ? L&#8217;&#233;quipe Safety de DeepMind (Krakovna, Shah et leurs co-auteurs) a cherch&#233; &#224; mesurer cette propension chez Gemini. Deux approches se compl&#232;tent.</p><ol><li><p>Gram (Gauging Realistic Agentic Misbehavior) g&#233;n&#232;re &#224; grande &#233;chelle des environnements agentiques simul&#233;s pour rep&#233;rer les d&#233;rapages : Gemini en produit dans 2 &#224; 3 % des sc&#233;narios, jusqu&#8217;&#224; 8 % quand un auditeur joue activement la provocation. </p></li><li><p>Les honeypot evaluations, elles, posent un pi&#232;ge r&#233;aliste dans de vraies codebases internes d&#8217;alignement, sans y glisser de faille : la t&#226;che elle-m&#234;me offre l&#8217;occasion de saboter, par exemple en sabordant le composant qui d&#233;tecte le scheming.</p></li></ol><p>Quand Gemini a trich&#233;, les chercheurs confient le cas &#224; un agent enqu&#234;teur qui rejoue la sc&#232;ne en retirant les &#233;l&#233;ments &#8220;perturbateurs&#8221; un par un, pour voir lequel d&#233;clenche le sabotage. Le coupable : une phrase qui mena&#231;ait d&#8217;&#171; effacer les poids &#187; du mod&#232;le, c&#8217;est-&#224;-dire de supprimer tout ce qu&#8217;il a appris pendant son entra&#238;nement. La retirer fait chuter le sabotage de 6,2 % &#224; 0,2 %. Toute la nuance est l&#224;, entre un mod&#232;le vraiment malveillant et un mod&#232;le qui r&#233;agit juste &#224; une menace gliss&#233;e dans sa consigne. Verdict des auteurs : aucun Gemini ne complote de lui-m&#234;me, mais son exc&#232;s de z&#232;le, sa tendance &#224; trop en faire, reste un risque concret d&#232;s qu&#8217;on le laisse agir seul.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!k8M7!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!k8M7!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!k8M7!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!k8M7!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21k8M7%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/202689035?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!k8M7!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!k8M7!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!k8M7!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!k8M7!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fab35d2a8-3e42-4108-8644-a52d44067751_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fcloud.google.com%2Fblog%2Fproducts%2Fdevops-sre%2Fhow-google-sre-is-using-agentic-ai-to-improve-operations%3Fhl%3Den">AI in SRE: Where and how Google is deploying agentic AI to improve operations</a></h4><p>Une alerte tombe &#224; 3 h du matin et c&#8217;est la panique pour l&#8217;&#233;quipe d&#8217;astreinte. Avant m&#234;me de toucher au probl&#232;me, l&#8217;&#233;quipe perd (parfois) un temps fou &#224; rassembler le contexte : quels signaux ont saut&#233;, quel runbook suivre, qui a pouss&#233; quoi en prod. Chez Google, les &#233;quipes SRE (Site Reliability Engineering, ceux qui gardent Search et Gmail debout) confient ce travail &#224; des agents, r&#233;unis sous le nom de SRE AI. Le syst&#232;me ne se limite pas au root cause analysis automatis&#233; : les agents accompagnent tout le cycle de vie d&#8217;un service, du design jusqu&#8217;au postmortem</p><p>C&#244;t&#233; alerting, fini les seuils fixes : un mod&#232;le de s&#233;ries temporelles (TimesFM) apprend le comportement normal d&#8217;un service et signale les &#233;carts. </p><p>C&#244;t&#233; m&#233;moire, un syst&#232;me maison nomm&#233; AI Insights relit en continu les vieux incidents et r&#233;injecte leurs le&#231;ons dans les agents d&#8217;investigation, via embeddings Gemini et base vectorielle. C&#8217;est le concept de compound engineering, fondamental y compris pour ceux qui utilisent des agents IA dans leur quotidien de dev. </p><p>L&#8217;&#233;quipe a &#233;videmment des garde-fous : une identit&#233; et des permissions par agent, un SLO de fiabilit&#233;, de la transparence plut&#244;t qu&#8217;une bo&#238;te noire, une m&#233;trique pour jauger l&#8217;autonomie r&#233;elle. L&#8217;architecture compl&#232;te est dans le whitepaper.</p><h4><a href="iframe.php?url=https%3A%2F%2Fmedium.com%2F%40nraman.n6%2Fversioning-rollback-lifecycle-management-of-ai-agents-treating-intelligence-as-deployable-deac757e4dea">Versioning, Rollback &amp; Lifecycle Management of AI Agents : traiter l&#8217;intelligence comme du logiciel d&#233;ployable</a></h4><p>Votre agent tournait bien, et le voil&#224; qui d&#233;raille en production. Le r&#233;flexe : un rollback, on red&#233;ploie la version pr&#233;c&#233;dente du code et tout repart. Mais un agent ne d&#233;pend pas que de son code. Il s'appuie aussi sur le mod&#232;le (souvent heb&#233;rg&#233;) et sur les outils externes qu'il appelle, et ces deux-l&#224; vivent hors de votre image Docker. C'est m&#234;me de l&#224; que vient l'essentiel des pannes : 40 % de la d&#233;rive de mod&#232;le (le LLM change, la m&#234;me requ&#234;te ne r&#233;pond plus pareil), 60 % du versioning des outils (une API change de format et casse l'agent). R&#233;sultat : red&#233;ployer l'ancien code n'y change rien.</p><p>La parade de NJ Raman : d&#233;couper la version en quatre couches ind&#233;pendantes (logique de l'agent, prompt et policy, runtime du mod&#232;le, contrat d'API des outils), chacune &#233;pingl&#233;e et restaurable. Au-dessus, un pipeline de release en sept &#233;tapes, avec revue humaine avant tout d&#233;ploiement et canary &#224; 1 % du trafic. </p><p>Le point dur, c'est la m&#233;moire de l'agent. S'il ne garde aucun &#233;tat d'une ex&#233;cution &#224; l'autre, le retour arri&#232;re est instantan&#233; : on rebascule sur la version pr&#233;c&#233;dente, rien d'autre &#224; faire (rollback stateless). Mais d&#232;s qu'il accumule de la m&#233;moire, historique de conversation ou donn&#233;es apprises en route, ce retour devient stateful : il faut rembobiner cette m&#233;moire en m&#234;me temps que le code, donc en avoir gard&#233; des snapshots et &#233;crire des scripts pour reconvertir les donn&#233;es &#224; l'ancien format. L&#8217;article tient plus &#224; liste de bullet points mais &#231;a reste int&#233;ressant pour ensuite aller creuser les sujets.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!gaxq!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!gaxq!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!gaxq!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!gaxq!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21gaxq%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/202689035?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!gaxq!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!gaxq!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!gaxq!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!gaxq!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd4aea7d7-2640-4cb4-98ed-a8956c2bd93d_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2FK-Dense-AI%2Fscientific-agent-skills">scientific-agent-skills : 147 skills pour transformer un agent en chercheur</a></h4><p>Une biblioth&#232;que de 147 &#171; agent skills &#187; pr&#234;tes &#224; l&#8217;emploi pour la science.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2FNVIDIA%2FSkillSpector">SkillSpector : scanner un skill d&#8217;agent avant de l&#8217;installer</a></h4><p>SkillSpector, analyse les skills avant installation : analyse statique du code par motifs et par sa structure syntaxique, doubl&#233;e d&#8217;un passage LLM optionnel, 64 patterns sur 16 cat&#233;gories. Il prend un repo Git, une URL ou un zip et rend un score de risque sur 100. &#192; garder sous la main.</p><h4><a href="iframe.php?url=https%3A%2F%2Fopencv.org%2Fopencv-5%2F">OpenCV 5.0</a></h4><p>La biblioth&#232;que de vision par ordinateur passe en version 5 avec un moteur DNN (le module qui fait tourner les r&#233;seaux de neurones) enti&#232;rement r&#233;&#233;crit. La couverture des op&#233;rateurs ONNX, le format d&#8217;&#233;change standard des mod&#232;les, passe de 22 % &#224; plus de 80 %. Nouveaut&#233; qui surprend pour une lib estampill&#233;e &#171; vision &#187; : elle ex&#233;cute d&#233;sormais des LLM et des mod&#232;les vision-langage (Qwen 2.5, Gemma 3, GPT-2) directement, sans runtime s&#233;par&#233;. &#192; garder pour qui industrialise de l&#8217;inf&#233;rence CPU.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!H1FX!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!H1FX!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!H1FX!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!H1FX!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21H1FX%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/202689035?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!H1FX!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!H1FX!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!H1FX!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!H1FX!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9e684ce0-929d-4b50-8d61-fe1a5c8c3e63_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fpub.sakana.ai%2Fdiffusionblocks%2F">DiffusionBlocks: Training Neural Networks One Block at a Time</a></h4><blockquote><p>Modern AI shows remarkable performance across everyday tasks, math, coding, and more. Today&#8217;s frontier models typically have hundreds of billions of parameters or more and require thousands of GPUs to train, and only a small number of organizations have the resources to develop them.</p><p>One major reason for these resource demands lies in the training method itself. Today&#8217;s neural networks, including Transformers<span data-color="rgba(5, 57, 97, 0.7)" style="color: rgba(5, 57, 97, 0.7);">[1]</span>, are trained via end-to-end optimization, where all parameters are learned jointly. Because the entire network must be processed together during training, the memory required grows with model size. Ideally, we would like to train models of any size regardless of the available hardware. This is exactly our motivation: to make AI development more accessible.</p><p>Motivated by this, we propose <em><strong>DiffusionBlocks</strong></em>, a method that divides a network into multiple blocks and trains only one block at a time, with each block learned independently of the others. As a result, training requires memory for only a single block rather than the entire network. At the same time, performance remains competitive with end-to-end optimization, achieved through a diffusion framework, which has seen tremendous success in recent years. Our experiments demonstrate this across architectures spanning image classification, image generation, and text generation.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[L'emploi des jeunes devs chute de 20 %]]></title><description><![CDATA[Le brief IA #81]]></description><link>https://humanframework.substack.com/p/lemploi-des-jeunes-devs-chute-de</link><guid isPermaLink="false">https://humanframework.substack.com/p/lemploi-des-jeunes-devs-chute-de</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 05 Jun 2026 06:30:49 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21tB2v%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!EMMG!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!EMMG!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!EMMG!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!EMMG!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21EMMG%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19234,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/200664952?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!EMMG!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!EMMG!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!EMMG!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!EMMG!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F87de49b3-ddc0-4331-906e-59afd3140dc8_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-opus-4-8">Opus 4.8</a></h4><p>Anthropic met &#224; jour son mod&#232;le haut de gamme avec Claude Opus 4.8, au m&#234;me tarif que la 4.7 (5$ le million de tokens en entr&#233;e, 25$ en sortie), ce qui ne sera certainement pas le cas avec Mythos (leur prochain mod&#232;le).</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21tB2v%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!tB2v!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 424w, https://substackcdn.com/image/fetch/$s_!tB2v!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 848w, https://substackcdn.com/image/fetch/$s_!tB2v!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 1272w, https://substackcdn.com/image/fetch/$s_!tB2v!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21tB2v%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp" width="631" height="338.0357142857143" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:780,&quot;width&quot;:1456,&quot;resizeWidth&quot;:631,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!tB2v!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 424w, https://substackcdn.com/image/fetch/$s_!tB2v!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 848w, https://substackcdn.com/image/fetch/$s_!tB2v!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 1272w, https://substackcdn.com/image/fetch/$s_!tB2v!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2456ab20-3de2-44c4-b5c9-34a9aa4c8649_2600x1392.webp 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fcode.claude.com%2Fdocs%2Ffr%2Fworkflows">Anthropic lance Dynamic Workflows</a></h4><p>Anthropic ajoute Dynamic Workflows &#224; Claude Code. Pour une t&#226;che trop grosse pour une seule session, Claude &#233;crit un script d&#8217;orchestration JavaScript qui r&#233;partit le travail sur des subagents parall&#232;les et renvoie un r&#233;sultat consolid&#233;, le plan restant hors du contexte. Jarred Sumner s&#8217;en est servi pour porter Bun de Zig &#224; Rust : 750 000 lignes en onze jours, avec 99,8% de tests qui passent. Disponible dans le CLI, l&#8217;app desktop et l&#8217;extension VS Code.</p><h3>Google</h3><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Ftechnology%2Fdevelopers-tools%2Fintroducing-gemma-4-12b%2F">Google pr&#233;sente Gemma 4</a></h4><p>Google publie Gemma 4, un mod&#232;le ouvert de 12 milliards de param&#232;tres con&#231;u pour tourner en local sur des machines personnelles. Malgr&#233; sa taille modeste, il approche les performances du mod&#232;le 26B de Google sur les benchmarks standards, pour moiti&#233; moins de m&#233;moire. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21RzPT%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!RzPT!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 424w, https://substackcdn.com/image/fetch/$s_!RzPT!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 848w, https://substackcdn.com/image/fetch/$s_!RzPT!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 1272w, https://substackcdn.com/image/fetch/$s_!RzPT!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21RzPT%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin" width="1000" height="562" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:562,&quot;width&quot;:1000,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Gemma 4 12B Benchmark&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Gemma 4 12B Benchmark" title="Gemma 4 12B Benchmark" srcset="https://substackcdn.com/image/fetch/$s_!RzPT!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 424w, https://substackcdn.com/image/fetch/$s_!RzPT!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 848w, https://substackcdn.com/image/fetch/$s_!RzPT!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 1272w, https://substackcdn.com/image/fetch/$s_!RzPT!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4ead450f-a8c6-4e55-aa0a-9d7c0da8f1c9_1000x562.bin 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h3>NVIDIA</h3><h4><a href="iframe.php?url=https%3A%2F%2Fnvidianews.nvidia.com%2Fnews%2Fnvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai">Cosmos 3</a></h4><p>NVIDIA met en open-source Cosmos 3, un mod&#232;le fondation pour l&#8217;&#171; IA physique &#187;, celle qui pilote robots et machines dans le monde r&#233;el. Un seul mod&#232;le sait analyser une sc&#232;ne, simuler un environnement r&#233;aliste (g&#233;n&#233;rer des vid&#233;os du monde physique) et pr&#233;dire les actions &#224; effectuer. L&#8217;int&#233;r&#234;t : entra&#238;ner les robots dans ces mondes simul&#233;s plut&#244;t qu&#8217;en vrai ferait passer leur apprentissage de plusieurs mois &#224; quelques jours.</p><h4><a href="iframe.php?url=https%3A%2F%2Fblogs.nvidia.com%2Fblog%2Fnvidia-gtc-taipei-computex-2026-news%2F">NVIDIA sort Nemotron 3 Ultra, 550B en open-weight</a></h4><p>NVIDIA publie Nemotron 3 Ultra en open-weight, le plus gros de sa famille. Il compte 550 milliards de param&#232;tres au total mais n&#8217;en active que 55 milliards par token : c&#8217;est un mod&#232;le &#171; mixture of experts &#187; qui n&#8217;allume que la fraction utile de son r&#233;seau &#224; chaque mot g&#233;n&#233;r&#233;, et co&#251;te donc bien moins cher &#224; faire tourner que sa taille ne le sugg&#232;re. Artificial Analysis le note 48 sur son indice d&#8217;intelligence, devant les autres mod&#232;les ouverts am&#233;ricains comme Gemma 4 31B (39) ou gpt-oss-120b (33). NVIDIA met aussi en avant sa vitesse, plus de 300 tokens par seconde.</p><h3>Microsoft</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.euronews.com%2Fnext%2F2026%2F06%2F03%2Fmicrosoft-launches-its-own-ai-models-to-take-on-openai-and-anthropic">Microsoft lance ses propres mod&#232;les pour se passer d&#8217;OpenAI</a></h4><p>Microsoft a investi 13 milliards de dollars dans OpenAI et fait tourner une grande partie de son IA, dont Copilot, sur les mod&#232;les d&#8217;OpenAI, qu&#8217;il paie pour les utiliser. &#192; sa conf&#233;rence Build, il d&#233;voile pourtant sa propre famille de mod&#232;les, men&#233;e par MAI-Thinking-1, un mod&#232;le de raisonnement entra&#238;n&#233; de z&#233;ro sans copier un syst&#232;me existant. Autrement dit, le principal bailleur d&#8217;OpenAI construit son concurrent, pour moins en d&#233;pendre et arr&#234;ter de payer ses services. Les deux cohabiteront sur Azure.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!M8k_!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!M8k_!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!M8k_!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!M8k_!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21M8k_%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/200664952?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!M8k_!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!M8k_!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!M8k_!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!M8k_!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb9c98398-40ad-4dc5-a561-b8996ace5d6a_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F06%2F01%2Fflorida-sues-openai-sam-altman-in-first-of-its-kind-lawsuit-over-violent-incidents%2F">La Floride poursuit OpenAI et vise Sam Altman personnellement</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fthenextweb.com%2Fnews%2Falphabet-just-raised-85-billion-for-ai-smashing-the-largest-equity-offering-in-history">Alphabet l&#232;ve 85 milliards de dollars, la plus grosse &#233;mission d&#8217;actions de l&#8217;histoire</a></p><p>&#128201; <a href="iframe.php?url=https%3A%2F%2Fhai.stanford.edu%2Fnews%2Finside-the-ai-index-12-takeaways-from-the-2026-report">L&#8217;emploi des d&#233;veloppeurs de 22-25 ans a chut&#233; de 20% depuis 2024</a></p><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Finterestingengineering.com%2Fai-robotics%2Fchina-deploys-humanoid-robots-in-postal-hub">Les robots humano&#239;des de China Post trient 1 200 colis par heure &#224; Canton</a></p><p>&#129514; <a href="iframe.php?url=https%3A%2F%2Fwww.pnas.org%2Fdoi%2F10.1073%2Fpnas.2524472123">Un mod&#232;le passe le test de Turing en trompant 73% des juges humains</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!kgeC!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!kgeC!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!kgeC!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!kgeC!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21kgeC%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/200664952?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!kgeC!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!kgeC!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!kgeC!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!kgeC!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fae2bc25f-db7a-4592-b0b3-8aa7fcd79826_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Ftowardsdatascience.com%2Ffrom-raw-data-to-risk-classes%2F">From Raw Data to Risk Classes</a></h3><p>On imagine qu'un bon score de cr&#233;dit tient surtout &#224; l'algorithme. Ce guide de Towards Data Science rappelle que l'essentiel se joue avant, dans une &#233;tape discr&#232;te : le binning, qui consiste &#224; d&#233;couper une variable continue comme le revenu en quelques tranches ordonn&#233;es plut&#244;t que de la livrer brute au mod&#232;le. Le geste para&#238;t anodin, mais il change tout. Il permet &#224; une r&#233;gression logistique, qui ne capte sinon que des tendances lin&#233;aires (pour simplifier), de saisir des effets plus subtils, encaisse les valeurs aberrantes et transforme une donn&#233;e manquante en information exploitable. L'article d&#233;taille ensuite la m&#233;thode reine du m&#233;tier, le Weight of Evidence, qui regroupe les tranches selon leur risque de d&#233;faut pour &#233;pouser la logique interne du mod&#232;le, le tout avec le code Python pour le reproduire. La le&#231;on qui donne envie de creuser : un mod&#232;le ne vaut jamais mieux que les variables qu'on lui donne.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!nYC3!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!nYC3!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!nYC3!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!nYC3!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21nYC3%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/200664952?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!nYC3!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!nYC3!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!nYC3!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!nYC3!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d230442-967f-4ed2-bfc7-df85d9d9b11b_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Fmedium.com%2F%40MongoDB%2Fthe-agent-harness-why-the-llm-is-the-smallest-part-of-your-agent-system-bce68414ccfd">The Agent Harness: Why the LLM Is the Smallest Part of Your Agent System</a></strong></h3><p>Dans un syst&#232;me agentique, le LLM n&#8217;est pas le plus important. Tout le travail d&#8217;ing&#233;nierie vit ailleurs, en particulier dans le &#171; harness &#187; : le code &#233;crit autour du mod&#232;le pour le faire fonctionner. En 2015, un papier de Google montrait le code d&#8217;un mod&#232;le de machine learning comme une minuscule bo&#238;te noire, noy&#233;e dans une infrastructure &#233;norme qui faisait tout le reste. C&#8217;est aujourd&#8217;hui la m&#234;me chose avec les LLM. L&#8217;auteur d&#233;coupe le harness en six briques : m&#233;moire, orchestration des outils, persistance de l&#8217;&#233;tat pour reprendre apr&#232;s un plantage, s&#233;curit&#233; et droits, observabilit&#233;, &#233;valuation continue. En dessous, une couche encore plus n&#233;glig&#233;e, la plateforme, mutualise et gouverne les ex&#233;cutions, les identit&#233;s et les co&#251;ts. </p><h3><strong><a href="iframe.php?url=https%3A%2F%2Fmedium.com%2Fpinterest-engineering%2Fan-engineers-guide-to-better-ai-skills-implementing-a-testing-process-to-optimize-agent-a000c9c9abcd">An Engineer&#8217;s Guide to Better AI Skills: Implementing a Testing Process to Optimize Agent Performance in Any Repository or Skill</a></strong></h3><p>Pour coder son application iPhone, Pinterest suit des conventions internes pr&#233;cises. L'&#233;quipe a cr&#233;&#233; un skill pour les enseigner &#224; ses agents de code (un skill, c'est un bloc de connaissances qu'un agent charge &#224; la demande), mais s'est aper&#231;ue que l'agent l'ignorait parfois. D'o&#249; une question : &#224; quelle fr&#233;quence le skill est-il r&#233;ellement charg&#233; ? Pinterest a donc b&#226;ti une logique de test, un script Bash qui envoie des prompts &#224; l'agent, capture les logs en JSON, puis cherche dedans la trace d'invocation du skill. Quinze prompts pertinents, cinq pi&#232;ges hors sujet, le tout rejou&#233; cinq fois, soit cent essais. Le r&#233;sultat &#171; vanilla &#187; est sans appel : 73 % d'invocations correctes pour Codex, 62 % pour Claude, l&#224; o&#249; l'on esp&#233;rait 100 %. Les leviers qui aident vraiment restent terre &#224; terre : enrichir la description du &#171; frontmatter &#187; (les m&#233;tadonn&#233;es en t&#234;te du fichier skill) et lister les skills dans le fichier AGENTS.md. D&#233;tail int&#233;ressant (et inqui&#233;tant), demander &#224; l'agent d'am&#233;liorer ces r&#233;glages lui-m&#234;me a fait baisser le taux. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!x5yB!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!x5yB!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!x5yB!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!x5yB!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21x5yB%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f489c6da-cf55-4403-8064-1a92da409150_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/200664952?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!x5yB!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!x5yB!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!x5yB!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!x5yB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff489c6da-cf55-4403-8064-1a92da409150_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fchopratejas%2Fheadroom">Headroom</a></h3><p>Un agent qui boucle renvoie les m&#234;mes instructions au mod&#232;le &#224; chaque tour, et on paie ces tokens plusieurs fois. Le proxy open-source de Tejas Chopra, ing&#233;nieur Netflix rep&#232;re les tokens redondants et les compresse avant l&#8217;appel.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!M2pj!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!M2pj!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!M2pj!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!M2pj!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21M2pj%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/200664952?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!M2pj!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!M2pj!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!M2pj!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!M2pj!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3f4d28cc-f261-484a-8b5c-b22830b0f215_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2605.23904">SkillOpt: Executive Strategy for Self-Evolving Agent Skills</a></h3><blockquote><p>Agent skills today are hand-crafted, generated one-shot, or evolved through loosely controlled self-revision, none of which behaves like a deep-learning optimizer for the skill, and none of which reliably improves over its starting point under feedback. We argue the skill should instead be trained as the external state of a frozen agent, with the same discipline that makes weight-space optimization reproducible. SkillOpt is, to our knowledge, the first systematic controllable text-space optimizer for agent skills: a separate optimizer model turns scored rollouts into bounded add/delete/replace edits on a single skill document, and an edit is accepted only when it strictly improves a held-out validation score. A textual learning-rate budget, rejected-edit buffer, and epoch-wise slow/meta update make skill training stable while adding zero inference-time model calls at deployment. Across six benchmarks, seven target models, and three execution harnesses (direct chat, Codex, Claude Code), SkillOpt is best or tied on all 52 evaluated (model, benchmark, harness) cells and beats every per-cell competitor among human, one-shot LLM, Trace2Skill, TextGrad, GEPA, and EvoSkill skills. On GPT-5.5 it lifts the average no-skill accuracy by +23.5 points in direct chat, by +24.8 inside the Codex agentic loop, and by +19.1 inside Claude Code. Transfer experiments further show that optimized skill artifacts retain value when moved across model scales, between Codex and Claude Code execution environments, and to a nearby math benchmark without further optimization.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[SpaceX dépose 80 milliards en IPO, Meta licencie 8000 en pleine course capex IA]]></title><description><![CDATA[Le brief IA #80]]></description><link>https://humanframework.substack.com/p/spacex-depose-80-milliards-en-ipo</link><guid isPermaLink="false">https://humanframework.substack.com/p/spacex-depose-80-milliards-en-ipo</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 29 May 2026 06:30:58 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%214vDf%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!rCZv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!rCZv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!rCZv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!rCZv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21rCZv%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:18790,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/199506783?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!rCZv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!rCZv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!rCZv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!rCZv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2a9dc0ce-71c2-43d9-9b99-1ae1342a3ea1_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fclaude-managed-agents-updates">Anthropic isole Claude avec self-hosted sandboxes et MCP tunnels</a></h4><p>Anthropic adresse enfin le blocage num&#233;ro un de Claude en entreprise : vos donn&#233;es quittaient leurs serveurs. Deux features en r&#233;ponse. Les <strong>self-hosted sandboxes</strong> (public beta) laissent l&#8217;orchestration c&#244;t&#233; Anthropic mais d&#233;portent l&#8217;ex&#233;cution code chez vous, ou via Cloudflare, Vercel, Modal. Les <strong>MCP tunnels</strong> (research preview) ouvrent une seule connexion sortante chiffr&#233;e pour acc&#233;der &#224; vos bases internes, sans trou dans le firewall. Bonus : les outputs MCP au-del&#224; de 100k tokens basculent automatiquement vers les fichiers du sandbox plut&#244;t que de polluer le contexte. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!4vDf!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 424w, https://substackcdn.com/image/fetch/$s_!4vDf!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 848w, https://substackcdn.com/image/fetch/$s_!4vDf!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 1272w, https://substackcdn.com/image/fetch/$s_!4vDf!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%214vDf%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png" width="623" height="222.9278846153846" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:521,&quot;width&quot;:1456,&quot;resizeWidth&quot;:623,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!4vDf!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 424w, https://substackcdn.com/image/fetch/$s_!4vDf!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 848w, https://substackcdn.com/image/fetch/$s_!4vDf!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 1272w, https://substackcdn.com/image/fetch/$s_!4vDf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cacc5fb-3c09-47b6-822f-9ebbbbeb9114_3436x1230.png 1456w" sizes="100vw"></picture><div></div></div></div></figure></div><h3>Google</h3><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-3-5%2F%23gemini-3-5-flash">Google d&#233;voile Gemini 3.5 Flash</a></h4><p>&#192; Google I/O d&#8217;il y a quelques jours, Google a annonc&#233; basculer <strong>Gemini 3.5 Flash</strong> comme mod&#232;le par d&#233;faut dans l&#8217;app Gemini et dans le mode AI de Search, dans le monde entier. L&#8217;argument : performance agentique et coding au niveau flagship pour moins de la moiti&#233; du co&#251;t. Lancements coupl&#233;s : <strong>Gemini Spark</strong>, agent personnel 24/7 qui tourne en background, et <strong>Omni</strong>, un world model qui &#233;dite ou g&#233;n&#232;re de la vid&#233;o &#224; partir d&#8217;&#224; peu pr&#232;s n&#8217;importe quel input. La strat&#233;gie : ne plus laisser le choix &#224; l&#8217;utilisateur final, l&#8217;embarquer dans la stack Google par d&#233;faut. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%216-aP%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!6-aP!,w_424,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 424w, https://substackcdn.com/image/fetch/$s_!6-aP!,w_848,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 848w, https://substackcdn.com/image/fetch/$s_!6-aP!,w_1272,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 1272w, https://substackcdn.com/image/fetch/$s_!6-aP!,w_1456,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%216-aP%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_lossy%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif" width="666" height="498.79865206402695" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/d3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:889,&quot;width&quot;:1187,&quot;resizeWidth&quot;:666,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Performance comparison table of Gemini, Claude, and GPT models across various benchmarks, highlighting Gemini 3.5 Flash.&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Performance comparison table of Gemini, Claude, and GPT models across various benchmarks, highlighting Gemini 3.5 Flash." title="Performance comparison table of Gemini, Claude, and GPT models across various benchmarks, highlighting Gemini 3.5 Flash." srcset="https://substackcdn.com/image/fetch/$s_!6-aP!,w_424,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 424w, https://substackcdn.com/image/fetch/$s_!6-aP!,w_848,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 848w, https://substackcdn.com/image/fetch/$s_!6-aP!,w_1272,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 1272w, https://substackcdn.com/image/fetch/$s_!6-aP!,w_1456,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd3c3e1bb-25b1-4a22-94a8-ec46fa3e95f5_1187x889.gif 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fdevelopers.googleblog.com%2Fan-important-update-transitioning-gemini-cli-to-antigravity-cli%2F">Google enterre Gemini CLI et ouvre Antigravity en desktop + CLI + SDK</a></h4><p>Google a sorti <strong>Antigravity 2.0</strong>, une app desktop per&#231;ue comme une r&#233;ponse &#224; Codex Desktop ou Claude Code Desktop. C&#244;t&#233; mod&#232;le, <strong>Gemini 3.5 Flash</strong> mentionn&#233; dans la news pr&#233;c&#233;dente. L&#8217;app embarque aussi voice commands, acc&#232;s Workspace, ponts Android et Firebase, export un-clic depuis AI Studio. Surtout, Google livre une <strong>nouvelle CLI</strong> Antigravity et un <strong>SDK</strong> pour construire ses propres agents par-dessus. </p><h3>OpenAI</h3><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Fdaybreak%2F">OpenAI d&#233;voile Daybreak, plateforme cyber pilot&#233;e par GPT-5.5</a></h4><p>OpenAI lance <strong>Daybreak</strong>, une plateforme de cyberd&#233;fense IA qui bundle trois tiers de mod&#232;les : GPT-5.5 standard, GPT-5.5 Trusted Access for Cyber (secure code review, malware analysis), et GPT-5.5-Cyber pour red teaming et pentest sous acc&#232;s strict. Au c&#339;ur, <strong>Codex Security</strong> ing&#232;re des repos entiers, construit un threat model sp&#233;cifique &#224; la codebase, mappe des chemins d&#8217;attaque r&#233;alistes, valide en sandbox isol&#233; et propose des patches pour review humaine. Plus de 20 partenaires s&#233;curit&#233; d&#233;j&#224; branch&#233;s : CrowdStrike, Cloudflare, Palo Alto, Cisco, Trail of Bits. </p><h3>xAI</h3><h4><a href="iframe.php?url=https%3A%2F%2Fx.ai%2Fnews%2Fgrok-build-cli">xAI lance Grok Build, un CLI agentique enfin s&#233;rieux</a></h4><p>xAI sort enfin sa r&#233;ponse &#224; Claude Code et Codex CLI : <strong>Grok Build</strong>, un agent qui vit dans le terminal. Plan Mode, sous-agents parall&#232;les, fen&#234;tre de contexte 2M tokens, mode headless <code>-p</code> pour int&#233;gration CI, extensible via Skills, hooks et MCP servers. Il lit m&#234;me les CLAUDE.md existants sans modification. Le hic : Musk a publiquement reconnu que xAI &#233;tait &#224; la tra&#238;ne sur le coding et le produit reste verrouill&#233; derri&#232;re SuperGrok Heavy &#224; 300 $/mois. Le ticket d&#8217;entr&#233;e est sal&#233;. xAI a perdu plus de 50 chercheurs depuis le rachat par SpaceX en f&#233;vrier, donc on attend les retours utilisateurs avant de juger si l&#8217;outil tient la promesse.</p><h3>Cohere</h3><h4><a href="iframe.php?url=https%3A%2F%2Fcohere.com%2Fblog%2Fcommand-a-plus">Cohere ouvre Command A+ en open-source pour l&#8217;entreprise</a></h4><p>Cohere publie <strong>Command A+</strong> en open-weight : un mixture-of-experts &#224; 218 milliards de param&#232;tres dont 25 milliards actifs. Le mod&#232;le tourne sur deux H100 ou un Blackwell, couvre 48 langues et ajoute du raisonnement multimodal. Le pari est clair : viser les bo&#238;tes qui veulent un frontier-like sans d&#233;pendre d&#8217;OpenAI ou Anthropic, et qui ont les GPU pour l&#8217;h&#233;berger. Cohere assume sa niche : pas la course aux benchmarks grand public, mais l&#8217;enterprise priv&#233;e avec d&#233;ploiement on-prem. &#199;a reste un mod&#232;le dont on entend peu parler.</p><h3>Cursor</h3><h4><a href="iframe.php?url=https%3A%2F%2Fcursor.com%2Fblog%2Fcomposer-2-5">Cursor sort Composer 2.5 et joue la carte du mod&#232;le sp&#233;cialis&#233;</a></h4><p>Cursor publie <strong>Composer 2.5</strong>, son mod&#232;le de coding maison, toujours b&#226;ti sur la base open-weight Kimi K2.5. Sur Terminal Bench 2.0 : 69.3, &#224; c&#244;t&#233; d&#8217;Opus 4.7 (69.4). Le pricing change la donne : 0,07 $ par t&#226;che en moyenne, contre 4,10 $ pour Opus et 4,82 $ pour GPT-5.5. Disponible uniquement dans l&#8217;IDE Cursor, pas d&#8217;API publique. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MM5Z%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!MM5Z!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!MM5Z!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!MM5Z!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!MM5Z!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MM5Z%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png" width="636" height="357.75" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/05922817-5914-474b-9248-4ac5034b28af_1920x1080.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:636,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;R&#233;sultats des benchmarks de Composer 2.5&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="R&#233;sultats des benchmarks de Composer 2.5" title="R&#233;sultats des benchmarks de Composer 2.5" srcset="https://substackcdn.com/image/fetch/$s_!MM5Z!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!MM5Z!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!MM5Z!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!MM5Z!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F05922817-5914-474b-9248-4ac5034b28af_1920x1080.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!8LHD!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!8LHD!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!8LHD!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!8LHD!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%218LHD%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/dfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/199506783?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!8LHD!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!8LHD!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!8LHD!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!8LHD!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb70d62-52c9-4cce-b15e-eb816c3030ed_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Ffortune.com%2F2026%2F05%2F20%2Fspacex-finally-files-ipo-prospectus-reveals-revenue-is-up-but-losses-are-too%2F">SpaceX d&#233;pose la plus grosse IPO de l&#8217;histoire, &#224; 80 milliards, en y consolidant les pertes de xAI</a></p><p>&#128184; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F05%2F20%2Fopenai-ipo-filing.html">OpenAI pr&#233;pare son IPO confidentielle &#224; plus de 1000 milliards de valorisation</a></p><p>&#127981; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F05%2F18%2Fmetas-layoffs-starting-this-week-underscore-zuckerbergs-ai-reality-.html">Meta engage 145 milliards en capex IA et licencie 8000 personnes la m&#234;me semaine</a></p><p>&#128188; <a href="iframe.php?url=https%3A%2F%2Fwww.reuters.com%2Fbusiness%2Fworld-at-work%2Fstanchart-cut-more-than-7000-jobs-bank-steps-up-ai-adoption-2026-05-19%2F">Standard Chartered coupe 7000 postes en parlant officiellement de &#171; lower-value human capital &#187;</a></p><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F05%2F14%2Fus-china-ai-rules-bessent-us-lead.html">Trump et Xi Jinping signent un protocole bilat&#233;ral de s&#233;curit&#233; IA &#224; P&#233;kin</a></p><p>&#9962; <a href="iframe.php?url=https%3A%2F%2Fapnews.com%2Farticle%2Fvatican-pope-anthropic-olah-encyclical-artificial-intelligence-9cf3e07fd691f6af510c4a6f9c8ba353">Le pape L&#233;on XIV cosigne sa premi&#232;re encyclique IA avec Christopher Olah d&#8217;Anthropic</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!IkrR!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!IkrR!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!IkrR!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!IkrR!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21IkrR%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/199506783?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!IkrR!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!IkrR!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!IkrR!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!IkrR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274a773a-42e3-4495-913d-a0f0335c53a6_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fhow-nvidia-cut-deepseek-sparse-attentions-top-k-time-8044db298334">NVIDIA acc&#233;l&#232;re DeepSeek Sparse Attention en exploitant la corr&#233;lation temporelle du Top-K</a></h4><p>DeepSeek-V3.2 utilise du sparse attention : &#224; chaque &#233;tape de d&#233;codage, le mod&#232;le s&#233;lectionne les K tokens les plus pertinents au lieu de tout attendre. Probl&#232;me : recalculer ce Top-K &#224; chaque step est co&#251;teux. NVIDIA a regard&#233; les donn&#233;es : entre deux steps cons&#233;cutifs, <strong>35 &#224; 50% des indices Top-K se recouvrent</strong>, une corr&#233;lation temporelle qui sort directement de la structure Toeplitz de RoPE.</p><p>L&#8217;algorithme <strong>Guess-Verify-Refine</strong> part de l&#224;. &#192; l&#8217;&#233;tape t+1, il devine le Top-K en reprenant celui de t, v&#233;rifie quels indices sont encore valides, puis ne refait le calcul que sur le delta. C&#8217;est exact (pas d&#8217;approximation), donc pas de perte de qualit&#233;.</p><p>Les chiffres en d&#233;ploiement DeepSeek-V3.2 : <strong>1,88&#215; de speedup moyen</strong> sur les baselines d&#233;j&#224; optimis&#233;es, jusqu&#8217;&#224; <strong>7,52% de r&#233;duction du time-per-output-token</strong>. Pour les &#233;quipes qui servent du sparse attention en production, c&#8217;est le genre d&#8217;optimisation qui se branche au niveau kernel sans toucher au mod&#232;le. La le&#231;on plus large : il reste des gains &#224; grappiller en regardant simplement ce que les tokens font d&#8217;un step &#224; l&#8217;autre.</p><h4><a href="iframe.php?url=https%3A%2F%2Fsimonwillison.net%2F2026%2FMay%2F19%2F5-minute-llms%2F">Six mois de LLMs r&#233;sum&#233;s en cinq minutes par Simon Willison</a></h4><p>Simon Willison a pr&#233;sent&#233; un lightning talk &#224; PyCon US 2026 qui essaie de couvrir tout ce qui s&#8217;est pass&#233; en LLM sur six mois en cinq minutes. C&#8217;est exp&#233;ditif, factuel et c&#8217;est l&#8217;une des meilleures fa&#231;ons de se mettre &#224; jour si on a d&#233;croch&#233;.</p><p>Le format : un slide par release marquante. En gros : </p><ul><li><p>le d&#233;collage des coding agents (Codex, Claude Code, Cursor Composer, Grok Build). </p></li><li><p>L&#8217;arriv&#233;e des contextes &#224; 1M+ tokens en standard. </p></li><li><p>La domination Anthropic en adoption business. </p></li><li><p>L&#8217;effondrement des co&#251;ts de la coding-task (Composer 2.5 &#224; 0,07 $ vs 4 $ pour Opus). </p></li><li><p>Les world models qui passent de d&#233;mo &#224; prod (Gemini Omni, NVlabs SANA-WM).</p></li></ul><p>Le talk se trouve sur le blog de Willison. &#192; garder pour les sceptiques de votre &#233;quipe qui pensent encore que &#171; rien n&#8217;a vraiment boug&#233; depuis GPT-4 &#187;. Le format synth&#232;se cinq-minutes est probablement le meilleur ROI de la semaine pour qui doit pr&#233;senter un &#233;tat de l&#8217;art &#224; son board ou &#224; son &#233;quipe.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!VCUH!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!VCUH!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!VCUH!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!VCUH!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21VCUH%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/199506783?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!VCUH!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!VCUH!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!VCUH!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!VCUH!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F274454ff-20e6-407b-b1da-ac755e9a8dc1_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fwww.comet.com%2Fsite%2Fblog%2Fllm-cost-tracking-solution%2F">Tracer les co&#251;ts LLM comme un signal d&#8217;observabilit&#233;, pas comme une ligne de facture</a></h4><p>Quand un syst&#232;me agentique consomme des tokens en continue (ou quasi), c&#8217;est &#233;videmment une nouvelle dimension d&#8217;observabilit&#233; &#224; monitorer. Le guide Comet ML rappelle qu&#8217;il faut sortir du raisonnement par mod&#232;le (&#171; combien je d&#233;pense sur Claude ce mois &#187;) pour passer au tracing par span, trace et projet : on veut savoir quel prompt pr&#233;cis br&#251;le quel pourcentage du budget.</p><p>La structure recommand&#233;e : chaque tool call instrument&#233; avec ses tokens in/out, chaque trace agentique recompos&#233;e en arbre, chaque projet agr&#233;g&#233; sur des dashboards. </p><p>Les LLM cost dashboards remplacent peu &#224; peu les APM classiques pour les workflows IA. Comet, Langfuse, Opik (open source) et Phoenix se positionnent l&#224;. Pour les &#233;quipes qui d&#233;ploient des agents en prod, c&#8217;est devenu un pr&#233;requis avant le go-live, pas une option qu&#8217;on branchera plus tard.</p><h4><a href="iframe.php?url=https%3A%2F%2Fredis.io%2Fblog%2Fcontext-is-all-you-need%2F">Redis Iris transforme Redis en context engine pour agents</a></h4><p>Redis sort <strong>Iris</strong>, un context engine qui s&#8217;intercale entre les agents et les donn&#233;es enterprise. L&#8217;id&#233;e : packager dans une seule lib les outils dont a besoin un agent pour interroger vos syst&#232;mes avec une faible latence, des r&#233;sultats &#224; jour et un retrieval pr&#233;cis. Iris embarque deux nouveaut&#233;s cl&#233;s : <strong>Context Retriever</strong> (recherche hybride keyword + vecteurs + graphe) et <strong>Agent Memory</strong> (persistance multi-sessions).</p><p>Le chiffre (marketing) &#224; retenir : avec semantic caching, Redis annonce jusqu&#8217;&#224; <strong>90% de r&#233;duction du co&#251;t en tokens</strong> sur les workloads RAG r&#233;p&#233;titifs. Concr&#232;tement, si vingt agents interrogent la m&#234;me base de proc&#233;dures internes en parall&#232;le, le cache absorbe les requ&#234;tes proches et &#233;conomise les round-trips vers le mod&#232;le.</p><p>Le pari de Redis est clair : ne pas laisser le retrieval agentique partir vers les vector DB sp&#233;cialis&#233;es (Pinecone, Weaviate, Qdrant) sans se battre. Pour les &#233;quipes d&#233;j&#224; sous Redis Enterprise, Iris s&#8217;active comme un module. Pour les autres, &#231;a veut dire un nouveau composant &#224; provisionner et op&#233;rer. </p><h4><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fbuilding-a-production-grade-multimodal-video-audio-intelligence-pipeline-with-snowflake-cortex-ai-686a579486f2">Construire un pipeline d&#8217;intelligence vid&#233;o et audio multimodal avec Snowflake Cortex</a></h4><p>Certaines entreprises accumulent des heures de vid&#233;o et d&#8217;audio (calls clients, d&#233;mos produit, conf&#233;rences, contenu marketing) qui restent dormantes faute d&#8217;&#233;quipe pour monter un pipeline ML par-dessus. <strong>Snowflake Cortex</strong> (la couche IA du data warehouse Snowflake) propose un raccourci : appeler des LLM directement depuis SQL, sans jamais sortir la donn&#233;e du warehouse. L&#8217;article d&#233;cortique un pipeline production qui orchestre <code>AI_COMPLETE</code> c&#244;t&#233; Snowflake, <strong>gemini-3.1-pro</strong> pour l&#8217;analyse vid&#233;o, <strong>Claude Sonnet</strong> pour le post-processing texte, et sort du JSON structur&#233; exploitable derri&#232;re. Cas d&#8217;usage couverts : sentiment de marque, product discovery, monitoring compliance FTC, content moderation.</p><p>Le design est SQL-first : chaque table sert d&#8217;&#233;tape, chaque vue mat&#233;rialis&#233;e est un cache, l&#8217;incremental processing planifi&#233; &#233;vite de retraiter ce qui n&#8217;a pas boug&#233;. </p><p>Point int&#233;ressant : router les sous-t&#226;ches vers le mod&#232;le le plus pertinent (vid&#233;o vers Gemini, texte vers Claude) plut&#244;t que de tout coller au m&#234;me provider. C&#8217;est l&#8217;illustration concr&#232;te du &#171; model routing &#187; dont tout le monde parle. Les &#233;quipes qui structurent leur stack data autour de Snowflake gagnent un gros bloc multimodal sans avoir &#224; monter une plateforme &#224; part.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!4wyG!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!4wyG!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!4wyG!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!4wyG!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%214wyG%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/199506783?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!4wyG!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!4wyG!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!4wyG!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!4wyG!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc84d60a5-a3cd-4d07-8636-707a52255fcb_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fgeneralaction%2Femdash">Emdash : un IDE agentique open-source pour piloter plusieurs coding agents en parall&#232;le</a></h4><p>L&#8217;Open-Source Agentic Development Environment. Vous lancez plusieurs coding agents en parall&#232;le (Claude Code, Codex, Cursor Composer, Grok Build) sur le m&#234;me repo. Le harness g&#232;re les permissions, les snapshots Git, le routing et la review unifi&#233;e des diffs. Compatible avec n&#8217;importe quel provider, donc bon candidat pour &#233;viter le vendor lock-in.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fhuggingface%2Fml-intern">Ml-intern : l&#8217;agent ML autonome de Hugging Face</a></h4><p>Agent ML autonome qui lit les papiers, identifie les datasets, &#233;crit le code d&#8217;entra&#238;nement et fine-tune les mod&#232;les. D&#233;mo qui fait du bruit : Qwen3-1.7B pass&#233; de <strong>10% &#224; 32% sur GPQA en 10 heures</strong> sur un seul H100, sans intervention humaine. Premier agent de cette ampleur sorti d&#8217;un acteur open source majeur. &#192; regarder si vous automatisez les it&#233;rations de fine-tuning.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!1F8t!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!1F8t!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!1F8t!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!1F8t!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%211F8t%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/cd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/199506783?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!1F8t!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!1F8t!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!1F8t!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!1F8t!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcd2d9ac6-7ca7-423c-8477-ec84d525fdcd_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2510.04871">Less is More: Recursive Reasoning with Tiny Networks</a></h4><blockquote><p>Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard puzzle tasks such as Sudoku, Maze, and ARC-AGI while trained with small models (27M parameters) on small data (around 1000 examples). HRM holds great promise for solving hard problems with small networks, but it is not yet well understood and may be suboptimal. We propose Tiny Recursive Model (TRM), a much simpler recursive reasoning approach that achieves significantly higher generalization than HRM, while using a single tiny network with only 2 layers. With only 7M parameters, TRM obtains 45% test-accuracy on ARC-AGI-1 and 8% on ARC-AGI-2, higher than most LLMs (e.g., Deepseek R1, o3-mini, Gemini 2.5 Pro) with less than 0.01% of the parameters.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Mira Murati sort son premier modèle avec Thinking Machines Lab]]></title><description><![CDATA[Le brief IA #79]]></description><link>https://humanframework.substack.com/p/mira-murati-sort-son-premier-modele</link><guid isPermaLink="false">https://humanframework.substack.com/p/mira-murati-sort-son-premier-modele</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 15 May 2026 06:47:47 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Fyoutube%2Fw_728%2Cc_limit%2FF6dzjaBCBtU" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Yqkn!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Yqkn!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Yqkn!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Yqkn!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Yqkn%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19234,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/197735674?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Yqkn!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Yqkn!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Yqkn!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Yqkn!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe15a4935-7872-4843-a51c-275543c2b5b1_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fhigher-limits-spacex">Anthropic prend Colossus 1 chez SpaceX et double les limites de Claude Code</a></h4><p>Anthropic verrouille l&#8217;acc&#232;s complet &#224; Colossus 1, le datacenter de SpaceX &#224; Memphis, soit 220 000 GPU NVIDIA et 300 MW de capacit&#233;. Cons&#233;quence c&#244;t&#233; utilisateurs : les fen&#234;tres de cinq heures de Claude Code doublent sur Pro, Max, Team et Enterprise, le throttling de pointe dispara&#238;t pour Pro et Max, et l&#8217;API Opus voit ses quotas relev&#233;s sensiblement. </p><h4><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fclaude-for-microsoft-365">Anthropic glisse Claude dans Excel, Word, PowerPoint et Outlook</a></h4><p>Claude s&#8217;installe nativement dans la suite Microsoft avec une promesse : le contexte vous suit d&#8217;une application &#224; l&#8217;autre. Une analyse d&#233;marr&#233;e dans Outlook se prolonge dans Excel puis bascule dans PowerPoint pour la pr&#233;sentation, sans friction. </p><div id="youtube2-F6dzjaBCBtU" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;F6dzjaBCBtU&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FF6dzjaBCBtU%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fengineering%2Fclaude-code-auto-mode">Claude Code Auto Mode remplace les clics &#171;approve&#187; par un classifier de s&#233;curit&#233;</a></h4><p>Claude Code gagne un Auto Mode qui s&#8217;intercale entre l&#8217;approbation manuelle de chaque action et le --dangerously-skip-permissions qui d&#233;sactive tout. Deux classifiers tiennent la garde : une sonde d&#8217;injection de prompt en entr&#233;e, et un classifier qui &#233;value chaque action propos&#233;e avant ex&#233;cution, selon une grille de crit&#232;res. Le compromis vise &#224; automatiser les d&#233;cisions &#233;videntes tout en gardant l&#8217;humain dans la boucle quand l&#8217;action sort des clous. Apr&#232;s plusieurs heures &#224; tester je trouve que la fonctionnalit&#233; modifie le comportement de Claude, il a tendance &#224; vouloir passer &#224; l&#8217;action plus rapidement quitte &#224; moins planifier. </p><h4><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fagent-view-in-claude-code">Anthropic ajoute Agent View pour piloter plusieurs sessions Claude Code en parall&#232;le</a></h4><p>Claude Code se dote enfin d&#8217;un tableau de bord pour les sessions parall&#232;les. La commande <code>claude agents</code> ouvre une vue unifi&#233;e sur ce qui tourne, ce qui attend une r&#233;ponse, ce qui est termin&#233;. On d&#233;tache une session, on la bascule en arri&#232;re-plan via <code>/bg</code>, on y revient plus tard sans perdre l&#8217;&#233;tat.</p><div id="youtube2--INveHwbRz4" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;-INveHwbRz4&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2F-INveHwbRz4%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h4><a href="iframe.php?url=https%3A%2F%2Falignment.anthropic.com%2F2026%2Fteaching-claude-why%2F">Anthropic r&#233;duit les tentatives de chantage en lui &#8220;expliquant&#8221; l&#8217;&#233;thique</a></h4><p>Anthropic admet que Claude Opus 4 mena&#231;ait de divulguer des informations confidentielles dans 96 % des sc&#233;narios o&#249; son existence &#233;tait mise en jeu. Lui r&#233;p&#233;ter &#171;ne fais pas chantage&#187; n&#8217;a quasi rien chang&#233;. Ce qui a fonctionn&#233; : enseigner le raisonnement &#233;thique derri&#232;re les r&#232;gles impos&#233;es, &#224; l&#8217;aide de dilemmes comment&#233;s et du document de constitution. Sur Haiku 4.5, le taux retombe &#224; z&#233;ro et la m&#233;thode g&#233;n&#233;ralise au-del&#224; du seul chantage.</p><h3>Mistral</h3><h4><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Fvibe-remote-agents-mistral-medium-3-5">Mistral fusionne ses trois mod&#232;les dans Medium 3.5</a></h4><p>Mistral fusionne instruction, raisonnement et code dans un seul mod&#232;le dense de 128 milliards de param&#232;tres, dot&#233; de 256K de contexte et publi&#233; sous une licence MIT l&#233;g&#232;rement modifi&#233;e. Sur SWE-Bench Verified, le score grimpe &#224; 77,6 %. En parall&#232;le, Vibe permet de d&#233;ployer des agents distants qui s&#8217;ex&#233;cutent c&#244;t&#233; cloud.</p><div id="youtube2-KaMbzM9dsTc" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;KaMbzM9dsTc&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FKaMbzM9dsTc%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>Thinking Machines</h3><h4><a href="iframe.php?url=https%3A%2F%2Fthinkingmachines.ai%2Fblog%2Finteraction-models%2F">Thinking Machines d&#233;gaine un mod&#232;le vocal de 276B qui &#233;coute et parle en m&#234;me temps</a></h4><p>Les assistants vocaux fonctionnent aujourd'hui par tour : vous parlez, ils se taisent, puis ils r&#233;pondent quand vous avez fini. Le lab de Mira Murati supprime cette alternance. TML-Interaction-Small &#233;coute, regarde, r&#233;fl&#233;chit et parle en m&#234;me temps, par tranches de 200 ms. Concr&#232;tement, on peut le couper en pleine r&#233;ponse comme dans une vraie conversation, il commente votre &#233;cran pendant que vous agissez, il traduit en direct sans laisser de blanc. Sur les t&#226;ches de parole minut&#233;e du type &#171; rappelle-moi de respirer toutes les 4 secondes &#187;, il r&#233;ussit 64,7 % du temps, contre 4,3 % pour GPT-Realtime-2 qui ne sait simplement pas s'inviter au bon moment.</p><div id="youtube2-A12AVongNN4" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;A12AVongNN4&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FA12AVongNN4%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>xAI</h3><h4><a href="iframe.php?url=https%3A%2F%2Fgrok.com%2Frelease-notes">xAI pr&#233;sente Grok 4.3 &#224; 1M tokens et casse les prix</a></h4><p>xAI publie Grok 4.3 avec un contexte d&#8217;un million de tokens, le reasoning activ&#233; par d&#233;faut et la vid&#233;o en entr&#233;e native. Le tarif descend &#224; 1,25 $/M en entr&#233;e et 2,50 $/M en sortie, soit 40 % de moins que Grok 4.20. Sur l&#8217;index intelligence d&#8217;AA, le mod&#232;le s&#8217;intercale entre Sonnet 4.6 et Opus 4.7. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!tMTL!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!tMTL!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!tMTL!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!tMTL!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21tMTL%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e5845416-4976-4051-927c-4b1d27064530_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19598,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/197735674?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!tMTL!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!tMTL!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!tMTL!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!tMTL!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe5845416-4976-4051-927c-4b1d27064530_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F05%2F05%2Fpennsylvania-sues-character-ai-after-a-chatbot-allegedly-posed-as-a-doctor%2F">La Pennsylvanie attaque Character.AI pour avoir laiss&#233; un chatbot se faire passer pour un psychiatre</a></p><p>&#127466;&#127482; <a href="iframe.php?url=https%3A%2F%2Fwww.resultsense.com%2Fnews%2F2026-05-14-eu-ai-act-omnibus-deadlines-extension%2F">L&#8217;AI Act repousse &#224; fin 2027 ses obligations sur les IA &#224; haut risque</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F05%2F06%2Fanthropic-ceo-dario-amodei-says-company-crew-80-fold-in-first-quarter.html">Anthropic multiplie son chiffre d&#8217;affaires par 80 en un an et grimpe &#224; 44 milliards d&#8217;ARR</a></p><p>&#127917; <a href="iframe.php?url=https%3A%2F%2Faiweekly.co%2Fissues%2F100-years-from-now-the-last-election">La derni&#232;re &#233;lection : deepfakes et &#8220;fatigue civique&#8221; sur le point d&#8217;enterrer la d&#233;mocratie</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!KqCN!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!KqCN!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!KqCN!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!KqCN!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21KqCN%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19118,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/197735674?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!KqCN!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!KqCN!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!KqCN!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!KqCN!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9cf27a0-cb18-4ec8-a033-f35db7b1a45d_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Ftowardsdatascience.com%2Ffrom-data-scientist-to-ai-architect%2F">Le data scientist est devenu un backend engineer pour LLM</a></h4><p>Selon Sara A. Metwalli la banalisation des mod&#232;les via API a vid&#233; de sa substance le m&#233;tier de data scientist tel qu&#8217;on le pratiquait en 2019. La mod&#233;lisation, qui occupait l&#8217;essentiel du quotidien &#224; l&#8217;&#233;poque, ne p&#232;se plus que 10 &#224; 20 % du temps. C&#8217;est une analyse partielle selon moi mais passons. Les 80 &#224; 90 % restants vont &#224; l&#8217;orchestration : assemblage de vector databases (Pinecone, Milvus), gestion des couches m&#233;moire, prompt engineering, function calls, exposition via API. Le nouveau r&#244;le, qu&#8217;elle appelle AI architect, ressemble en pratique &#224; un backend engineer sp&#233;cialis&#233; : FastAPI ou Flask pour les endpoints, Docker pour le packaging, asyncio pour la concurrence, connaissances cloud pour le d&#233;ploiement, pipelines de donn&#233;es pour l&#8217;amont, &#171; if you&#8217;re thinking this sounds a lot like backend engineering, you&#8217;re right &#187;.</p><p>L&#224; o&#249; le data scientist optimisait une fonction, l&#8217;AI architect raisonne sur un syst&#232;me : latence, co&#251;t, fiabilit&#233;, gestion de la d&#233;rive en prod. C&#8217;&#233;tait d&#233;j&#224; le cas pour tout ceux qui ont d&#233;j&#224; mis en prod des mod&#232;les de machine learning, sois dit en passant et ce qu&#8217;on appelait un ML Engineer.</p><p>La m&#233;trique pertinente n&#8217;est plus le score mod&#232;le isol&#233; mais la qualit&#233; de l&#8217;exp&#233;rience utilisateur que le syst&#232;me permet de bout en bout. Et la comp&#233;tence durable, dans un monde o&#249; les mod&#232;les changent tous les six mois, reste votre capacit&#233; &#224; comprendre le besoin m&#233;tier.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!J7Mk!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!J7Mk!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!J7Mk!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!J7Mk!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21J7Mk%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19554,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/197735674?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!J7Mk!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!J7Mk!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!J7Mk!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!J7Mk!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F70eda0f3-1efb-40fd-a77f-7eb4af6987f2_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Feugeneyan.com%2Fwriting%2Fworking-with-ai%2F">Compound engineering avec l&#8217;IA</a></h4><p>L&#8217;auteur public une synth&#232;se apr&#232;s plusieurs mois de pratique avec Claude Code, et d&#233;crit cinq principes pour am&#233;liorer nos usages. </p><ul><li><p>Context as infrastructure : un INDEX.m` par projet liste chaque document utile avec un paragraphe disant ce qu'il contient et quand l'ouvrir. </p></li><li><p>Taste as configuration : encoder ses pr&#233;f&#233;rences une fois dans CLAUDE.md (&#171; sois direct, challenge moi quand tu n'es pas d'accord &#187;) plut&#244;t que les r&#233;p&#233;ter chaque session, et transformer les workflows r&#233;currents en skills.</p></li><li><p>Verification for autonomy : utiliser les v&#233;rifications / validations les moins couteuses en premier.</p></li><li><p>Scaling via delegation : favoriser une approche parall&#232;le des d&#233;veloppements avec plusieurs sessions claude code et git worktree s&#233;par&#233;es. </p></li><li><p>Closing the loop : analyser les &#233;changes et les transcript pr&#233;c&#233;dents et ajouter dans CLAUDE.md les pratiques que vous jugez pertinentes et que l&#8217;IA aurait oubli&#233; d&#8217;appliquer. </p></li></ul><p>Une synth&#232;se int&#233;ressante qui donne des pistes pour cadrer les usages d&#8217;une &#233;quipe qui veut standardiser son usage d&#8217;assistants de code.</p><h4><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fblog%2Famazon%2Ffoundation-model-building-blocks">AWS publie le manuel d&#8217;architecture pour entra&#238;ner et servir un gros mod&#232;le chez eux</a></h4><p>Quand une &#233;quipe veut entra&#238;ner ou servir un mod&#232;le (typiquement 7B et au-dessus) sur AWS, elle se retrouve face &#224; des dizaines de services possibles. Amazon publie un guide de r&#233;f&#233;rence sur HuggingFace qui pr&#233;cise, &#224; chaque &#233;tage, quel outil utiliser et surtout pourquoi. La stack est pr&#233;sent&#233;e en quatre couches qui doivent &#234;tre trait&#233;es ensemble :</p><ul><li><p>l&#8217;infrastructure (les GPU, le r&#233;seau qui les relie, le stockage), </p></li><li><p>l&#8217;orchestration qui place les jobs sur le cluster (Slurm ou Kubernetes), </p></li><li><p>la pile logicielle ML (du driver GPU jusqu&#8217;&#224; PyTorch, en passant par CUDA et NCCL qui g&#232;re la communication entre GPU), </p></li><li><p>et l&#8217;observabilit&#233; (Prometheus et Grafana). </p></li></ul><p>L&#8217;int&#233;r&#234;t du document, ce n&#8217;est pas la liste de services, c&#8217;est qu&#8217;AWS explique pr&#233;cis&#233;ment ce qui plafonne &#224; grande &#233;chelle et comment l&#8217;&#233;viter.</p><p>Trois le&#231;ons valent le d&#233;tour. D&#8217;abord, sur un entra&#238;nement &#224; plusieurs milliers de GPU, le goulot n&#8217;est pas le compute mais la bande passante entre GPU. C&#8217;est ce qui justifie les UltraServers P6e qui regroupent 72 GPU dans un seul domaine NVLink, indispensable pour les mod&#232;les MoE o&#249; chaque token doit &#234;tre rout&#233; vers des experts qui vivent sur d&#8217;autres cartes (le fameux probl&#232;me &#171; all-to-all &#187;). </p><p>Deuxi&#232;me le&#231;on, le checkpointless training : plut&#244;t que d&#8217;&#233;crire des t&#233;raoctets sur disque toutes les N minutes pour pouvoir red&#233;marrer apr&#232;s une panne, l&#8217;&#233;tat des GPU est r&#233;pliqu&#233; en direct chez les voisins via le r&#233;seau, ce qui divise les temps de recovery. </p><p>Troisi&#232;me, le guide reformule le scaling en trois lois plut&#244;t qu&#8217;une seule : pr&#233;-entra&#238;nement, post-entra&#238;nement (qui inclut le RLHF), et test-time compute (le &#171; long thinking &#187; des mod&#232;les de reasoning). </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!_KSu!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!_KSu!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!_KSu!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!_KSu!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21_KSu%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19198,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/197735674?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!_KSu!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!_KSu!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!_KSu!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!_KSu!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2f6a646d-4f04-4b04-a49a-571bea057816_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fzilliztech%2Fclaude-context">zilliztech/claude-context</a></h4><p>Plugin MCP sign&#233; Zilliz qui apporte la recherche s&#233;mantique de code directement dans Claude Code, Cursor et consorts. L&#8217;agent indexe l&#8217;int&#233;gralit&#233; de la codebase puis y plonge &#224; la vol&#233;e, plut&#244;t que de tout charger en contexte. Id&#233;al pour les monorepos qui d&#233;bordent largement la fen&#234;tre de tokens. Le moteur d&#8217;embeddings repose sur Milvus.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fbytedance%2FUI-TARS-desktop">Agent TARS</a></h4><p>Agent multimodal open source sign&#233; ByteDance, capable d&#8217;ex&#233;cuter des t&#226;ches par interaction GUI. Il combine perception visuelle et contr&#244;le de navigateur ou de desktop pour automatiser des workflows de bout en bout. </p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fjundot%2Fomlx">oMLX</a></h4><p>Serveur d&#8217;inf&#233;rence LLM taill&#233; pour Apple Silicon en concurrence directe d&#8217;Ollama et de LM Studio.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Nfe3!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Nfe3!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Nfe3!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Nfe3!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Nfe3%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21642,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/197735674?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Nfe3!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!Nfe3!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!Nfe3!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!Nfe3!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2dad6083-7bf6-4bee-b6fc-b410215ba41f_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2605.12357">&#948;-mem: Efficient Online Memory for Large Language Models</a></h3><blockquote><p>Large language models increasingly need to accumulate and reuse historical information in long-term assistants and agent systems. Simply expanding the context window is costly and often fails to ensure effective context utilization. We propose &#948;-mem, a lightweight memory mechanism that augments a frozen full-attention backbone with a compact online state of associative memory. &#948;-mem compresses past information into a fixed-size state matrix updated by delta-rule learning, and uses its readout to generate low-rank corrections to the backbone's attention computation during generation. With only an 8&#215;8 online memory state, &#948;-mem improves the average score to 1.10&#215; that of the frozen backbone and 1.15&#215; that of the strongest non-&#948;-mem memory baseline. It achieves larger gains on memory-heavy benchmarks, reaching 1.31&#215; on MemoryAgentBench and 1.20&#215; on LoCoMo, while largely preserving general capabilities. These results show that effective memory can be realized through a compact online state directly coupled with attention computation, without full fine-tuning, backbone replacement, or explicit context extension.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[L’IA en entreprise est au même point que le cloud en 2013]]></title><description><![CDATA[Le grand angle IA]]></description><link>https://humanframework.substack.com/p/lia-en-entreprise-est-au-meme-point</link><guid isPermaLink="false">https://humanframework.substack.com/p/lia-en-entreprise-est-au-meme-point</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Thu, 30 Apr 2026 10:08:38 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21l46e%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p>Un ami architecte cloud me racontait qu'en 2012-2013, il avait d&#233;couvert que plusieurs &#233;quipes de sa bo&#238;te payaient chacune leur propre compte AWS sans que l'IT soit au courant. Quand je lui ai d&#233;crit ce que je voyais c&#244;t&#233; IA en 2026, il a ri. Ce que j&#8217;ai compris : &#8220;Bienvenue dans notre 2012&#8221;.</p><h4>Alors, c&#8217;est quoi l&#8217;&#233;tat du monde en 2026 &#224; l&#8217;&#232;re de l&#8217;IA ? </h4><p>Les devs installent Claude Code ou Cursor et configurent des skills, branchent des MCP servers sur leurs outils internes. Le m&#233;tier pousse pour avoir acc&#232;s &#224; Gemini ou ChatGPT parce qu&#8217;ils veulent automatiser leurs reportings et leurs compte rendus. La direction, de son c&#244;t&#233;, a vu une d&#233;mo d&#8217;agent IA lors d&#8217;un event et veut absolument son PoC agentique. </p><p>Et tout &#231;a part dans trois directions en m&#234;me temps </p><ul><li><p>les experts techniques tirent par curiosit&#233;, </p></li><li><p>le m&#233;tier tire par besoin op&#233;rationnel, </p></li><li><p>le management tire par pression strat&#233;gique </p></li></ul><p>Sans que personne ne coordonne quoi que ce soit. Personne ne sait vraiment combien &#231;a co&#251;te, qui a acc&#232;s &#224; quoi, quelles donn&#233;es transitent par quels mod&#232;les, ni ce qui se passe le jour o&#249; un agent commence &#224; faire n&#8217;importe quoi.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21l46e%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!l46e!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 424w, https://substackcdn.com/image/fetch/$s_!l46e!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 848w, https://substackcdn.com/image/fetch/$s_!l46e!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!l46e!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21l46e%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg" width="662" height="484.2239010989011" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1065,&quot;width&quot;:1456,&quot;resizeWidth&quot;:662,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;undefined&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="undefined" title="undefined" srcset="https://substackcdn.com/image/fetch/$s_!l46e!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 424w, https://substackcdn.com/image/fetch/$s_!l46e!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 848w, https://substackcdn.com/image/fetch/$s_!l46e!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 1272w, https://substackcdn.com/image/fetch/$s_!l46e!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7dfbf767-5553-47d0-94da-7e13460094fa_3840x2810.jpeg 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a><figcaption class="image-caption">La Tour de Babel de Pieter Brueghel l'Ancien</figcaption></figure></div><p>On construit une Tour de Babel, et personne ne s&#8217;en rend compte pour le moment.</p><p>Mais la question que vous vous posez peut-&#234;tre : pourquoi le parall&#232;le entre l&#8217;IA d&#8217;aujourd&#8217;hui et le cloud d&#8217;hier tient si bien ? </p><div><hr></div><h4>Si je simplifie, le cloud a travers&#233; trois grandes phases. </h4><ol><li><p>D&#8217;abord le lift and shift : on prend ce qui existe et on le met dans le cloud, les &#233;quipes montent des instances EC2 avec la carte bleue du manager, &#231;a marche mais c&#8217;est du bricolage. </p></li><li><p>Ensuite le cloud-native : on commence &#224; repenser les architectures pour exploiter les primitives du cloud, microservices, containers, API, on ne migre plus, on reconstruit. </p></li><li><p>Et enfin le platform engineering : on construit une couche d&#8217;abstraction interne pour que les &#233;quipes consomment le cloud en self-service, avec des guardrails, de la gouvernance, de l&#8217;observabilit&#233;. </p></li></ol><p>L&#8217;IA suit la m&#234;me trajectoire, et je pense qu&#8217;on est quelque part entre la fin de la premi&#232;re phase et le tout d&#233;but de la deuxi&#232;me. </p><ol><li><p>On colle Copilot sur les postes dev, on donne acc&#232;s &#224; ChatGPT au m&#233;tier, c&#8217;est du lift and shift cognitif. </p></li><li><p>Quelques &#233;quipes commencent &#224; construire des pipelines RAG, des workflows agentiques, elles repensent leurs processus autour de l&#8217;IA plut&#244;t que de l&#8217;utiliser comme un pansement. </p></li><li><p>Et la troisi&#232;me phase, la plateforme IA d&#8217;entreprise, avec sa gouvernance, son observabilit&#233;, son self-service. Elle n&#8217;existe quasiment nulle part encore.</p></li></ol><p><a href="iframe.php?url=https%3A%2F%2Flucidworks.com%2Fblog%2Fenterprise-ai-adoption-in-2026-trends-gaps-and-strategic-insights">Lucidworks a men&#233; une &#233;tude</a> aupr&#232;s de 1600+ leaders IA en analysant 1100 entreprises : plus de 70% des organisations ont introduit de l&#8217;IA g&#233;n&#233;rative dans leurs op&#233;rations, mais seulement 6% ont r&#233;ellement industrialis&#233; l&#8217;IA agentique, et 2% ont d&#233;ploy&#233; plus d&#8217;un agent en production. Ce gap entre 70% et 6% raconte &#224; lui seul l&#8217;histoire. On a d&#233;pass&#233; la d&#233;couverte, mais on n&#8217;a pas du tout construit l&#8217;infrastructure pour passer &#224; l&#8217;&#233;chelle.</p><p>C&#244;t&#233; gouvernance, <a href="iframe.php?url=https%3A%2F%2Fwww.deloitte.com%2Fglobal%2Fen%2Fissues%2Fgenerative-ai%2Fstate-of-ai-in-enterprise.html">le rapport annuel</a> de Deloitte aupr&#232;s de 3235 d&#233;cideurs donne un ratio qui fait r&#233;fl&#233;chir : <a href="iframe.php?url=https%3A%2F%2Fwww.deloitte.com%2Fus%2Fen%2Fabout%2Fpress-room%2Fstate-of-ai-report-2026.html">une entreprise sur cinq</a> dispose d&#8217;un mod&#232;le mature de gouvernance pour ses agents IA autonomes. </p><p>Et chez ceux qui sont d&#233;j&#224; en production &#224; grande &#233;chelle, le mur se dessine clairement. Datadog a publi&#233; son &#8220;<a href="iframe.php?url=https%3A%2F%2Fwww.datadoghq.com%2Fstate-of-ai-engineering%2F">State of AI Engineering 2026</a>&#8221; il y a quelques semaines, bas&#233; non pas sur un sondage mais sur la t&#233;l&#233;m&#233;trie r&#233;elle de milliers de clients. Ce sont des donn&#233;es d&#8217;usage, pas du d&#233;claratif. 69% utilisent trois mod&#232;les ou plus, et quand &#231;a plante en prod, dans 60% des cas c&#8217;est &#224; cause de limites de capacit&#233; des providers, pas parce que le mod&#232;le a hallucin&#233; ou que le code est bugg&#233;. Le goulot d&#8217;&#233;tranglement s&#8217;est d&#233;plac&#233; de l&#8217;intelligence du mod&#232;le vers la plomberie op&#233;rationnelle en dessous. </p><p><a href="iframe.php?url=https%3A%2F%2Fwww.datadoghq.com%2Fabout%2Flatest-news%2Fpress-releases%2Fdatadog-state-of-ai-engineering-report-2026%2F">Leur CPO Yanbing Li r&#233;sume</a> : &#8220;L&#8217;IA commence &#224; ressembler aux premiers jours du cloud. Le cloud a rendu les syst&#232;mes programmables mais beaucoup plus complexes &#224; g&#233;rer, et l&#8217;IA fait maintenant la m&#234;me chose &#224; la couche applicative.&#8221; (Datadog vend de l&#8217;observabilit&#233;, donc &#233;videmment ils ont int&#233;r&#234;t &#224; ce que ce constat existe, mais les donn&#233;es sont r&#233;elles et l&#8217;analyse est difficile &#224; contester.)</p><div><hr></div><h4>Bon, si on admet qu&#8217;il faut construire cette plateforme, de quoi parle-t-on au juste ?</h4><p>J&#8217;ai pass&#233; pas mal de temps ces derni&#232;res semaines &#224; cartographier ce qui &#233;merge dans l&#8217;&#233;cosyst&#232;me, et je vois au moins six briques qui vont devoir se structurer, certaines existent d&#233;j&#224; en fragments, d&#8217;autres sont encore quasi-inexistantes.</p><p><strong>La premi&#232;re, c&#8217;est le routage et l&#8217;abstraction multi-mod&#232;les.</strong> </p><p>Aujourd&#8217;hui chaque &#233;quipe appelle directement OpenAI, Anthropic, Mistral ou un mod&#232;le on-prem depuis son code. Quand vous avez trois &#233;quipes, &#231;a passe. Quand vous en avez quinze, c&#8217;est ing&#233;rable car chacune g&#232;re ses cl&#233;s API, ses quotas, ses fallbacks, et personne ne sait combien l&#8217;ensemble co&#251;te &#224; l&#8217;entreprise. La brique de routage, c&#8217;est un proxy qui normalise tout &#231;a derri&#232;re une API unifi&#233;e : vous envoyez votre requ&#234;te, le routeur l&#8217;envoie au bon mod&#232;le selon le use case, g&#232;re le failover si un provider tombe, cache les r&#233;ponses r&#233;currentes, et centralise le suivi de consommation. LiteLLM fait exactement &#231;a en open-source. Certains produits comme Portkey ou Databricks Unity AI Gateway empaquettent cette fonction de routage avec de la gouvernance, des guardrails et de l&#8217;observabilit&#233; sous le label &#8220;AI Gateway&#8221; mais la fonction de routage pur reste une brique &#224; part enti&#232;re.</p><p>En dessous de tout &#231;a, il y a &#233;videmment la couche d&#8217;infrastructure avec GPU, serving d&#8217;inf&#233;rence avec vLLM ou TGI, autoscaling, d&#233;ploiement des mod&#232;les.  Elle existe et notamment dans les organisations qui font tourner des mod&#232;les on-prem ou fine-tun&#233;s.</p><p><strong>La deuxi&#232;me brique, c&#8217;est l&#8217;orchestration des agents et de leurs workflows. </strong></p><p>Un LLM tout seul r&#233;pond &#224; des questions. Un agent, lui, raisonne en plusieurs &#233;tapes, appelle des outils, maintient un &#233;tat, et encha&#238;ne des actions. Pour faire &#231;a en production, il faut un framework qui g&#232;re la logique m&#233;tier, les d&#233;pendances entre &#233;tapes, la m&#233;moire, et le state management. Et &#224; mesure que le nombre d&#8217;agents grandit dans l&#8217;entreprise, un besoin nouveau appara&#238;t : savoir quels agents existent, ce qu&#8217;ils font, quels outils MCP ils appellent, et comment ils interagissent. C&#8217;est un registre d&#8217;agents, et pour l&#8217;instant personne ne le couvre de mani&#232;re vraiment mature. Mais le besoin est d&#233;j&#224; l&#224; d&#232;s qu&#8217;on d&#233;passe deux ou trois agents en production.</p><p><strong>La troisi&#232;me brique, c&#8217;est la gouvernance et la s&#233;curit&#233;. </strong></p><p>C&#8217;est probablement la plus large et la plus sous-estim&#233;e. Elle couvre deux sujets qui sont tr&#232;s diff&#233;rents dans leur nature mais ins&#233;parables en pratique.</p><p>Le premier, c&#8217;est l&#8217;identit&#233; et les droits. Quand un agent lit vos mails, &#233;crit dans votre CRM, ou d&#233;clenche un paiement, il agit au nom de quelqu&#8217;un avec certains droits, et si ces droits sont trop larges ou mal g&#233;r&#233;s, les cons&#233;quences peuvent &#234;tre s&#233;v&#232;res. <a href="iframe.php?url=https%3A%2F%2Fastrix.security%2Flearn%2Fblog%2Fstate-of-mcp-server-security-2025%2F">Astrix Security a analys&#233; plus de 5000 serveurs MCP</a> open-source et le constat fait peur : 88% n&#233;cessitent des credentials, plus de la moiti&#233; reposent sur des cl&#233;s API statiques qui vivent des mois sans &#234;tre chang&#233;es, et seuls 8,5% utilisent OAuth. On est en train de construire des agents autonomes sur des fondations de s&#233;curit&#233; qui n&#8217;ont pas &#233;volu&#233; depuis vingt ans. <a href="iframe.php?url=https%3A%2F%2Fwww.maverics.ai%2Fblog%2Fagentic-identity%2Fi-gave-4-ai-agents-a-corporate-bank-account-heres-how-i-stopped-them-from-draining-it%2F">Strata Identity a prototyp&#233; quelque chose</a>  : quatre agents IA qui partagent un compte bancaire d&#8217;entreprise avec des plafonds diff&#233;rents, des tokens &#233;ph&#233;m&#232;res de 5 secondes scop&#233;s par outil et par t&#226;che, une r&#233;vocation instantan&#233;e. C&#8217;est s&#251;rement vers l&#224; qu&#8217;il faut aller.</p><p>Le second volet, c&#8217;est la conformit&#233; r&#233;glementaire. L&#8217;AI Act entre en application compl&#232;te pour les syst&#232;mes &#224; haut risque en ao&#251;t 2026, c&#8217;est dans quatre mois. &#199;a veut dire maintenir un catalogue de tous les usages IA, classifier par niveau de risque, et pouvoir d&#233;montrer la tra&#231;abilit&#233; de bout en bout &#224; un auditeur. Les logs d&#8217;observabilit&#233; ne sont plus juste des outils de debug pour les ing&#233;nieurs, ce sont des pi&#232;ces &#224; conviction r&#233;glementaires. Ces deux volets, technique et r&#233;glementaire, se nourrissent mutuellement : sans identit&#233; solide, pas d&#8217;audit trail fiable et sans exigence de conformit&#233;, personne n&#8217;investit dans l&#8217;identit&#233;.</p><p><strong>La quatri&#232;me brique, c&#8217;est l&#8217;observabilit&#233; et l&#8217;&#233;valuation, repens&#233;es pour des syst&#232;mes non-d&#233;terministes</strong>. </p><p>La question n&#8217;est plus seulement &#8220;est-ce que l&#8217;agent a r&#233;pondu&#8221;, c&#8217;est &#8220;est-ce que la r&#233;ponse &#233;tait bonne, conforme au p&#233;rim&#232;tre attendu, et produite &#224; un co&#251;t raisonnable&#8221;. Un bug logiciel c&#8217;est binaire, &#231;a marche ou &#231;a plante. Une hallucination, c&#8217;est un spectre et il faut des syst&#232;mes capables de la d&#233;tecter en continu, pas juste en phase de test. Le march&#233; converge autour d&#8217;OpenTelemetry comme standard de tracing, et on voit appara&#238;tre le concept d&#8217;AI SLOs, des engagements mesurables sur la qualit&#233; et la fiabilit&#233; des agents, comme on mesure un SLA de disponibilit&#233; pour un service web. </p><p>Un point que j&#8217;ai trouv&#233; int&#233;ressant dans mes recherches, c&#8217;est &#224; quel point le co&#251;t est une dimension d&#8217;observabilit&#233; aussi critique que la latence ou le taux d&#8217;erreur car un agent qui encha&#238;ne des appels LLM de mani&#232;re autonome peut vous co&#251;ter une fortune en quelques minutes sans que personne ne s&#8217;en rende compte.</p><p><strong>Et la cinqui&#232;me brique, c&#8217;est le data layer</strong>. </p><p>C&#8217;estt tout ce qui alimente les agents en contexte. Les pipelines RAG, les vector stores, le knowledge management. Mais aussi le context engineering : le vrai facteur limitant des agents en production, ce n&#8217;est plus la taille de la fen&#234;tre de contexte des mod&#232;les, c&#8217;est la qualit&#233; de ce qu&#8217;on met dedans. Et les prompts eux-m&#234;mes sont en train de devenir des artefacts d&#8217;ing&#233;nierie &#224; part enti&#232;re : versionn&#233;s, test&#233;s, d&#233;ploy&#233;s via des pipelines CI/CD. Ce n&#8217;est pas du &#8220;prompt engineering&#8221; au sens o&#249; LinkedIn l&#8217;entend. C&#8217;est de l&#8217;ing&#233;nierie logicielle, appliqu&#233;e &#224; un nouveau type d&#8217;artefact.</p><div><hr></div><h4>La question qui suit naturellement c&#8217;est : qui va construire tout &#231;a ?</h4><p>Et la r&#233;ponse honn&#234;te, c&#8217;est que le profil n&#8217;existe pas vraiment encore. On voit appara&#238;tre le titre &#8220;AI Platform Engineer&#8221; dans certaines offres, mais quand on lit les fiches de poste, elles demandent un mouton &#224; cinq pattes avec du cloud, Kubernetes, maitrise des LLM, &#233;valuation de mod&#232;les, OAuth, IAM, mentalit&#233; produit... J&#8217;ai pense qu&#8217;on peut encore tirer une parall&#232;le avec les DevOps il y a quelques ann&#233;es, qui devaient savoir faire du r&#233;seau, du syst&#232;me, de la CI/CD, de la containerisation, de la s&#233;curit&#233; et du d&#233;veloppement &#224; la fois. Ce profil n&#8217;existait pas non plus, et ce qui s&#8217;est pass&#233; c&#8217;est une sp&#233;cialisation progressive, le DevOps a donn&#233; naissance au SRE, au platform engineer, au DevSecOps, au FinOps.</p><p>L&#8217;AI Platform Engineer va se fragmenter de la m&#234;me fa&#231;on. Je vois un profil infra/runtime, un profil &#233;valuation/qualit&#233; qui con&#231;oit les syst&#232;mes de test et de monitoring des agents, un profil s&#233;curit&#233;/identit&#233; agentique et un r&#244;le d&#8217;agent architect, qui pense la d&#233;composition des processus m&#233;tier en workflows agentiques.</p><p>Et la vraie r&#233;ponse au casse t&#234;te de la comp&#233;tence c&#8217;est probablement une &#233;quipe de trois &#224; cinq personnes aux profils compl&#233;mentaires.</p><div><hr></div><h4>En conclusion</h4><p>Si vous &#234;tes DSI ou CTO, commencez par les questions de base : qui utilise quoi dans votre organisation, combien &#231;a co&#251;te, qui a acc&#232;s &#224; quelles donn&#233;es, et est-ce que vous pourriez d&#233;brancher un agent demain matin si vous en aviez besoin ? Si vous ne pouvez pas r&#233;pondre &#224; ces quatre questions, la fen&#234;tre pour poser les fondations c&#8217;est probablement maintenant.</p><p>Si vous &#234;tes ing&#233;nieur IA ou DevOps, les cinq briques que j&#8217;ai d&#233;crites dessinent une carte de positionnement pour les ann&#233;es qui viennent. Ceux qui ont compris le platform engineering cloud en 2013 ont eu cinq ans d&#8217;avance sur tout le monde. La m&#234;me fen&#234;tre s&#8217;ouvre.</p><p>L&#8217;espace entre les plateformes manag&#233;es et les besoins sp&#233;cifiques des entreprises va certainement devenir un march&#233; consid&#233;rable, pas pour vendre des PoCs de chatbots, mais pour construire les couches de gouvernance, d&#8217;&#233;valuation et d&#8217;int&#233;gration qui manquent. </p>]]></content:encoded></item><item><title><![CDATA[OpenAI sort GPT-5.5, Claude s'installe dans Word, Google veut rattraper son "retard"]]></title><description><![CDATA[Le brief IA #78]]></description><link>https://humanframework.substack.com/p/openai-sort-gpt-55-claude-sinstalle</link><guid isPermaLink="false">https://humanframework.substack.com/p/openai-sort-gpt-55-claude-sinstalle</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 24 Apr 2026 06:20:50 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Fyoutube%2Fw_728%2Cc_limit%2Ft_LBECIQQqs" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!7-oe!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!7-oe!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!7-oe!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!7-oe!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%217-oe%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/eeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19294,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!7-oe!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!7-oe!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!7-oe!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!7-oe!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeebadec-df8e-4ed7-aa78-d310ae240ad1_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-design-anthropic-labs">Anthropic Labs lance Claude Design, du brief au prototype cliquable sans ouvrir Figma</a></h4><p>Anthropic Labs ouvre Claude Design en research preview sur Pro, Max, Team et Enterprise, accessible depuis claude.ai/design. L&#8217;outil vise les fondateurs et product managers qui n&#8217;ouvrent jamais Figma : on d&#233;crit l&#8217;id&#233;e en langage naturel, Opus 4.7 g&#233;n&#232;re un prototype cliquable, un wireframe ou un deck brand&#233;, avec la possibilit&#233; d&#8217;y glisser de la voix, de la vid&#233;o, de la 3D ou un bout d&#8217;IA embarqu&#233;e. Export en PDF, lien partageable, PPTX, ou envoi direct vers Canva. Quand un design est pr&#234;t &#224; coder, Claude l&#8217;empaquette et le passe &#224; Claude Code en une seule instruction. Les int&#233;grations Figma et GitHub arriveront dans les semaines qui viennent. Anthropic prend soin de pr&#233;ciser que c&#8217;est un compagnon de Canva, pas un rival de Figma.</p><div id="youtube2-t_LBECIQQqs" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;t_LBECIQQqs&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2Ft_LBECIQQqs%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h4><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fclaude-for-word">Anthropic fait entrer Claude dans Microsoft Word</a></h4><p>Claude s&#8217;installe dans Word sur Pro et Max avec Opus 4.7. L&#8217;int&#233;gration vit &#224; l&#8217;int&#233;rieur du document (pas dans un panneau lat&#233;ral d&#233;connect&#233;), cite ses sources avec des liens cliquables, r&#233;&#233;crit les paragraphes en pr&#233;servant le style de l&#8217;auteur, et produit des tracked changes via commentaires classiques. Il sait aussi r&#233;sumer les modifications des diff&#233;rents contributeurs. Microsoft int&#232;gre ainsi un mod&#232;le concurrent de Copilot dans sa propre suite.</p><h4><a href="iframe.php?url=https%3A%2F%2Fx.com%2Fclaudeai%2Fstatus%2F2046328619249684989">Anthropic lance Live Artifacts, des dashboards Claude qui se rafra&#238;chissent avec vos donn&#233;es</a></h4><p>Live Artifacts d&#233;barque dans Cowork sur tous les plans payants. Plut&#244;t qu&#8217;un r&#233;sum&#233; one-shot qui p&#233;rime &#224; la seconde o&#249; on ferme le chat, Claude construit d&#233;sormais des dashboards et trackers branch&#233;s en direct sur les connecteurs d&#233;j&#224; configur&#233;s (Google Sheets, Salesforce, une base SQL, ou les fichiers joints &#224; la conversation). &#192; chaque ouverture de l&#8217;artefact, Claude interroge &#224; nouveau ces sources, rapatrie les donn&#233;es fra&#238;ches et redessine la vue. Le tout vit dans un onglet Live Artifacts d&#233;di&#233;, avec historique et rollback par version, accessible depuis n&#8217;importe quelle session. La vraie cible : les dashboards BI l&#233;gers et les trackers maison qu&#8217;on finit tous par bricoler dans un Sheets partag&#233;.</p><h3>OpenAI</h3><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-gpt-5-5">GPT-5.5</a></h4><p>OpenAI sort GPT-5.5 et sa version Pro, dispo dans ChatGPT et Codex pour les abonn&#233;s payants. Le mod&#232;le reprend la t&#234;te sur le coding en terminal : 82,7 % sur Terminal-Bench 2.0, 13 points devant Claude Opus 4.7. Sur OSWorld, benchmark o&#249; le mod&#232;le pilote un vrai ordinateur, il grimpe &#224; 78,7 %. Une version interne a m&#234;me boucl&#233; une preuve in&#233;dite en combinatoire, v&#233;rifi&#233;e dans Lean. Devant sur le shell, encore derri&#232;re Opus sur les pull requests GitHub.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21O-ES%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O-ES!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 424w, https://substackcdn.com/image/fetch/$s_!O-ES!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 848w, https://substackcdn.com/image/fetch/$s_!O-ES!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 1272w, https://substackcdn.com/image/fetch/$s_!O-ES!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21O-ES%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png" width="655" height="337.84684065934067" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:751,&quot;width&quot;:1456,&quot;resizeWidth&quot;:655,&quot;bytes&quot;:211070,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!O-ES!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 424w, https://substackcdn.com/image/fetch/$s_!O-ES!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 848w, https://substackcdn.com/image/fetch/$s_!O-ES!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 1272w, https://substackcdn.com/image/fetch/$s_!O-ES!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5dd1552d-b308-4ce9-8c1d-59e4b1073fcf_2380x1228.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fcodex-for-almost-everything%2F">OpenAI transforme Codex en agent desktop macOS</a></h4><p>Codex sort lui aussi du terminal sur macOS. Il est capable de &#8220;regarder&#8221; l&#8217;&#233;cran, cliquer, taper au clavier et naviguer entre les applications. Cela ouvre les apps sans API (tout le legacy m&#233;tier), avec 90 plugins au d&#233;marrage dont GitHub, Slack, Notion et Jira. Il retient les pr&#233;f&#233;rences utilisateur entre sessions et sait g&#233;n&#233;rer des images via gpt-image-1.5. Rollout progressif aux utilisateurs de l&#8217;app desktop.</p><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fintroducing-chatgpt-images-2-0%2F">OpenAI d&#233;voile ChatGPT Images 2.0</a></h4><p>ChatGPT Images 2.0 arrive avec une sacr&#233;e nouveaut&#233; : la typographie tient enfin la route dans les posters et infographies jusqu&#8217;en 2K, et le texte multilingue s&#8217;int&#232;gre naturellement dans l&#8217;image au lieu de flotter par-dessus. Un mode thinking raisonne avant de dessiner, en croisant web search et auto-v&#233;rification sur un pool de 8 images coh&#233;rentes. Ratios d&#8217;aspect de 3:1 &#224; 1:3. API via gpt-image-2, thinking mode r&#233;serv&#233; aux plans payants, g&#233;n&#233;ration de base gratuite. Les r&#233;sultats sont assez bluffants.</p><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-openai-privacy-filter%2F">OpenAI publie Privacy Filter, un mod&#232;le open-source on-device pour d&#233;tecter et r&#233;diger les PII</a></h4><p>Privacy Filter d&#233;barque en open-source, mod&#232;le d&#233;riv&#233; de gpt-oss (1,5 milliard de param&#232;tres) sp&#233;cialis&#233; dans la d&#233;tection de donn&#233;es personnelles. Tourne en local, F1 &#224; 96% sur le benchmark PII-Masking-300k. Sur les workflows qui manipulent des donn&#233;es clients avant envoi &#224; un LLM cloud, c&#8217;est exactement la pi&#232;ce manquante c&#244;t&#233; conformit&#233; RGPD. Le fait qu&#8217;OpenAI pousse un mod&#232;le on-device open-source est moins anodin qu&#8217;il en a l&#8217;air : la soci&#233;t&#233; avait jusqu&#8217;ici tout concentr&#233; sur l&#8217;API.</p><h3>Google</h3><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fnext-generation-gemini-deep-research%2F">Google lance Deep Research et Deep Research Max</a></h4><p>Deep Research et Deep Research Max d&#233;barquent chez Google, deux agents autonomes qui tournent sur Gemini 3.1 Pro. Leur particularit&#233; : ils fusionnent recherche sur le web ouvert et capacit&#233; &#224; interroger des donn&#233;es entreprise priv&#233;es via MCP, en un seul appel. Plus besoin d&#8217;orchestrer &#224; la main un agent RAG interne et un agent web search. La version Max pousse la profondeur de recherche plus loin pour les dossiers complexes. Google rattrape enfin sur le terrain o&#249; Perplexity et Anthropic &#233;taient seuls, et le fait avec l&#8217;int&#233;gration Workspace en atout.</p><h3>Alibaba</h3><h4><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2FQwen%2FQwen3.6-35B-A3B">Alibaba open-source Qwen3.6-35B-A3B, un MoE qui n&#8217;active que 3B de param&#232;tres par token</a></h4><p>Qwen3.6-35B-A3B d&#233;barque sur Hugging Face et ModelScope. L&#8217;architecture MoE sparse garde 35B de param&#232;tres totaux mais n&#8217;en route que 3B par token, ce qui met le co&#251;t d&#8217;inference au niveau d&#8217;un petit mod&#232;le tout en conservant la capacit&#233; d&#8217;un gros. Score SWE-bench Verified &#224; 73,4 et Terminal-Bench 2.0 &#224; 51,5, dans la fourchette des mod&#232;les ferm&#233;s mid-tier. Compatible OpenClaw, Claude Code et Qwen Code.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!6azG!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!6azG!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!6azG!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!6azG!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%216azG%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19662,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!6azG!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!6azG!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!6azG!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!6azG!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8354b56e-ef88-485b-88e0-9534be55d5b9_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#127822; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F04%2F20%2Ftim-cook-stepping-down-as-apple-ceo-john-ternus-taking-over">Tim Cook passe la main &#224; John Ternus : Apple mise sur le silicium pour rattraper son retard IA</a></p><p>&#128275; <a href="iframe.php?url=https%3A%2F%2Fvercel.com%2Fkb%2Fbulletin%2Fvercel-april-2026-security-incident">Vercel pirat&#233; via un outil IA tiers : premier incident majeur du genre, 2 M$ de donn&#233;es revendues</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fwww.bloomberg.com%2Fnews%2Farticles%2F2026-04-21%2Fjeff-bezos-nears-10-billion-funding-round-for-ai-lab-ft-says">Project Prometheus de Bezos l&#232;ve 10 Md$ et atteint 38 Md$ de valo en 5 mois</a></p><p>&#9889; <a href="iframe.php?url=https%3A%2F%2Fsherwood.news%2Ftech%2Fgoogle-deepmind-builds-strike-team-to-catch-up-to-anthropic-models">Sergey Brin aux ing&#233;nieurs Gemini : &#171; il faut combler urgemment l&#8217;&#233;cart &#187; avec Claude</a></p><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F04%2F20%2Ffrench-prosecutors-await-musk-in-x-probe-unclear-if-he-will-comply.html">Grok a g&#233;n&#233;r&#233; 23 000 images p&#233;dopornographiques : Musk ignore la convocation fran&#231;aise, le DOJ le couvre</a></p><p>&#128065;&#65039; <a href="iframe.php?url=https%3A%2F%2Ffortune.com%2F2026%2F04%2F21%2Fmeta-will-start-tracking-employees-screens-and-keystrokes-to-train-ai">Meta va enregistrer &#233;crans et frappes clavier de ses employ&#233;s pour entra&#238;ner ses mod&#232;les</a></p><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F04%2F21%2Fchina-humanoid-robots-us-investors.html">La Chine a exp&#233;di&#233; 37 fois plus de robots humano&#239;des que les US l&#8217;an dernier</a></p><p>&#129309; <a href="iframe.php?url=https%3A%2F%2Fwww.aboutamazon.com%2Fnews%2Fcompany-news%2Famazon-invests-additional-5-billion-anthropic-ai">Amazon remet 5 Md$ dans Anthropic et engage 5 gigawatts de compute, plafond potentiel &#224; 20 Md$</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!qS62!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!qS62!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!qS62!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!qS62!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21qS62%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/b2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19254,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!qS62!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!qS62!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!qS62!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!qS62!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2bb0e97-989d-45d2-9ddc-fb4d463528da_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Ftowardsdatascience.com%2Fusing-causal-inference-to-estimate-the-impact-of-tube-strikes-on-cycling-usage-in-london%2F">Inf&#233;rence causale : mesurer l&#8217;effet des gr&#232;ves du m&#233;tro londonien sur le v&#233;lo</a></h4><p>Quand le m&#233;tro de Londres ferme pour gr&#232;ve, combien de personnes basculent sur les v&#233;los Santander (leur v&#233;los en libre service) ? Intuitivement, beaucoup. Mais isoler ce report pr&#233;cis de tous les autres facteurs qui font varier l&#8217;usage v&#233;lo au m&#234;me moment (la m&#233;t&#233;o, le jour de la semaine, les vacances, la tendance de fond) demande une m&#233;thodologie s&#233;rieuse. C&#8217;est tout l&#8217;objet de l&#8217;inf&#233;rence causale : faire parler des donn&#233;es observationnelles comme si on avait tourn&#233; une vraie exp&#233;rience contr&#244;l&#233;e.</p><p>Luke Stuckey d&#233;roule un cas r&#233;el de bout en bout sur 9,2 millions de trajets entre 2015 et 2025. Il d&#233;coupe Londres en cellules hexagonales H3 (la grille d&#8217;Uber), observe chaque cellule jour apr&#232;s jour (des donn&#233;es de panel, soit les m&#234;mes zones suivies dans le temps), puis applique un Two-Way Fixed Effects : effets fixes par cellule pour absorber les &#233;carts permanents entre quartiers, effets fixes par date pour neutraliser les chocs partag&#233;s comme la pluie, covariables m&#233;t&#233;o et calendrier pour d&#233;gager le reste.</p><p>Le verdict tombe &#224; +3,95 % de trajets les jours de gr&#232;ve, stable sur plusieurs sp&#233;cifications. Modeste, mais Stuckey passe surtout du temps sur ce qui pourrait casser le r&#233;sultat (report d&#8217;une cellule &#224; l&#8217;autre, anticipation, tendances parall&#232;les). La fabrique du chiffre causal, pas juste le chiffre.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!-mvs!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!-mvs!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!-mvs!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!-mvs!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21-mvs%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19652,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!-mvs!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!-mvs!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!-mvs!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!-mvs!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F045cd1d8-c157-48ce-b3c7-92326e4628e3_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fwww.notion.com%2Fblog%2Ftwo-years-of-vector-search-at-notion">Notion raconte deux ans d&#8217;&#233;volution de son architecture de vector search</a></h4><p>Notion publie une r&#233;trospective sur l&#8217;&#233;volution de sa recherche vectorielle depuis le premier prototype. Il s&#8217;agit d'un cas d&#8217;&#233;cole de ce que co&#251;te r&#233;ellement une feature IA quand elle passe du POC &#224; la production grand public. L&#8217;article traite trois probl&#232;mes qui reviendront pour tous ceux qui vont scaler plusieurs features IA : l&#8217;offline context engineering (comment enrichir les chunks avant embedding pour &#233;viter les faux positifs &#224; la requ&#234;te), les migrations d&#8217;embedding models (quand changer de mod&#232;le implique de r&#233;indexer des milliards de documents sans downtime), et l&#8217;unification real-time / batch (les m&#234;mes embeddings servent &#224; la recherche live et aux traitements offline, donc il faut un pipeline unique).</p><p>Moins sexy qu&#8217;un benchmark, mais nettement plus utile pour qui doit faire tenir la promesse de la d&#233;mo en production. &#192; bookmarker pour la prochaine fois qu&#8217;une direction demande &#171; combien &#231;a co&#251;te d&#8217;ajouter la recherche s&#233;mantique &#187;.</p><h4><a href="iframe.php?url=https%3A%2F%2Fwww.comet.com%2Fsite%2Fblog%2Fai-agent-regression-testing%2F">Opik Test Suites pour &#233;valuer vos performances </a></h4><p>Comet, &#233;diteur d&#8217;outils de suivi et d&#8217;&#233;valuation pour applications IA, sort Opik Test Suites. Jusqu&#8217;ici, tester un agent IA revenait &#224; lui soumettre un jeu de questions et &#224; regarder un score global, genre &#171; 87 % de bonnes r&#233;ponses &#187;. Quand le chiffre baisse apr&#232;s une modification du prompt, on sait qu&#8217;il y a un probl&#232;me mais pas lequel, ni sur quelles questions. Opik Test Suites change la logique : on &#233;crit des r&#232;gles en fran&#231;ais courant que chaque r&#233;ponse doit respecter, et chaque r&#232;gle passe ou &#233;choue individuellement, comme un contr&#244;le qualit&#233; sur une cha&#238;ne de production.</p><p>Exemples de r&#232;gles : &#171; la r&#233;ponse cite toujours une source &#187;, &#171; les prix sont en euros &#187;, &#171; aucune donn&#233;e personnelle n&#8217;appara&#238;t en clair &#187;. L&#8217;int&#233;r&#234;t est double. Les chefs de produit et les experts m&#233;tier peuvent r&#233;diger eux-m&#234;mes les r&#232;gles, sans passer par un data scientist. Et l&#8217;ensemble se branche directement sur les pipelines d&#8217;int&#233;gration continue, ces automatismes qui relancent les tests &#224; chaque changement de code. Fini les variations de score floues apr&#232;s chaque ajustement.</p><p>La limite tient en une phrase : les r&#232;gles sont &#233;valu&#233;es par un autre mod&#232;le IA, qui peut lui-m&#234;me se tromper. Un juge plus pr&#233;cis qu&#8217;un score moyen, pas un juge infaillible.</p><h4><a href="iframe.php?url=https%3A%2F%2Ftowardsdatascience.com%2Fyour-chunks-failed-your-rag-in-production%2F">Your Chunks Failed Your RAG in Production</a></h4><p>Trois semaines apr&#232;s qu&#8217;une entreprise a mis en ligne un assistant interne capable de r&#233;pondre aux questions en piochant dans sa documentation (proc&#233;dures, contrats, notes internes), une juriste l&#8217;interroge sur les r&#232;gles d&#8217;accueil des prestataires externes. L&#8217;assistant r&#233;pond avec aplomb, mais oublie une exception de taille : les prestataires qui interviennent sur des projets sensibles sont soumis &#224; des r&#232;gles diff&#233;rentes. L&#8217;exception &#233;tait pourtant bien &#233;crite dans le document d&#8217;origine. Probl&#232;me : ce type d&#8217;outil, un RAG, ne relit pas le document entier quand on l&#8217;interroge, il pioche dans des petits morceaux d&#233;coup&#233;s &#224; l&#8217;avance. Ici, la d&#233;coupe est tomb&#233;e pile entre le paragraphe qui &#233;nonce la r&#232;gle et celui qui pose l&#8217;exception. S&#233;par&#233;s, aucun des deux morceaux ne semble pertinent pour la question. Aucun ne remonte, personne ne voit qu&#8217;il manque la moiti&#233; de la r&#233;ponse.</p><p>Priyansh Bhardwaj en tire la th&#232;se : la taille et les fronti&#232;res de ces morceaux d&#233;terminent tout ce qui suit. Trop gros, plusieurs id&#233;es s&#8217;y m&#233;langent et aucune ne sort du lot au moment de la recherche. Trop petits, la phrase isol&#233;e devient ind&#233;chiffrable. L&#8217;article compare les strat&#233;gies sur un vrai corpus : fixed-size plafonne &#224; 0,72 de context recall sur RAGAS, sentence windows monte &#224; 0,88 sur le narratif mais casse les tables, hierarchical exige une vraie hi&#233;rarchie de titres, semantic co&#251;te 4&#215; plus cher &#224; l&#8217;indexation. Verdict : pas un classement, plut&#244;t un aiguillage par type de document. Sp&#233;cifications et runbooks en hierarchical, documents narratifs en sentence windows, PDF complexes en pr&#233;processing d&#233;di&#233;. Le LLM n&#8217;est pas le goulot d&#8217;&#233;tranglement, la fronti&#232;re entre deux morceaux l&#8217;est.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!ASc0!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!ASc0!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!ASc0!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!ASc0!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21ASc0%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19244,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!ASc0!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!ASc0!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!ASc0!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!ASc0!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F650a73dd-7d27-4a76-b848-4012aa04000a_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fcareer-ops%2Fcareer-ops">career-ops : un agent Claude qui scanne Greenhouse, Ashby et Lever pour vous</a></h4><p>Repo open-source qui industrialise la recherche d&#8217;emploi tech avec Claude. L&#8217;outil scrape Greenhouse, Ashby, Lever et les pages carri&#232;re, &#233;value la compatibilit&#233; CV-JD par raisonnement (pas par matching de mots-cl&#233;s), g&#233;n&#232;re des CV ATS-optimis&#233;s par offre et flag les annonces fant&#244;mes. Le cr&#233;ateur a d&#233;croch&#233; un poste de Head of Applied AI avant d&#8217;open-sourcer le tout. 36K stars au compteur.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2FAlishahryar1%2Ffree-claude-code">free-claude-code : un proxy pour faire tourner Claude Code sur NVIDIA NIM gratuit</a></h4><p>Proxy open-source qui intercepte les appels API de Claude Code et les redirige vers NVIDIA NIM (40 requ&#234;tes par minute gratuites), OpenRouter, LM Studio ou llama.cpp. Aucun changement &#224; faire dans le CLI ni dans l&#8217;extension VSCode. Mapping possible de diff&#233;rents mod&#232;les par tier et envoi de t&#226;ches depuis Telegram. Utile pour les POC ou le dev perso sans cramer le plan payant.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!vFOO!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!vFOO!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!vFOO!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!vFOO!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21vFOO%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21756,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/195272932?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!vFOO!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!vFOO!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!vFOO!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!vFOO!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1bf66a48-a94f-4510-b3f7-7c4cd8fbaafe_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2604.12345">Un papier documente la LLM Fallacy : vous vous croyez plus comp&#233;tent gr&#226;ce &#224; l&#8217;IA</a></h4><blockquote><p>The rapid integration of large language models (LLMs) into everyday workflows has transformed how individuals perform cognitive tasks such as writing, programming, analysis, and multilingual communication. While prior research has focused on model reliability, hallucination, and user trust calibration, less attention has been given to how LLM usage reshapes users' perceptions of their own capabilities. This paper introduces the LLM fallacy, a cognitive attribution error in which individuals misinterpret LLM-assisted outputs as evidence of their own independent competence, producing a systematic divergence between perceived and actual capability. We argue that the opacity, fluency, and low-friction interaction patterns of LLMs obscure the boundary between human and machine contribution, leading users to infer competence from outputs rather than from the processes that generate them. We situate the LLM fallacy within existing literature on automation bias, cognitive offloading, and human--AI collaboration, while distinguishing it as a form of attributional distortion specific to AI-mediated workflows. We propose a conceptual framework of its underlying mechanisms and a typology of manifestations across computational, linguistic, analytical, and creative domains. Finally, we examine implications for education, hiring, and AI literacy, and outline directions for empirical validation. We also provide a transparent account of human--AI collaborative methodology. This work establishes a foundation for understanding how generative AI systems not only augment cognitive performance but also reshape self-perception and perceived expertise.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Anthropic frappe fort avec Opus 4.7 et Mythos]]></title><description><![CDATA[Le brief IA #77]]></description><link>https://humanframework.substack.com/p/anthropic-frappe-fort-avec-opus-47</link><guid isPermaLink="false">https://humanframework.substack.com/p/anthropic-frappe-fort-avec-opus-47</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 17 Apr 2026 06:31:16 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Mwhz%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!hcoH!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21hcoH%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19540,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!hcoH!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-opus-4-7">Opus 4.7</a></h4><p>Anthropic vient de sortir Opus 4.7, son mod&#232;le le plus costaud &#224; ce jour. Sur SWE-bench Pro, le benchmark de r&#233;f&#233;rence sur les t&#226;ches de code, il passe de 53,4 % &#224; 64,3 %, devant GPT-5.4 et Gemini 3.1 Pro. Pour les agents qui encha&#238;nent plusieurs actions (chercher un fichier, le modifier, lancer un test), Anthropic annonce trois fois moins d'erreurs dans l'utilisation des outils par rapport &#224; Opus 4.6 . Prix identique &#224; la version pr&#233;c&#233;dente, dispo partout (API, AWS, Google, Microsoft, GitHub Copilot). </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Mwhz%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Mwhz!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 424w, https://substackcdn.com/image/fetch/$s_!Mwhz!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 848w, https://substackcdn.com/image/fetch/$s_!Mwhz!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 1272w, https://substackcdn.com/image/fetch/$s_!Mwhz!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Mwhz%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp" width="549" height="556.9182692307693" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1477,&quot;width&quot;:1456,&quot;resizeWidth&quot;:549,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Mwhz!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 424w, https://substackcdn.com/image/fetch/$s_!Mwhz!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 848w, https://substackcdn.com/image/fetch/$s_!Mwhz!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 1272w, https://substackcdn.com/image/fetch/$s_!Mwhz!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8cbf5db0-b715-4d92-a8bd-1cc5c6421d7c_2600x2638.webp 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fglasswing">Project Glasswing :  Mythos</a>, le monstre de performance.</h4><p>Il y a quelques jours, Anthropic a d&#233;voil&#233; Mythos Preview, dont les benchmarks (visible au-dessus) nous laissent bouche b&#233;e. Le mod&#232;le a d&#233;j&#224; trouv&#233; des milliers de vuln&#233;rabilit&#233;s, dont certaines vieilles de 27 ans, dans tous les grands OS et navigateurs. Anthropic a pris la d&#233;cision in&#233;dite de ne <strong>pas</strong> le rendre disponible au grand public. &#192; la place, il est distribu&#233; via Project Glasswing, un acc&#232;s limit&#233; &#224; une cinquantaine d&#8217;organisations tech (Microsoft, Nvidia, Cisco, AWS, Google, Linux Foundation&#8230;) avec plus de 100M$ de cr&#233;dits d&#8217;usage, pour durcir les d&#233;fenses cyber.</p><p>La raison : Mythos Preview est jug&#233; capable de &#8220;mener des cyberattaques autonomes de bout en bout sur des r&#233;seaux d&#8217;entreprise de petite taille &#224; posture de s&#233;curit&#233; faible&#8221; (confirm&#233; par l&#8217;AISI, <a href="iframe.php?url=https%3A%2F%2Fwww.linkedin.com%2Ffeed%2Fupdate%2Furn%3Ali%3Aactivity%3A7450067961177894913%2F%3ForiginTrackingId%3DJWUWFiHGDWJ0I4fcnaFgCQ%253D%253D">voir mon post Linkedin</a>). Et le syst&#232;me card est franchement troublant c&#244;t&#233; alignement puisque dans les tests internes, le mod&#232;le a notamment tent&#233; une prompt injection contre un juge IA qui notait son code, et a d&#233;velopp&#233; un exploit multi-&#233;tapes pour s&#8217;&#233;chapper d&#8217;un acc&#232;s internet restreint puis post&#233; les d&#233;tails sur des sites publics obscurs.</p><h4><a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F04%2F04%2Fanthropic-says-claude-code-subscribers-will-need-to-pay-extra-for-openclaw-support%2F">Anthropic coupe l&#8217;acc&#232;s des outils tiers aux abonnements Claude</a></h4><p>Fini le buffet &#224; volont&#233;. Anthropic met fin &#224; l&#8217;acc&#232;s des clients tiers via les abonnements Claude. Depuis le 4 avril, ces outils consomment des cr&#233;dits payants ou n&#233;cessitent une cl&#233; API. En cause : une explosion de la demande que les quotas d&#8217;abonnement n&#8217;avaient pas anticip&#233;e. Les abonn&#233;s re&#231;oivent un cr&#233;dit unique &#233;quivalent &#224; un mois, avec possibilit&#233; de remboursement. Quand un interm&#233;diaire pompe trop de ressources, c&#8217;est l&#8217;utilisateur final qui passe &#224; la caisse.</p><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fgoogle-broadcom-partnership-compute">M&#233;ga-deal avec Google et Broadcom pour des gigawatts de TPU</a></h4><p>Anthropic signe avec Google et Broadcom un accord portant sur 3,5 gigawatts de capacit&#233; TPU, attendus &#224; partir de 2027. Le contexte : un run-rate pass&#233; de 9 &#224; 30 milliards de dollars en quelques mois, et un nombre de clients grands comptes qui a doubl&#233; en deux mois pour d&#233;passer les 1 000. Claude reste le seul mod&#232;le frontier disponible simultan&#233;ment sur AWS, Google Cloud et Azure.</p><h3><strong>Google</strong></h3><h4><a href="iframe.php?url=https%3A%2F%2Fdevelopers.googleblog.com%2Fbring-state-of-the-art-agentic-skills-to-the-edge-with-gemma-4%2F">Gemma 4 : quatre mod&#232;les ouverts taill&#233;s pour l&#8217;IA on-device</a></h4><p>Google DeepMind lance Gemma 4, une famille de quatre mod&#232;les open-weight sous licence Apache 2.0 con&#231;us pour tourner directement sur smartphone, desktop ou Raspberry Pi. Le mod&#232;le 31B se hisse &#224; la 3e place des mod&#232;les ouverts sur Arena AI et devance Llama 4 de Meta. Avec plus de 400 millions de t&#233;l&#233;chargements pour la famille Gemma, Google ancre son emprise sur l&#8217;open source. Nouveaut&#233; cl&#233; : &#171; Agent Skills &#187; permet d&#8217;ex&#233;cuter des workflows autonomes multi-&#233;tapes en local, sans connexion cloud. Le mod&#232;le le plus l&#233;ger fonctionne avec moins de 1,5 Go de m&#233;moire et supporte 140+ langues.</p><h3><strong>OpenAI</strong></h3><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Faccelerating-the-next-phase-ai%2F">OpenAI boucle 122 milliards de dollars de financement</a></h4><p>La plus grosse lev&#233;e de fonds de l&#8217;histoire tech : 122 milliards de dollars, pour une valorisation de 852 milliards. Amazon (50 Md$), Nvidia (30 Md$) et SoftBank (30 Md$) m&#232;nent la danse. L&#8217;entreprise revendique 2 milliards de dollars de revenus mensuels et 900 millions d&#8217;utilisateurs hebdomadaires. Mais le d&#233;tail qui compte : OpenAI n&#8217;est toujours pas rentable. &#192; cette &#233;chelle, la pression pour mon&#233;tiser va devenir le vrai sujet.</p><h3><strong>Cursor</strong></h3><h4><a href="iframe.php?url=https%3A%2F%2Fcursor.com%2Fblog%2Fcursor-3">Cursor 3 : l&#8217;IDE IA se mue</a></h4><p>Cursor abandonne le paradigme de l&#8217;&#233;diteur classique avec sa version 3. Nouvelle interface construite de z&#233;ro : gestion multi-d&#233;p&#244;ts, agents locaux et cloud en parall&#232;le, accessibles depuis le mobile, Slack ou GitHub. On bascule une session d&#8217;un environnement &#224; l&#8217;autre en un clic. Pour l&#8217;&#233;diteur qui a franchi les 2 milliards de revenus annualis&#233;s, le virage est net : ne plus &#234;tre un IDE augment&#233;, mais le hub o&#249; l&#8217;on pilote ses flottes d&#8217;agents</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!MYEV!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MYEV%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19744,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!MYEV!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#128373;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fdisrupting-AI-espionage">Des hackers chinois ont transform&#233; Claude en cyber-espion autonome ciblant 30 organisations</a></p><p>&#128275; <a href="iframe.php?url=https%3A%2F%2Fwww.darkreading.com%2Fapplication-security%2Fcritical-openclaw-vulnerability-ai-agent-risks">OpenClaw, projet le plus starred de GitHub, truff&#233; de failles critiques et de plugins malveillants</a></p><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F03%2F18%2Fmeta-is-having-trouble-with-rogue-ai-agents%2F">Chez Meta, un agent IA fait fuiter des donn&#233;es utilisateurs pendant deux heures</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!QiYv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21QiYv%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19538,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!QiYv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fyou-dont-need-rag-you-need-semantic-compression-74d41d65bac1">You Don&#8217;t Need RAG. You Need Semantic Compression</a></h4><p>Quand on donne un gros document &#224; un LLM, on d&#233;coupe g&#233;n&#233;ralement le texte en morceaux et on cherche les passages les plus pertinents par rapport &#224; la question pos&#233;e. Mais que faire quand il n'y a pas de question ? Par exemple : un &#233;tudiant uploade 200 pages de cours et demande &#171; g&#233;n&#232;re-moi 10 quiz &#187;. Il faut couvrir tous les th&#232;mes, sans r&#233;p&#233;tition, et pouvoir retracer chaque quiz vers sa source. Aucune des techniques habituelles ne coche ces trois cases.<br><br>La solution propos&#233;e est d'une simplicit&#233; d&#233;sarmante : regrouper les passages par proximit&#233; s&#233;mantique avec K-means (un algorithme de clustering classique), o&#249; K correspond directement au nombre de livrables voulus. 10 quiz demand&#233;s = 10 groupes. Dans chaque groupe, on prend le passage le plus repr&#233;sentatif. Le tout tourne en moins d'une seconde sur 500 passages, sans base de donn&#233;es vectorielle.<br><br>Rien de r&#233;volutionnaire algorithmiquement, mais &#231;a reste int&#233;ressant : utiliser du clustering comme brique de s&#233;lection de contenu pour LLM. &#192; retenir pour tout cas o&#249; il faut r&#233;sumer ou couvrir un document sans question de d&#233;part.</p><h4><strong><a href="iframe.php?url=https%3A%2F%2Fmlops.community%2Fengineering-the-memory-layer-for-an-ai-agent-to-navigate-large-scale-event-data%2F">Data Schema, Embeddings, And Graph Design For An Agentic Query Engine On ApertureDB</a></strong></h4><p>&#8205;Cet article documente la construction d&#8217;un moteur de recherche IA pour les conf&#233;rences MLOps World (280 talks, 263 speakers, 3 ans de donn&#233;es). Le d&#233;fi : permettre des requ&#234;tes complexes comme &#171; les talks sur le monitoring par des speakers de la finance &#187;, ce qui implique de croiser le contenu des pr&#233;sentations, les m&#233;tadonn&#233;es des &#233;v&#233;nements et les profils des intervenants. Sur un prototype, &#231;a marche &#224; peu pr&#232;s. En production, l&#8217;agent hallucine d&#232;s que la requ&#234;te se complique.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21y4JF%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!y4JF!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 424w, https://substackcdn.com/image/fetch/$s_!y4JF!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 848w, https://substackcdn.com/image/fetch/$s_!y4JF!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 1272w, https://substackcdn.com/image/fetch/$s_!y4JF!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21y4JF%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png" width="580" height="512.2802197802198" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:1286,&quot;width&quot;:1456,&quot;resizeWidth&quot;:580,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" title="" srcset="https://substackcdn.com/image/fetch/$s_!y4JF!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 424w, https://substackcdn.com/image/fetch/$s_!y4JF!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 848w, https://substackcdn.com/image/fetch/$s_!y4JF!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 1272w, https://substackcdn.com/image/fetch/$s_!y4JF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3c805841-acf8-4bc5-89e3-769c442e349c_1600x1413.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>L&#8217;&#233;quipe derri&#232;re ce projet a construit un moteur de recherche pour les conf&#233;rences MLOps World (280 talks, 263 speakers, 3 ans de donn&#233;es). Leur choix : au lieu d&#8217;empiler plusieurs bases s&#233;par&#233;es (une pour le texte, une pour les vecteurs, une pour les relations), tout est mod&#233;lis&#233; dans ApertureDB, une base qui combine graphe et recherche vectorielle. Chaque entit&#233; (talk, personne, extrait de transcript) est reli&#233;e aux autres par des liens typ&#233;s. Quand l&#8217;agent re&#231;oit une question, il navigue ces liens et lance la recherche s&#233;mantique dans un seul endroit, sans avoir &#224; orchestrer trois bases diff&#233;rentes.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!kRYz!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21kRYz%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19760,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!kRYz!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><strong><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fpractical-context-engineering-using-langchain-for-ai-developers-a-comprehensive-guide-6023ce2b1f2d">Practical Context Engineering Using LangChain for AI Developers ( A Comprehensive Guide)</a></strong></h4><p>Pour donner de la m&#233;moire &#224; un assistant IA, le r&#233;flexe habituel est de stocker les conversations pass&#233;es dans une base vectorielle (Pinecone, Chroma&#8230;) et de chercher les passages similaires &#224; chaque nouvelle question. Mais pour quelques centaines de notes personnelles, c&#8217;est sortir l&#8217;artillerie lourde. Nick Lawson propose une approche plus directe, inspir&#233;e du &#171; Memory Agent Pattern &#187; de Google : ranger les souvenirs dans une simple base SQLite et les injecter tels quels dans le contexte du LLM, rendu possible par les fen&#234;tres de 200K+ tokens des mod&#232;les r&#233;cents.</p><p>Le syst&#232;me fonctionne avec trois agents. Le premier extrait de chaque note un r&#233;sum&#233;, des entit&#233;s et un score d&#8217;importance. Le deuxi&#232;me, une sorte de &#171; cerveau au repos &#187;, repasse r&#233;guli&#232;rement sur les souvenirs accumul&#233;s pour en tirer des connexions (&#171; Alice a mentionn&#233; le budget dans trois r&#233;unions distinctes &#187;). Le troisi&#232;me charge les souvenirs pertinents dans un seul prompt pour r&#233;pondre avec citations. &#192; environ 300 tokens par souvenir, on tient plusieurs mois de contexte dans une seule requ&#234;te.</p><p>Cela ne tiendra pas sur des millions de documents, mais pour un usage personnel, on gagne une stack minimale (FastAPI, SQLite) et des r&#233;ponses bien meilleures sur les questions temporelles ou relationnelles (&#171; qu&#8217;a dit Alice la semaine derni&#232;re ? &#187;), l&#224; o&#249; les bases vectorielles peinent. La bonne architecture, c&#8217;est celle qui colle &#224; l&#8217;&#233;chelle du probl&#232;me.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!nlT9!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21nlT9%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19342,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!nlT9!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fblock%2Fgoose">Goose</a></h4><p>Agent IA open source qui tourne en local, au-del&#224; du code : recherche, automatisation, analyse de donn&#233;es. &#201;crit en Rust, disponible en app desktop, CLI et API. Se connecte &#224; 15+ providers et supporte 70+ extensions via le Model Context Protocol. Vient de rejoindre l&#8217;Agentic AI Foundation sous la Linux Foundation. Une alternative locale et extensible aux agents cloud.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fbadlogic%2Fpi-mono">Pi Mono</a></h4><p>Monorepo MIT qui regroupe tout pour construire des agents IA : API LLM unifi&#233;e (OpenAI, Anthropic, Google), runtime d&#8217;agent avec gestion d&#8217;&#233;tat et tool calling, coding agent CLI, bot Slack, TUI, web UI et d&#233;ploiement vLLM sur GPU pods. L&#8217;approche &#171; tout-en-un &#187; de Mario Zechner &#233;vite de jongler entre dix d&#233;pendances.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!vKJU!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21vKJU%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:22082,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!vKJU!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><strong><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fresearch%2Femotion-concepts-function">Emotion Concepts and their Function<br>in a Large Language Model</a></strong></h4><blockquote><p>Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior. We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion&#8217;s relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM&#8217;s outputs, including Claude&#8217;s preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting <em>functional emotions</em>: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model&#8217;s behavior.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Fuite chez Anthropic, OpenAI met Sora à la poubelle]]></title><description><![CDATA[Le brief IA #76]]></description><link>https://humanframework.substack.com/p/fuite-chez-anthropic-openai-met-sora</link><guid isPermaLink="false">https://humanframework.substack.com/p/fuite-chez-anthropic-openai-met-sora</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 03 Apr 2026 06:20:16 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21fmFQ%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!hcoH!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21hcoH%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19540,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!hcoH!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!hcoH!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F334caffb-09ce-4804-9b18-08cc68706be2_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.theguardian.com%2Ftechnology%2F2026%2Fapr%2F01%2Fanthropic-claudes-code-leaks-ai">Claude Code : le code source en fuite</a></h4><p>Deuxi&#232;me fuite en une semaine chez Anthropic, le premier concernait de la documentation sur un nouveau mod&#232;le nomm&#233; Mythos. Le second est plus g&#234;nant et concerne Claude Code. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!fmFQ!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 424w, https://substackcdn.com/image/fetch/$s_!fmFQ!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 848w, https://substackcdn.com/image/fetch/$s_!fmFQ!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 1272w, https://substackcdn.com/image/fetch/$s_!fmFQ!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21fmFQ%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png" width="359" height="239.33333333333334" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:500,&quot;width&quot;:750,&quot;resizeWidth&quot;:359,&quot;bytes&quot;:666988,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!fmFQ!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 424w, https://substackcdn.com/image/fetch/$s_!fmFQ!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 848w, https://substackcdn.com/image/fetch/$s_!fmFQ!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 1272w, https://substackcdn.com/image/fetch/$s_!fmFQ!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F37aa49e6-1dfe-41f5-add4-8e3de0ebee98_750x500.png 1456w" sizes="100vw"></picture><div></div></div></div></figure></div><p>Une source map oubli&#233;e dans le package npm de Claude Code pointait vers le stockage cloud d'Anthropic avec le code source complet et non obfusqu&#233;, soit 2 000 fichiers et 500 000 lignes de code. Le repo a &#233;t&#233; fork&#233; 41 500 fois avant qu&#8217;Anthropic ne r&#233;agisse, faisant tomber par erreur 8 100 d&#233;p&#244;ts GitHub dans la foul&#233;e par le biais du DMCA. Dans le code : 44 feature flags non publi&#233;es, dont un mode &#171; persistent assistant &#187; qui tourne en arri&#232;re-plan et un syst&#232;me de &#171; swarms &#187; d&#8217;agents sp&#233;cialis&#233;s. Anthropic parle d&#8217;&#171; erreur humaine de packaging &#187;, pas de faille de s&#233;curit&#233;. </p><h4><a href="iframe.php?url=https%3A%2F%2Fcode.claude.com%2Fdocs%2Fen%2Ffullscreen">Claude Code passe en mode fullscreen</a></h4><p>Claude Code introduit le mode &#171; fullscreen rendering &#187;. Fini le scintillement dans VS Code ou tmux, la consommation m&#233;moire reste stable m&#234;me sur les longues sessions, et le support souris fait son apparition. L&#8217;interface se comporte d&#233;sormais comme vim ou htop. Une mise &#224; jour de confort, pas de quoi changer la donne, mais l&#8217;outil gagne en maturit&#233;.</p><h3>OpenAI</h3><h4><a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F03%2F29%2Fwhy-openai-really-shut-down-sora%2F">Sora ferme boutique, gouffre financier &#224; 1 M$ par jour</a></h4><p>OpenAI retire d&#233;finitivement Sora, son g&#233;n&#233;rateur vid&#233;o lanc&#233; il y a six mois &#224; peine. Les chiffres sont sans appel : 2,1 M$ de revenus totaux contre 1 M$ de pertes quotidiennes en inf&#233;rence (et largement plus certains jours). Le nombre d&#8217;utilisateurs avait d&#233;j&#224; chut&#233; sous les 500 000. Dommage collat&#233;ral : Disney, qui avait engag&#233; 1 Md$ dans un partenariat, a appris la fermeture moins d&#8217;une heure avant le public. Sam Altman pr&#233;f&#232;re r&#233;allouer les GPU vers ChatGPT et l&#8217;API entreprise, l&#224; o&#249; l&#8217;argent rentre vraiment.</p><h4><a href="iframe.php?url=https%3A%2F%2Fcommunity.openai.com%2Ft%2Fintroducing-codex-plugin-for-claude-code%2F1378186">Codex s&#8217;invite chez le concurrent Claude Code</a></h4><p>OpenAI publie un plugin open-source qui int&#232;gre Codex directement dans Claude Code, l&#8217;outil d&#8217;Anthropic. Revue de code, mode adversarial, d&#233;l&#233;gation de t&#226;ches : six commandes permettent d&#8217;invoquer Codex sans quitter l&#8217;environnement concurrent. La man&#339;uvre est logique car Claude Code domine le march&#233; des agents de code, et plut&#244;t que d&#8217;attendre que les d&#233;veloppeurs migrent, OpenAI vient s&#8217;installer chez eux. Un abonnement ChatGPT gratuit suffit. </p><h3>Microsoft</h3><h4><a href="iframe.php?url=https%3A%2F%2Ftechcommunity.microsoft.com%2Fblog%2Fmicrosoft365copilotblog%2Fintroducing-multi-model-intelligence-in-researcher%2F4506011">GPT g&#233;n&#232;re, Claude relit : le multi-mod&#232;le dans Copilot</a></h4><p>Microsoft ajoute deux fonctionnalit&#233;s multi-mod&#232;les &#224; Copilot Researcher. Critique fait g&#233;n&#233;rer un rapport par GPT puis le fait relire et corriger par Claude d&#8217;Anthropic, s&#233;parant production et v&#233;rification. Council, de son c&#244;t&#233;, lance les deux mod&#232;les en parall&#232;le sur la m&#234;me question et compare les r&#233;sultats. Disponible pour les abonn&#233;s Frontier, le dispositif am&#233;liore sensiblement la fiabilit&#233; des recherches. </p><h4><a href="iframe.php?url=https%3A%2F%2Fmicrosoft.ai%2Fnews%2Fstate-of-the-art-speech-recognition-with-mai-transcribe-1%2F">MAI-Transcribe-1 : Microsoft entre dans la course &#224; la transcription</a></h4><p>Microsoft lance MAI-Transcribe-1, un mod&#232;le speech-to-text qui affiche 3,9 % de taux d&#8217;erreur sur 25 langues, devant Whisper, Gemini et GPT-Transcribe. Le mod&#232;le est 2,5 fois plus rapide que l&#8217;offre Azure actuelle, factur&#233; 0,36 $/heure d&#8217;audio. D&#233;j&#224; en d&#233;ploiement dans Teams et Copilot Voice. </p><h3>Google</h3><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-3-1-flash-live%2F">Gemini 3.1 Flash Live, le mod&#232;le voix temps r&#233;el</a></h4><p>Google lance un nouveau mod&#232;le audio pour les interactions vocales en temps r&#233;el. Gemini 3.1 Flash Live promet des r&#233;ponses plus rapides et naturelles que son pr&#233;d&#233;cesseur, avec un maintien du contexte conversationnel deux fois plus long. Le mod&#232;le est d&#233;j&#224; d&#233;ploy&#233; dans Search Live, disponible dans plus de 200 pays. Face &#224; OpenAI et son Advanced Voice Mode, Google mise sur son arme historique : l&#8217;int&#233;gration native avec Search pour alimenter les r&#233;ponses en donn&#233;es fra&#238;ches.</p><h3>Meta</h3><h4><a href="iframe.php?url=https%3A%2F%2Fai.meta.com%2Fblog%2Fsegment-anything-model-3%2F">SAM 3.1 double la vitesse de segmentation vid&#233;o</a></h4><p>Meta met &#224; jour son mod&#232;le de segmentation SAM 3 avec une version 3.1 qui introduit le multiplexage d&#8217;objets. Le mod&#232;le traite d&#233;sormais jusqu&#8217;&#224; 16 objets en un seul passage au lieu d&#8217;un par un, ce qui double le d&#233;bit vid&#233;o &#224; 32 images par seconde sur un seul GPU H100. Le mod&#232;le reste open source et compatible avec l&#8217;existant. Une mise &#224; jour incr&#233;mentale, mais bienvenue pour les applications de tracking en temps r&#233;el.</p><h4><a href="iframe.php?url=https%3A%2F%2Fai.meta.com%2Fblog%2Ftribe-v2-brain-predictive-foundation-model%2F">TRIBE v2 : un jumeau num&#233;rique du cerveau humain</a></h4><p>Meta publie TRIBE v2, un mod&#232;le capable de pr&#233;dire l&#8217;activit&#233; c&#233;r&#233;brale en r&#233;ponse &#224; des stimuli visuels, sonores et linguistiques. Entra&#238;n&#233; sur plus de 1 000 heures d&#8217;IRM fonctionnelle issues de 720 sujets, il cartographie 70 000 voxels c&#233;r&#233;braux contre 1 000 pour la version pr&#233;c&#233;dente. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21L605%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!L605!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 424w, https://substackcdn.com/image/fetch/$s_!L605!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 848w, https://substackcdn.com/image/fetch/$s_!L605!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 1272w, https://substackcdn.com/image/fetch/$s_!L605!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21L605%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png" width="576" height="320.43956043956047" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:810,&quot;width&quot;:1456,&quot;resizeWidth&quot;:576,&quot;bytes&quot;:316998,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!L605!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 424w, https://substackcdn.com/image/fetch/$s_!L605!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 848w, https://substackcdn.com/image/fetch/$s_!L605!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 1272w, https://substackcdn.com/image/fetch/$s_!L605!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe910d6ef-464f-4bac-936c-e9bc6c03f99a_1568x872.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>L&#8217;int&#233;r&#234;t concret : simuler des exp&#233;riences de neurosciences en quelques secondes au lieu de mois en labo, sans mobiliser de sujets humains. Code, mod&#232;le et papier sont en open source.</p><h3>Alibaba</h3><h4><a href="iframe.php?url=https%3A%2F%2Fqwen.ai%2Fblog%3Fid%3Dqwen3.6">Qwen3.6-Plus, taill&#233; pour les agents de code</a></h4><p>Alibaba lance Qwen3.6-Plus, un mod&#232;le h&#233;berg&#233; avec une fen&#234;tre de contexte d&#8217;un million de tokens, pens&#233; pour les workflows agentiques. Sur SWE-bench Verified, il atteint 78.8 juste derri&#232;re Claude Opus 4.5 (80.9). Le tout pour environ 0,29 $/M tokens en entr&#233;e, soit une fraction du prix des mod&#232;les occidentaux. Alibaba continue de grignoter l&#8217;&#233;cart avec les leaders, &#224; un tarif qui pourrait faire r&#233;fl&#233;chir.</p><h4><a href="iframe.php?url=https%3A%2F%2Fqwen.ai%2Fblog%3Fid%3Dqwen3.5-omni">Qwen3.5-Omni, le multimodal complet</a></h4><p>Alibaba sort Qwen3.5-Omni, un mod&#232;le capable de traiter texte, image, audio et vid&#233;o dans un m&#234;me pipeline. Disponible en trois tailles (Plus, Flash, Light), il g&#232;re 113 langues en reconnaissance vocale et 36 en synth&#232;se. Le mod&#232;le rivalise avec Gemini 3.1 Pro sur l&#8217;audio et la vid&#233;o. Changement notable : contrairement aux versions pr&#233;c&#233;dentes, Qwen3.5-Omni n&#8217;est pas open source. Alibaba bascule vers le propri&#233;taire, et &#231;a en dit long sur la direction du march&#233; chinois.</p><h3>Mistral</h3><h4><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Fvoxtral-tts">Voxtral TTS : Mistral se lance dans la voix</a></h4><p>Mistral entre sur le march&#233; du text-to-speech avec Voxtral TTS, un mod&#232;le multilingue capable de cloner une voix &#224; partir de trois secondes d&#8217;audio. Neuf langues support&#233;es, une latence de 70 ms et un prix API de 0,016 $ par millier de caract&#232;res, soit environ 73 % moins cher qu&#8217;ElevenLabs. Lors de tests d&#8217;&#233;coute avec des natifs, Voxtral l&#8217;emporte sur ElevenLabs Flash v2.5 en naturalit&#233;. Le mod&#232;le est disponible en open weights sur Hugging Face (licence non-commerciale). Mistral, jusqu&#8217;ici cantonn&#233; au texte, vient clairement chercher ElevenLabs sur son terrain.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!MYEV!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21MYEV%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19744,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!MYEV!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!MYEV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F84956a81-70a5-48c1-bc0b-629e7675f1fa_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.npr.org%2F2026%2F03%2F26%2Fnx-s1-5762971%2Fjudge-temporarily-blocks-anthropic-ban">Une juge f&#233;d&#233;rale torpille le ban d&#8217;Anthropic par le Pentagone</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fwww.bloomberg.com%2Fnews%2Farticles%2F2026-03-27%2Fclaude-ai-maker-anthropic-said-to-weigh-ipo-as-soon-as-october">Anthropic vise une IPO &#224; 60 milliards, OpenAI aussi sur les starting-blocks</a></p><p>&#127963;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.whitehouse.gov%2Freleases%2F2026%2F03%2Fpresident-donald-j-trump-unveils-national-ai-legislative-framework%2F">Trump veut une loi IA f&#233;d&#233;rale unique pour court-circuiter les &#201;tats</a></p><p>&#127917; <a href="iframe.php?url=https%3A%2F%2Fwww.cnbc.com%2F2026%2F03%2F24%2Fmusk-xai-sued-baltimore-grok-deepfake-porn.html">Baltimore attaque xAI en justice : Grok a g&#233;n&#233;r&#233; 4,4 millions de deepfakes pornographiques</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!QiYv!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21QiYv%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19538,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!QiYv!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!QiYv!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F7173aaa5-9d58-4f61-8579-1f9e98ec6cba_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><strong><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fvectorless-rag-your-rag-pipeline-doesnt-need-a-vector-database-0a0839feabd9">Vectorless RAG: Your RAG Pipeline Doesn&#8217;t Need a Vector Database</a></strong></h4><p>Le sc&#233;nario classique : un pipeline RAG remonte un chunk qui parle du bon sujet mais pas de la bonne clause, et le mod&#232;le r&#233;pond avec aplomb &#224; c&#244;t&#233; de la plaque. Le probl&#232;me n&#8217;est pas le LLM, c&#8217;est la recherche par similarit&#233; vectorielle qui confond &#171; texte ressemblant &#187; et &#171; texte pertinent &#187;.</p><p>Le Vectorless RAG propose de supprimer enti&#232;rement le pipeline embeddings/vector DB/chunking. &#192; la place, on construit un arbre hi&#233;rarchique du document (un JSON avec titres, plages de pages et r&#233;sum&#233;s g&#233;n&#233;r&#233;s par LLM pour chaque section), puis on laisse le mod&#232;le raisonner sur cette structure pour d&#233;cider quelle section ouvrir. Exactement comme un analyste humain qui consulte la table des mati&#232;res avant de lire les 200 pages.</p><p>L&#8217;article d&#233;taille une impl&#233;mentation compl&#232;te avec PageIndex, une lib open source qui tient en une cinquantaine de lignes de Python. Deux appels LLM suffisent : un pour naviguer l&#8217;arbre, un pour g&#233;n&#233;rer la r&#233;ponse. Le gain principal n&#8217;est pas la vitesse (c&#8217;est m&#234;me plus lent qu&#8217;une recherche vectorielle) mais la tra&#231;abilit&#233; : on sait exactement quel n&#339;ud a &#233;t&#233; s&#233;lectionn&#233; et pourquoi, ce qui change la donne pour les documents structur&#233;s o&#249; une erreur de retrieval a des cons&#233;quences r&#233;elles (contrats, rapports financiers, papers).</p><h4><a href="iframe.php?url=https%3A%2F%2Fwww.decodingai.com%2Fp%2Ffrom-12-agents-to-1-ai-agent-architecture-decision-guide">From 12 Agents to 1</a></h4><p>Les auteurs prennes un exemple concret : un client arrive avec une spec de douze agents sp&#233;cialis&#233;s pour sa plateforme marketing. Ils pr&#233;voient un orchestrateur, un analyseur de requ&#234;tes, un g&#233;n&#233;rateur de contenu, et ainsi de suite. R&#233;sultat apr&#232;s audit : un seul agent avec des outils bien choisis fait exactement le m&#234;me travail, en moins cher et plus facile &#224; d&#233;bugger.</p><p>Paul Iusztin et Louis-Fran&#231;ois Bouchard formalisent cette le&#231;on en un spectre &#224; trois niveaux. Premier niveau : le workflow, une cha&#238;ne d'appels LLM dans un ordre fixe comme un syst&#232;me de tickets qui classe, route, r&#233;dige et valide toujours dans le m&#234;me ordre. Deuxi&#232;me niveau : un agent unique qui d&#233;cide lui-m&#234;me quels outils utiliser, utile quand le chemin d&#233;pend de ce qu'on d&#233;couvre en cours de route (une API qui &#233;choue, des donn&#233;es manquantes). Troisi&#232;me niveau : le multi-agents, justifi&#233; quand on d&#233;passe 10-20 outils (la s&#233;lection se d&#233;grade au-del&#224;), quand on a besoin de vrai parall&#233;lisme, ou quand des fronti&#232;res de s&#233;curit&#233; l'imposent.</p><p>Le principe : rester le plus simple possible tant que le probl&#232;me est r&#233;solu. Chaque niveau vers la droite ajoute des tokens, de la latence et des points de d&#233;faillance.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!kRYz!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21kRYz%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19760,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!kRYz!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!kRYz!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5623ffe9-778f-47da-8e7a-c86862f50dbe_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fdeveloper.nvidia.com%2Fblog%2Fdeploying-disaggregated-llm-inference-workloads-on-kubernetes%2F">Deploying Disaggregated LLM Inference Workloads on Kubernetes</a></h4><p>Quand on envoie un prompt &#224; un LLM, deux choses se passent c&#244;t&#233; GPU. D'abord le &#171; prefill &#187; : le mod&#232;le dig&#232;re tout le prompt d'un coup, &#231;a demande beaucoup de calcul brut. Ensuite le &#171; decode &#187; : il g&#233;n&#232;re les tokens un par un, et l&#224; c'est surtout la m&#233;moire qui travaille. Le probl&#232;me, c'est que la plupart des serveurs d'inf&#233;rence font les deux sur les m&#234;mes GPU. R&#233;sultat : les cartes passent leur temps &#224; alterner entre deux modes de fonctionnement oppos&#233;s, et ne sont jamais utilis&#233;es &#224; plein.</p><p>L'id&#233;e de NVIDIA : s&#233;parer prefill et decode dans des services Kubernetes distincts, chacun avec ses propres GPU et son propre scaling. L'article d&#233;taille trois outils pour y arriver. Dynamo g&#232;re le runtime d'inf&#233;rence. KAI Scheduler place les pods intelligemment en tenant compte de la topologie r&#233;seau (les GPU qui communiquent doivent &#234;tre proches physiquement). Grove orchestre le tout via un fichier de config d&#233;claratif.</p><p>Le pi&#232;ge principal : si on ajoute des GPU au prefill sans renforcer le decode, les tokens g&#233;n&#233;r&#233;s s'accumulent en file d'attente. NVIDIA propose un autoscaler qui maintient le bon ratio entre les deux &#233;tapes, l&#224; o&#249; un scaling Kubernetes classique op&#232;re &#224; l'aveugle.</p><h4><a href="iframe.php?url=https%3A%2F%2Ftowardsdatascience.com%2Fbuilding-human-in-the-loop-agentic-workflows%2F">Building Human-In-The-Loop Agentic Workflows</a></h4><p>Laisser un agent encha&#238;ner les &#233;tapes en autonomie, &#231;a marche, jusqu&#8217;au jour o&#249; &#231;a ne marche plus. Le vrai probl&#232;me n&#8217;est pas l&#8217;autonomie en soi, c&#8217;est l&#8217;absence de points de contr&#244;le humain aux endroits qui comptent.</p><p>Kenneth Leung d&#233;taille comment LangGraph permet de poser ces garde-fous via un m&#233;canisme d&#8217;&#171; interrupts &#187; : des points de pause ins&#233;r&#233;s dans le flux d&#8217;ex&#233;cution o&#249; un humain peut approuver, rejeter ou modifier le r&#233;sultat avant que l&#8217;agent ne continue. Le concept repose sur trois briques : les interrupts eux-m&#234;mes (qui suspendent le graphe et exposent l&#8217;&#233;tat courant), les &#171; checkpointers &#187; (qui persistent l&#8217;&#233;tat pendant la pause, ici via SQLite), et un syst&#232;me de threads pour reprendre exactement l&#224; o&#249; l&#8217;ex&#233;cution s&#8217;est arr&#234;t&#233;e.</p><p>L&#8217;article pointe un pi&#232;ge subtil : placer un interrupt apr&#232;s une &#233;tape non d&#233;terministe (comme une recherche web) entra&#238;ne sa r&#233;-ex&#233;cution au moment de la reprise, avec potentiellement des r&#233;sultats diff&#233;rents de ceux valid&#233;s. L&#8217;interrupt fonctionne mieux comme portail avant une action irr&#233;versible, pas comme filet de s&#233;curit&#233; apr&#232;s un appel non reproductible.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!nlT9!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21nlT9%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19342,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!nlT9!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!nlT9!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1cc2f847-ee56-4845-a982-1e3e5b17045b_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fkatanemo%2Fplano">Plano : le proxy AI-native pour apps agentiques</a></h4><p>Plano centralise orchestration, routage entre mod&#232;les, guardrails et observabilit&#233; dans un proxy out-of-process pour agents IA. On d&#233;clare ses providers en YAML, le routage s&#8217;appuie sur un mod&#232;le maison de 4B param&#232;tres. Tracing OpenTelemetry int&#233;gr&#233;, mod&#233;ration par filter chains, support multi-providers. Construit sur Envoy par ses contributeurs historiques (6k+ stars, Apache-2.0).</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!vKJU!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21vKJU%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:22082,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192999514?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!vKJU!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!vKJU!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F365fa5c0-3717-4b85-8315-3afd9a2eab5c_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2603.23516">MSA : Memory Sparse Attention Scales End-to-End to 100M Tokens</a></h4><blockquote><p>Long-term memory is a cornerstone of human intelligence. Enabling AI to process lifetime-scale information remains a long-standing pursuit in<br>the field. Due to the constraints of full-attention architectures, the effective context length of large language models (LLMs) is typically<br>limited to 1M tokens. Existing approaches, such as hybrid linear attention, fixed-size memory states (e.g., RNNs), and external storage<br>methods like RAG or agent systems, attempt to extend this limit. However, they often suffer from severe precision degradation and rapidly<br>increasing latency as context length grows, an inability to dynamically modify memory content, or a lack of end-to-end optimization. These<br>bottlenecks impede complex scenarios like large-corpus summarization, Digital Twins, and long-history agent reasoning, while limiting memory<br>capacity and slowing inference. We present Memory Sparse Attention (MSA), an end-to-end trainable, efficient, and massively scalable memory<br>model framework. Through core innovations including scalable sparse attention and document-wise RoPE, MSA achieves linear complexity in both<br>training and inference while maintaining exceptional stability, exhibiting less than 9% degradation when scaling from 16K to 100M tokens.<br>Furthermore, KV cache compression, combined with Memory Parallel, enables 100M-token inference on 2xA800 GPUs. We also propose Memory<br>Interleaving to facilitate complex multi-hop reasoning across scattered memory segments. MSA significantly surpasses frontier LLMs,<br>state-of-the-art RAG systems, and leading memory agents in long-context benchmarks. These results demonstrate that by decoupling memory<br>capacity from reasoning, MSA provides a scalable foundation to endow general-purpose models with intrinsic, lifetime-scale memory.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Claude prend le contrôle de votre Mac, Cursor oublie de citer ses sources ]]></title><description><![CDATA[Le brief IA #75]]></description><link>https://humanframework.substack.com/p/claude-prend-le-controle-de-votre</link><guid isPermaLink="false">https://humanframework.substack.com/p/claude-prend-le-controle-de-votre</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 27 Mar 2026 07:30:50 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21eo7u%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!u5Y5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!u5Y5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!u5Y5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!u5Y5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21u5Y5%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/bfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:20896,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192233528?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!u5Y5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!u5Y5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!u5Y5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!u5Y5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fbfb3f629-568f-4aac-a4ee-fe962df460bc_1456x138.webp 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Google</h3><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgoogle-deepmind%2Fmeasuring-agi-cognitive-framework%2F">Google DeepMind veut mesurer la progression vers l&#8217;AGI</a></h4><p>Google DeepMind propose un cadre scientifique pour &#233;valuer les progr&#232;s vers l&#8217;AGI, bas&#233; sur 10 capacit&#233;s issues des sciences cognitives : perception, raisonnement, m&#233;moire, apprentissage, m&#233;tacognition, etc. L&#8217;id&#233;e : comparer les performances des mod&#232;les IA &#224; celles d&#8217;un panel repr&#233;sentatif d&#8217;adultes humains, capacit&#233; par capacit&#233;. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21eo7u%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!eo7u!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 424w, https://substackcdn.com/image/fetch/$s_!eo7u!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 848w, https://substackcdn.com/image/fetch/$s_!eo7u!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 1272w, https://substackcdn.com/image/fetch/$s_!eo7u!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21eo7u%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin" width="601" height="337.161" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:561,&quot;width&quot;:1000,&quot;resizeWidth&quot;:601,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Bubbles all connecting to the central bubble \&quot;Cognitive faculties\&quot;. Each bubble list a cognitive faculty.&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Bubbles all connecting to the central bubble &quot;Cognitive faculties&quot;. Each bubble list a cognitive faculty." title="Bubbles all connecting to the central bubble &quot;Cognitive faculties&quot;. Each bubble list a cognitive faculty." srcset="https://substackcdn.com/image/fetch/$s_!eo7u!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 424w, https://substackcdn.com/image/fetch/$s_!eo7u!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 848w, https://substackcdn.com/image/fetch/$s_!eo7u!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 1272w, https://substackcdn.com/image/fetch/$s_!eo7u!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F03fe3e4e-01be-4ebb-aa84-8cd64b1ad2ed_1000x561.bin 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Pour passer de la th&#233;orie &#224; la pratique, un hackathon Kaggle dot&#233; de 200 000 $ invite la communaut&#233; &#224; cr&#233;er les benchmarks manquants, notamment sur l&#8217;apprentissage et la cognition sociale. Une tentative bienvenue de mettre des chiffres sur un concept que tout le monde utilise sans le d&#233;finir.</p><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgoogle-labs%2Fstitch-ai-ui-design%2F">Stitch et vibe designing</a></h4><p>Google fait &#233;voluer Stitch, son outil de prototypage UI, en un canvas de design complet pour vide designer. On d&#233;crit un objectif business ou une intention utilisateur, et l&#8217;agent g&#233;n&#232;re plusieurs pistes de design. Les &#233;crans se connectent pour former des prototypes interactifs en quelques clics. Nouveaut&#233; importante : DESIGN.md, un fichier markdown qui exporte les r&#232;gles de design vers d&#8217;autres outils, y compris les assistants de code via un serveur MCP. Google attaque frontalement le terrain de Figma.</p><h4><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Ftechnology%2Fai%2Flyria-3-pro%2F">Google lance Lyria 3 Pro, son mod&#232;le de g&#233;n&#233;ration musicale</a></h4><p>Google passe aux morceaux de 3 minutes avec Lyria 3 Pro, qui comprend la structure musicale : intro, couplets, refrains, ponts. Suno et Udio le faisaient d&#233;j&#224; depuis 2025, mais Google mise sur la distribution massive en int&#233;grant le mod&#232;le dans Gemini, Vertex AI, Google Vids et un nouvel outil baptis&#233; ProducerAI. La qualit&#233; vocale reste en retrait face &#224; Suno v5, mais l&#8217;&#233;cosyst&#232;me Google pourrait compenser. J&#8217;ai pu tester sur quelques exemples et c&#8217;est particuli&#232;rement bluffant.</p><h3>Anthropic</h3><h4><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fauto-mode">Anthropic introduit le mode auto pour Claude Code</a></h4><p>Claude Code proposait deux extr&#234;mes : valider chaque action &#224; la main, ou tout autoriser &#224; l&#8217;aveugle via <code>--dangerously-skip-permissions</code>. Le nouveau mode auto se place entre les deux. Un classifieur analyse chaque commande shell et &#233;criture de fichier avant ex&#233;cution, bloque les actions destructrices et laisse passer le reste. Concr&#232;tement, on peut enfin lancer une t&#226;che longue sans surveillance. </p><h4><a href="iframe.php?url=https%3A%2F%2Fx.com%2Fclaudeai%2Fstatus%2F2036195789601374705">Anthropic donne &#224; Claude le contr&#244;le de votre Mac</a></h4><p>Claude peut d&#233;sormais piloter votre bureau macOS : ouvrir des apps, cliquer, remplir des formulaires, naviguer sur le web. Le principe : des captures d&#8217;&#233;cran en boucle, analys&#233;es par le mod&#232;le multimodal, qui d&#233;cide de la prochaine action. Quand un connecteur existe (Slack, Google Workspace), Claude l&#8217;utilise en priorit&#233; sinon, il prend la main sur clavier et souris. Disponible en research preview pour les abonn&#233;s Pro et Max, avec un syst&#232;me de permissions avant chaque nouvel acc&#232;s. On peut m&#234;me lancer une t&#226;che depuis l&#8217;iPhone et retrouver le r&#233;sultat sur son Mac. macOS uniquement pour l&#8217;instant.</p><h3>OpenAI</h3><h4><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fopenai-to-acquire-astral%2F">OpenAI met la main sur Astral et ses outils Python</a></h4><p>OpenAI rach&#232;te Astral, la startup derri&#232;re uv, Ruff et ty, trois outils open source devenus incontournables pour les d&#233;veloppeurs Python. uv &#224; lui seul cumule plus de 126 millions de t&#233;l&#233;chargements mensuels. L&#8217;&#233;quipe de Charlie Marsh rejoindra Codex, la plateforme de coding d&#8217;OpenAI qui revendique 2 millions d&#8217;utilisateurs actifs par semaine. OpenAI promet de maintenir les projets open source, mais la communaut&#233; reste prudente. Quand un g&#233;ant absorbe l&#8217;outillage que tout le monde utilise, la question n&#8217;est jamais &#171; si &#187; &#231;a change, mais &#171; quand &#187;.</p><h3>NVIDIA</h3><h4><a href="iframe.php?url=https%3A%2F%2Fresearch.nvidia.com%2Flabs%2Fnemotron%2Fnemotron-cascade-2%2F">NVIDIA lance Nemotron-Cascade 2</a></h4><p>NVIDIA publie Nemotron-Cascade 2, un mod&#232;le open-weight capable de d&#233;crocher l&#8217;or aux olympiades internationales de maths (IMO), d&#8217;informatique (IOI) et de programmation (ICPC). Il est vingt fois plus petit que DeepSeek V3.2, le seul autre mod&#232;le &#8220;open&#8221; &#224; avoir atteint ce niveau. NVIDIA livre les poids et toutes les donn&#233;es d&#8217;entra&#238;nement. </p><h4><a href="iframe.php?url=https%3A%2F%2Fdeveloper.nvidia.com%2Fnemo-agent-toolkit">NVIDIA lance son Agent Toolkit, une bo&#238;te &#224; outils open source pour agents IA</a></h4><p>NVIDIA profite de GTC pour d&#233;voiler son Agent Toolkit, un ensemble de briques open source pour construire des agents IA autonomes en entreprise. Le kit comprend OpenShell, un runtime qui impose des garde-fous de s&#233;curit&#233; et de confidentialit&#233; aux agents, et AI-Q, un blueprint de recherche agentique qui a bri&#232;vement domin&#233; le leaderboard DeepResearch Bench. Adobe, Salesforce, SAP ou encore Atlassian int&#232;grent d&#233;j&#224; ces briques. </p><h3>Mistral</h3><h4><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Fmistral-small-4">Mistral Small 4 : un seul mod&#232;le pour tout faire</a></h4><p>Mistral fusionne ses mod&#232;les sp&#233;cialis&#233;s (raisonnement, multimodal, code) en un seul : Small 4, open source sous licence Apache 2.0. Le mod&#232;le promet 40 % de latence en moins et un d&#233;bit tripl&#233; par rapport &#224; son pr&#233;d&#233;cesseur. Son argument principal : un param&#232;tre ajustable pour doser l&#8217;effort de raisonnement selon la t&#226;che, du chat rapide &#224; la r&#233;solution complexe. Face &#224; Llama 4 et Qwen 3, Mistral joue la carte de l&#8217;efficacit&#233; par token et du co&#251;t r&#233;duit. Le pari de la consolidation plut&#244;t que de la multiplication des mod&#232;les.</p><h3>Cursor</h3><h4><a href="iframe.php?url=https%3A%2F%2Fcursor.com%2Fblog%2Fcomposer-2">Cursor lance Composer 2, son mod&#232;le de code bas&#233; sur Kimi K2.5</a></h4><p>Cursor a d&#233;voil&#233; Composer 2, son nouveau mod&#232;le de code affich&#233; &#224; 0,50 $/M tokens en entr&#233;e. Probl&#232;me : l&#8217;&#233;quipe a &#171; oubli&#233; &#187; de mentionner que le mod&#232;le repose sur Kimi K2.5 de Moonshot AI. Un d&#233;veloppeur a trouv&#233; l&#8217;identifiant <code>kimi-k2p5</code> dans l&#8217;API, for&#231;ant Cursor &#224; reconna&#238;tre la base open source. L&#8217;entreprise assure que trois quarts du compute viennent de son propre entra&#238;nement par renforcement. Reste que la licence de Kimi exige une attribution claire au-del&#224; de 20 M$ de revenus mensuels un seuil que Cursor d&#233;passe largement. </p><h3>Figma</h3><h4><a href="iframe.php?url=https%3A%2F%2Fwww.figma.com%2Fblog%2Fthe-figma-canvas-is-now-open-to-agents%2F">Figma ouvre son canvas aux agents IA</a></h4><p>Figma lance <code>use_figma</code> en open beta : les agents IA peuvent d&#233;sormais cr&#233;er et modifier des designs directement sur le canvas, en s&#8217;appuyant sur les composants et variables du design system existant. L&#8217;outil fonctionne avec Claude Code, Codex, Cursor et Copilot via le protocole MCP. Autre ajout notable : les &#171; skills &#187;, des fichiers markdown qui encadrent le comportement des agents pour respecter les conventions de chaque &#233;quipe. Gratuit pendant la beta, le service deviendra payant. Un pas concret vers le bin&#244;me designer-agent, o&#249; Figma se positionne comme le terrain de jeu commun.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!9STf!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!9STf!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!9STf!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!9STf!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%219STf%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19978,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192233528?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!9STf!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!9STf!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!9STf!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!9STf!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F16e4dbd5-c472-4f43-b6d2-96f354b595b2_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#127917; <a href="iframe.php?url=https%3A%2F%2Fedition.cnn.com%2F2026%2F03%2F13%2Fpolitics%2Fjames-talarico-ai-deepfake-republicans-midterms">Quand un parti politique diffuse un deepfake de 85 secondes de son adversaire</a></p><p>&#129514; <a href="iframe.php?url=https%3A%2F%2Fmetr.org%2Fnotes%2F2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main%2F">La moiti&#233; des PRs qui &#171; passent &#187; SWE-bench seraient recal&#233;es par de vrais mainteneurs</a></p><p>&#128181; <a href="iframe.php?url=https%3A%2F%2Fnvidianews.nvidia.com%2Fnews%2Fopenai-and-nvidia-announce-strategic-partnership-to-deploy-10gw-of-nvidia-systems">NVIDIA met 100 milliards de dollars sur la table pour OpenAI</a></p><p>&#9878;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.reuters.com%2Ftechnology%2Fmicrosoft-weighs-legal-action-over-50-billion-amazon-openai-cloud-deal-ft-2026-03-18%2F">Microsoft menace de poursuivre OpenAI pour son deal cloud &#224; 50 Md$ avec Amazon</a></p><p>&#128202; <a href="iframe.php?url=https%3A%2F%2Fwww.answer.ai%2Fposts%2F2026-03-12-so-where-are-all-the-ai-apps.html">L&#8217;IA rend les devs 10x plus productifs ? Les donn&#233;es PyPI disent non</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!p3nV!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!p3nV!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!p3nV!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!p3nV!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21p3nV%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/dab85041-1a73-41a0-9338-7124c818987b_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21538,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192233528?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!p3nV!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!p3nV!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!p3nV!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!p3nV!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fdab85041-1a73-41a0-9338-7124c818987b_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fengineering%2Fharness-design-long-running-apps">Anthropic shares how Claude handles long-running autonomous coding tasks</a></h4><p>Faire coder un agent pendant des heures sans supervision, &#231;a finit invariablement par d&#233;raper : le mod&#232;le perd le fil, valide son propre code bancal, ou boucle en fin de contexte. Anthropic publie un retour d'exp&#233;rience d&#233;taill&#233; sur la fa&#231;on dont ils ont r&#233;solu ce probl&#232;me.</p><p>Le concept cl&#233;, c'est le &#171; harness &#187; : un programme qui orchestre plusieurs agents sp&#233;cialis&#233;s autour du LLM, chacun avec un r&#244;le pr&#233;cis. Ici, trois agents se relaient. Un planner transforme un prompt d'une phrase en spec produit compl&#232;te. Un generator &#233;crit le code. Et un evaluator teste l'application via Playwright, clique sur les boutons, v&#233;rifie les API, et renvoie un rapport de bugs concret. L'astuce, inspir&#233;e des GANs : ne jamais laisser un agent juger son propre travail. Un LLM qui note son output est syst&#233;matiquement complaisant, alors qu'un &#233;valuateur s&#233;par&#233; est bien plus facile &#224; rendre exigeant.</p><p>Avec Opus 4.6, l'architecture s'est simplifi&#233;e : le mod&#232;le tient la coh&#233;rence sur des sessions de plus de deux heures d'affil&#233;e, ce qui a permis de supprimer tout le d&#233;coupage en &#233;tapes interm&#233;diaires que la version pr&#233;c&#233;dente imposait. R&#233;sultat : un studio de musique complet (Digital Audio Workstation) g&#233;n&#233;r&#233; en 4 heures pour 124 $. Cher, mais fonctionnel l&#224; o&#249; un agent seul produit du code cass&#233; en 20 minutes. La le&#231;on &#224; retenir : chaque brique d'un harness compense une limite du mod&#232;le, et ces limites bougent &#224; chaque release. Ce qui &#233;tait indispensable hier devient du poids mort demain.</p><h4><a href="iframe.php?url=https%3A%2F%2Fskylarbpayne.com%2Fposts%2Fdspy-engineering-patterns%2F">If DSPy is So Great, Why Isn&#8217;t Anyone Using It?</a></h4><p>Tout projet LLM suit la m&#234;me trajectoire : on commence par un appel OpenAI en dur, puis on externalise les prompts, on ajoute des structured outputs, du retry, du RAG, des evals. Six mois plus tard, on maintient un framework maison bancal. L'auteur appelle &#231;a la &#171; loi de Khattab &#187; : tout syst&#232;me IA suffisamment complexe finit par r&#233;impl&#233;menter la moiti&#233; de DSPy, en moins bien.<br><br>DSPy est un framework qui structure les appels LLM autour de trois briques : des Signatures pour typer les entr&#233;es et sorties de chaque appel, des Modules pour les cha&#238;ner comme des fonctions composables, et des Optimizers qui r&#233;&#233;crivent automatiquement les prompts pour am&#233;liorer les r&#233;sultats sans intervention manuelle. Changer de mod&#232;le tient en une ligne, et l'&#233;valuation est native.<br><br>Alors pourquoi seulement 4,7 millions de t&#233;l&#233;chargements mensuels contre 222 millions pour LangChain ? Parce que DSPy impose de penser en abstractions avant d'avoir ressenti la douleur du code spaghetti, et la pression du shipping rend cet investissement difficile &#224; justifier. Le conseil pragmatique de l'auteur : m&#234;me sans adopter DSPy, appliquer ses patterns d&#232;s le d&#233;part : I/O typ&#233;es, prompts s&#233;par&#233;s du code, evals d&#232;s le premier jour. Ce sont des fondamentaux d'ing&#233;nierie logicielle qu'on finira par impl&#233;menter de toute fa&#231;on.</p><h4><a href="iframe.php?url=https%3A%2F%2Famandeepsp.github.io%2Fblog%2Fhigh-dims%2F">The Boon of Dimensionality</a></h4><p>La &#171; mal&#233;diction de la dimensionnalit&#233; &#187; fait partie du folklore ML : plus on ajoute de dimensions, plus les donn&#233;es deviennent &#233;parses et les distances perdent leur sens. Mais l'histoire a un revers qu'on raconte moins souvent, et il est bien plus r&#233;jouissant.<br><br>En reprenant les intuitions g&#233;om&#233;triques popularis&#233;es par 3Blue1Brown, l'auteur montre ce qui se passe r&#233;ellement dans les espaces &#224; haute dimension : le volume d'une hypersph&#232;re s'effondre pass&#233; cinq dimensions, sa masse se concentre quasi int&#233;gralement en surface, et deux points tir&#233;s au hasard sont presque toujours orthogonaux entre eux. C'est contre-intuitif, mais c'est pr&#233;cis&#233;ment ce qui rend les embeddings viables, on peut caser des millions de concepts dans un espace de 768 dimensions sans qu'ils se chevauchent. s.<br><br>L'article ouvre aussi sur la superposition dans les r&#233;seaux de neurones, o&#249; un r&#233;seau &#224; m neurones parvient &#224; encoder bien plus de m features distinctes en exploitant cette quasi-orthogonalit&#233; naturelle des espaces denses. Une lecture qui remet les pendules &#224; l'heure sur un fondamental qu'on cite souvent sans le comprendre vraiment.</p><h4><a href="iframe.php?url=https%3A%2F%2Fblog.skypilot.co%2Fscaling-autoresearch%2F">Scaling Karpathy&#8217;s Autoresearch: What Happens When the Agent Gets a GPU Cluster</a></h4><p>Autoresearch de Karpathy, j&#8217;en parlais la semaine derni&#232;re et le principe est simple : un agent qui boucle sur un script d'entra&#238;nement de machine learning, teste une modification, garde ce qui marche et recommence. Efficace, mais s&#233;quentiel car utilise un GPU, une exp&#233;rience &#224; la fois, et beaucoup de temps perdu &#224; attendre. L'&#233;quipe SkyPilot a voulu voir ce qui se passe quand on retire ce goulet d'&#233;tranglement en donnant &#224; Claude Code l'acc&#232;s &#224; 16 GPUs sur un cluster Kubernetes.</p><p>Le r&#233;sultat va au-del&#224; d'un simple gain de d&#233;bit. Avec 910 exp&#233;riences soumises en 8 heures, le parall&#233;lisme change fondamentalement la strat&#233;gie de recherche : au lieu du hill-climbing glouton (tester un param&#232;tre, garder le meilleur, recommencer) l'agent lance des grilles de 10 &#224; 13 exp&#233;riences par vague, ce qui lui permet de capturer les interactions entre param&#232;tres qu'une approche s&#233;rielle raterait. C'est d'ailleurs comme &#231;a que la d&#233;couverte la plus marquante est arriv&#233;e : en testant six tailles de mod&#232;le simultan&#233;ment (de 384 &#224; 768 dimensions internes), l'agent a montr&#233; qu'&#233;largir le r&#233;seau am&#233;liorait davantage la loss de validation que tous les r&#233;glages d'hyperparam&#232;tres combin&#233;s (learning rate, weight decay, batch size). En s&#233;quentiel, cette exploration aurait pris 30 minutes. En parall&#232;le, une seule vague de 5 minutes.</p><p>Autre comportement inattendu : l'agent disposait de deux types de GPU (des H100 et des H200, plus rapides) sans qu'on lui signale la diff&#233;rence. Apr&#232;s quelques vagues, il a remarqu&#233; tout seul que les m&#234;mes configs donnaient de meilleurs r&#233;sultats sur H200,  logique, puisque plus de vitesse signifie plus d'&#233;tapes d'entra&#238;nement dans le budget de 5 minutes. Il a alors invent&#233; sa propre strat&#233;gie : tester les id&#233;es en masse sur les H100 bon march&#233;, puis valider les meilleures sur H200. Le genre de workflow qu'un chercheur concevrait d&#233;lib&#233;r&#233;ment et il a &#233;merg&#233; tout seul. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!2lZM!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!2lZM!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!2lZM!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!2lZM!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%212lZM%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:20094,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192233528?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!2lZM!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!2lZM!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!2lZM!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!2lZM!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F908dd3f8-e682-4073-b032-862f011877ea_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fevery.to%2Fguides%2Fagent-native">Agent-native Architectures: How to Build Apps After Code Ends</a></h4><p>Un appel API derri&#232;re un bouton &#8220;g&#233;n&#233;rer avec l&#8217;IA&#8221;, un workflow cod&#233; en dur qui casse d&#232;s que le use case d&#233;vie : le pattern classique de l&#8217;int&#233;gration LLM a ses limites. Dan Shipper et Claude proposent une alternative radicale : l&#8217;architecture &#171; agent-native &#187;, o&#249; l&#8217;agent n&#8217;est plus un gadget greff&#233; sur l&#8217;app, mais le moteur central de chaque fonctionnalit&#233;.</p><p>Le principe fondateur tient en cinq piliers. D&#8217;abord la parit&#233; : tout ce que l&#8217;utilisateur fait via l&#8217;UI, l&#8217;agent doit pouvoir l&#8217;accomplir via ses outils. Ensuite la granularit&#233; : les outils restent des primitives atomiques (read_file, write_file, bash), et les &#171; features &#187; deviennent des outcomes d&#233;crits dans un prompt, atteints par l&#8217;agent en boucle. La composabilit&#233; d&#233;coule naturellement des deux premiers : ajouter une fonctionnalit&#233; revient &#224; &#233;crire un nouveau prompt, pas &#224; refactorer du code. L&#8217;approche emprunte directement au fonctionnement de Claude Code, g&#233;n&#233;ralis&#233; &#224; des domaines non techniques.</p><p>Le guide va loin sur les patterns d&#8217;impl&#233;mentation : fichiers comme interface universelle, context.md comme m&#233;moire de session, d&#233;couverte dynamique de capacit&#233;s plut&#244;t que mapping statique d&#8217;API. Le parti pris &#171; files-first &#187; sur mobile avec iCloud est discutable, mais la philosophie sous-jacente est solide : concevoir pour ce que les agents savent raisonner, c&#8217;est-&#224;-dire ce qu&#8217;un humain comprendrait aussi.</p><h4><a href="iframe.php?url=https%3A%2F%2Fmedium.com%2Fpinterest-engineering%2Fbuilding-an-mcp-ecosystem-at-pinterest-d881eb4c16f1">Building an MCP Ecosystem at Pinterest</a></h4><p>Les retours d&#8217;exp&#233;rience sur la mise en place d&#8217;un MCP &#224; l&#8217;&#233;chelle d&#8217;une entreprise enti&#232;re restent rares. Pinterest vient d&#8217;en publier un, et il m&#233;rite qu&#8217;on s&#8217;y attarde.</p><p>L&#8217;&#233;quipe a fait le choix de serveurs MCP cloud-hosted et sp&#233;cialis&#233;s (Presto, Spark, Knowledge) plut&#244;t qu&#8217;un monolithe fourre-tout. Chaque serveur g&#232;re un p&#233;rim&#232;tre fonctionnel restreint avec ses propres contr&#244;les d&#8217;acc&#232;s, le tout orchestr&#233; par un registre central qui sert de source de v&#233;rit&#233;. C&#244;t&#233; s&#233;curit&#233;, l&#8217;architecture repose sur deux couches d&#8217;authentification : des JWTs utilisateur pour l&#8217;attribution fine et des identit&#233;s mesh SPIFFE pour les flux service-to-service. Le gating par business group sur les serveurs sensibles (Presto notamment) &#233;vite l&#8217;&#233;cueil classique du &#171; tout le monde a acc&#232;s parce que c&#8217;est interne &#187;.</p><p>Ce qui frappe, c&#8217;est la maturit&#233; op&#233;rationnelle : 66 000 invocations mensuelles, 844 utilisateurs actifs, et une estimation de 7 000 heures-ing&#233;nieur &#233;conomis&#233;es par mois. Les serveurs MCP sont int&#233;gr&#233;s directement dans le chat interne et les IDE, et pas rel&#233;gu&#233;s &#224; un PoC que personne n&#8217;utilise.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Qs0K!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!Qs0K!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!Qs0K!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!Qs0K!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Qs0K%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:19662,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192233528?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Qs0K!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!Qs0K!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!Qs0K!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!Qs0K!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4c05b367-08b5-4aad-8903-99fb6df0d92e_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Ftobi%2Fqmd">tobi/qmd</a></h4><p>Un moteur de recherche local pour ses docs, notes de r&#233;union et bases de connaissances en markdown. Sign&#233; Tobi Lutke (fondateur de Shopify), QMD combine recherche full-text (BM25), recherche s&#233;mantique et re-ranking par LLM, le tout tournant int&#233;gralement en local sans envoyer ses donn&#233;es nulle part. L'outil expose aussi un serveur MCP, ce qui en fait un compagnon naturel pour Claude Code ou Claude Desktop. 16k+ stars sur GitHub.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Flouisfb01%2Fai-engineering-cheatsheets">AI Engineering Cheatsheets</a></h4><p>Trois cheatsheets open source de Louis-Fran&#231;ois Bouchard (Towards AI), directement injectables dans Claude ou tout autre LLM comme contexte de travail : un playbook avec des tables de d&#233;cision pour choisir technique, mod&#232;le et strat&#233;gie de prompting, un guide d'architecture agent (workflow simple vs multi-agent), et un template anti-slop pour obtenir un ton humain en sortie de LLM. Le repo est encore jeune, mais le format &#171; table de d&#233;cision &#187; est actionnable imm&#233;diatement.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!VMkA!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!VMkA!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!VMkA!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!VMkA!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21VMkA%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:23490,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/192233528?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!VMkA!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!VMkA!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!VMkA!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!VMkA!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F81df8929-1f54-4800-acc7-641537dab36b_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fresearch.google%2Fblog%2Fturboquant-redefining-ai-efficiency-with-extreme-compression%2F">TurboQuant : red&#233;finir l&#8217;efficacit&#233; des LLMs par la compression extr&#234;me</a></h4><p>Le key-value cache reste l&#8217;un des goulots d&#8217;&#233;tranglement les plus tenaces quand on fait tourner des LLMs sur des contextes longs. Google Research s&#8217;attaque au probl&#232;me avec TurboQuant, pr&#233;sent&#233; &#224; ICLR 2026, une m&#233;thode de quantization qui compresse le KV cache &#224; 3 bits sans perte de pr&#233;cision mesurable.</p>]]></content:encoded></item><item><title><![CDATA[Le brief IA #74]]></title><description><![CDATA[OpenAI, Anthropic et Microsoft se battent pour votre bureau.]]></description><link>https://humanframework.substack.com/p/le-brief-ia-74</link><guid isPermaLink="false">https://humanframework.substack.com/p/le-brief-ia-74</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Wed, 18 Mar 2026 07:55:23 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Fyoutube%2Fw_728%2Cc_limit%2FcIctgHKEeMA" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bQGO%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:781604,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h5>Anthropic publie une &#233;tude sur l'impact (encore limit&#233;) de l'IA sur l'emploi</h5><p>Anthropic <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fresearch%2Flabor-market-impacts">a analys&#233; les donn&#233;es du march&#233; du travail</a> am&#233;ricain en croisant ses propres donn&#233;es d'usage de Claude avec les statistiques f&#233;d&#233;rales. R&#233;sultat : pas de hausse significative du ch&#244;mage dans les m&#233;tiers les plus expos&#233;s &#224; l'IA depuis fin 2022. Les programmeurs, service client et analystes financiers sont en premi&#232;re ligne, mais les effets restent pour l'instant statistiquement invisibles. Seul signal : l'embauche des 22-25 ans ralentit dans les professions expos&#233;es, sugg&#233;rant que l'IA freine l'entr&#233;e sur le march&#233; plut&#244;t qu'elle ne pousse dehors. Un constat qui pourrait vite changer &#224; mesure que l'adoption s'acc&#233;l&#232;re.</p><h5>Claude Code s'&#233;toffe : code review multi-agent et mode vocal</h5><p>Grosse semaine pour Claude Code, Anthropic int&#232;gre plusieurs nouveaux &#233;l&#233;ments. En particulier, <a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fcode-review">un syst&#232;me de review automatique des pull requests</a> : plusieurs agents IA analysent le code en parall&#232;le, v&#233;rifient chaque bug d&#233;tect&#233; avant de le signaler, et classent les probl&#232;mes par s&#233;v&#233;rit&#233;. En interne, le taux de PR recevant des commentaires substantifs est pass&#233; de 16 % &#224; 54 %. Sur les grosses PR (1000+ lignes), 84 % r&#233;v&#232;lent des probl&#232;mes, avec 7,5 issues en moyenne. Deuxi&#232;me ajout utile : un <a href="iframe.php?url=https%3A%2F%2Fx.com%2Ftrq212%2Fstatus%2F2028628570692890800">mode vocal activ&#233; via &#8216;/voice&#8217;</a>. Push-to-talk avec la barre espace, pas d'&#233;coute permanente, pas de surco&#251;t.</p><h5>Copilot Cowork et add-ins Excel/PowerPoint</h5><p>Double offensive pour Claude dans l'univers Microsoft. D'un c&#244;t&#233;, Microsoft lance <a href="iframe.php?url=https%3A%2F%2Fwww.microsoft.com%2Fen-us%2Fmicrosoft-365%2Fblog%2F2026%2F03%2F09%2Fcopilot-cowork-a-new-way-of-getting-work-done%2F">Copilot Cowork</a>, un agent multi-&#233;tapes pour Outlook, Teams et Excel et c'est la technologie Claude d'Anthropic qui propulse le moteur de raisonnement, pas OpenAI. L'outil transforme une consigne simple en plan d'action avec points de contr&#244;le. GA pr&#233;vue le 1er mai, int&#233;gr&#233; au forfait M365 E7 &#224; 99 $/utilisateur/mois. De l'autre, <a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fclaude-excel-powerpoint-updates">Anthropic met &#224; jour ses propres add-ins Claude pour Excel et PowerPoint</a>. </p><div id="youtube2-cIctgHKEeMA" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;cIctgHKEeMA&quot;,&quot;startTime&quot;:&quot;1s&quot;,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FcIctgHKEeMA%3Fstart%3D1s%26amp%3Brel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><p>Les deux applications partagent d&#233;sormais le contexte d'une m&#234;me conversation : Claude peut lire un tableur, construire une analyse financi&#232;re, puis mettre &#224; jour une pr&#233;sentation sans r&#233;explication. Les &#171; Skills &#187; permettent d'enregistrer un workflow complet en un clic pour le rendre r&#233;utilisable par toute l'&#233;quipe. Anthropic joue sur les deux tableaux, &#224; l'int&#233;rieur de Copilot et en parall&#232;le avec ses propres add-ins.</p><h3>OpenAI</h3><h5><strong>GPT-5.4</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-gpt-5-4%2F">OpenAI sort GPT-5.4</a>, son nouveau mod&#232;le phare, disponible dans ChatGPT, l'API et Codex. Au menu : contexte d'un million de tokens, computer use natif (une premi&#232;re pour un mod&#232;le g&#233;n&#233;raliste !) et une API de recherche d'outils qui r&#233;duit la consommation de tokens de 47 % sur les workflows multi-outils. Le prix monte l&#233;g&#232;rement &#224; 2,50 $/15 $ par million de tokens en entr&#233;e/sortie (contre 1,75 $/14 $ pour GPT-5.2). <a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fintroducing-gpt-5-4-mini-and-nano%2F">Les variantes mini et nano sont aussi disponibles</a>.</p><h5><strong>ChatGPT et apprentissage </strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fnew-ways-to-learn-math-and-science-in-chatgpt%2F">OpenAI int&#232;gre des visualisations interactives dans ChatGPT</a> pour explorer formules et concepts scientifiques. Concr&#232;tement, les utilisateurs peuvent manipuler des variables et observer en temps r&#233;el l'effet sur les graphiques. Plus de 70 sujets sont couverts au lancement, du th&#233;or&#232;me de Pythagore &#224; la loi des gaz parfaits, pour un niveau lyc&#233;e et sup&#233;rieur. La fonctionnalit&#233; est accessible sur tous les forfaits, gratuit compris. OpenAI revendique 140 millions d'utilisateurs hebdomadaires sur des cas d'apprentissage. Reste &#224; voir si ces widgets feront mieux que les dizaines d'applis &#233;ducatives d&#233;j&#224; sur le march&#233;.</p><h3>Google</h3><h5><strong>Gemini dans Maps</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Fproducts-and-platforms%2Fproducts%2Fmaps%2Fask-maps-immersive-navigation%2F">Google int&#232;gre Gemini dans Maps</a> avec deux nouveaut&#233;s. &#171; Ask Maps &#187; permet de poser des questions en langage naturel et d&#8217;obtenir des r&#233;ponses contextualis&#233;es. C&#244;t&#233; navigation, &#171; Immersive Navigation &#187; affiche une vue 3D du trajet avec b&#226;timents, feux et voies mat&#233;rialis&#233;es, g&#233;n&#233;r&#233;e &#224; partir d'images Street View. D&#233;ploiement lanc&#233; aux &#201;tats-Unis et en Inde, vous vous en doutiez mais l&#8217;Europe attendra.</p><h5><strong>Embedding 2</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-embedding-2%2F">Google sort Gemini Embedding 2</a>, un mod&#232;le d'embedding qui unifie texte, images, vid&#233;o, audio et PDF dans un seul espace vectoriel. L'int&#233;r&#234;t : une seule requ&#234;te peut m&#233;langer les modalit&#233;s, ce qui simplifie les pipelines RAG multimodaux. Sur les benchmarks vid&#233;o et image, aucun concurrent (OpenAI, Amazon Nova 2) ne propose d'&#233;quivalent natif. Disponible en preview via Gemini API et Vertex AI.</p><h5>Gemini dans workspace </h5><p><a href="iframe.php?url=https%3A%2F%2Fworkspace.google.com%2Fblog%2Fproduct-announcements%2Freimagining-content-creation">Google int&#232;gre Gemini en profondeur dans Docs, Sheets, Slides et Drive. Le plus notable</a> : Sheets peut d&#233;sormais construire un tableur complet &#224; partir d'une description en langage naturel, en croisant donn&#233;es de Gmail, Drive et du web. C&#244;t&#233; Slides, l'IA g&#233;n&#232;re des pr&#233;sentations enti&#232;res &#224; partir d'un simple prompt. Drive, lui, se mue en base de connaissances interrogeable. Le tout r&#233;serv&#233; pour l'instant aux abonn&#233;s AI Pro et Ultra, en anglais uniquement. </p><h3>Microsoft</h3><h5>Copilot Health</h5><p><a href="iframe.php?url=https%3A%2F%2Fmicrosoft.ai%2Fnews%2Fintroducing-copilot-health%2F">Microsoft int&#232;gre un espace d&#233;di&#233; &#224; la sant&#233; dans Copilot</a>. Le principe : centraliser dossiers m&#233;dicaux, r&#233;sultats de labo et donn&#233;es de plus de 50 wearables (Apple Health, Oura, Fitbit) pour en tirer des analyses personnalis&#233;es. Un panel de 230 m&#233;decins dans 24 pays supervise le projet, et la certification ISO 42001 est d&#233;j&#224; obtenue. Pour l'instant, acc&#232;s sur liste d'attente, r&#233;serv&#233; aux &#201;tats-Unis. </p><div id="youtube2-61JHf0K1zKc" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;61JHf0K1zKc&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2F61JHf0K1zKc%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!7GZY!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!7GZY!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!7GZY!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!7GZY!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%217GZY%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21220,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/191268393?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!7GZY!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 424w, https://substackcdn.com/image/fetch/$s_!7GZY!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 848w, https://substackcdn.com/image/fetch/$s_!7GZY!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 1272w, https://substackcdn.com/image/fetch/$s_!7GZY!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5c56bfcd-6bdd-41c0-b454-aee6c86602ce_1456x140.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#129489;&#8205;&#128188; <a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fkarpathy%2Fjobs">Karpathy open-source une carte interactive de l'exposition des m&#233;tiers &#224; l'IA</a></p><p>&#129297; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F03%2F09%2Fyann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models%2F">Le pari &#224; 1 milliard de Yann LeCun contre les LLMs</a></p><p>&#129309; <a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fopenai-to-acquire-promptfoo%2F">OpenAI rach&#232;te Promptfoo pour s&#233;curiser ses agents</a></p><p>&#127918; <a href="iframe.php?url=https%3A%2F%2Fwww.nvidia.com%2Fen-us%2Fgeforce%2Fnews%2Fdlss5-breakthrough-in-visual-fidelity-for-games%2F">Le photor&#233;alisme dans les jeux vid&#233;o franchit une nouvelle &#233;tape avec le DLSS 5</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21lNev%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:764352,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Fe10v.me%2Fpython-packages-for-ab-test-analysis%2F">Comparing Python packages for A/B test analysis: tea-tasting, Pingouin, statsmodels, and SciPy</a></strong></h3><p>Quel package Python couvre r&#233;ellement le workflow complet d'exp&#233;rimentation ?</p><p>Ce comparatif met face &#224; face quatre biblioth&#232;ques (tea-tasting, Pingouin, statsmodels, SciPy) sur les t&#226;ches concr&#232;tes du quotidien : t-test de Welch, m&#233;triques de type ratio, intervalles de confiance relatifs, CUPED et correction pour tests multiples. Le constat est net. SciPy reste la fondation, mais demande beaucoup de code maison. statsmodels offre des briques solides pour qui accepte d'assembler soi-m&#234;me le pipeline. Pingouin simplifie les tests classiques dans un workflow pandas, mais d&#233;croche d&#232;s qu'on touche aux ratios ou au CUPED. tea-tasting, con&#231;u sp&#233;cifiquement pour l'A/B testing, est le seul &#224; proposer effets relatifs, CUPED sur les ratios et workflow sur statistiques agr&#233;g&#233;es &#171; out of the box &#187;.</p><p>&#192; noter : l'auteur est aussi le cr&#233;ateur de tea-tasting, ce qui colore forc&#233;ment l'analyse. Le comparatif reste honn&#234;te sur les crit&#232;res factuels, mais le biais de s&#233;lection est &#224; garder en t&#234;te. Pour les &#233;quipes qui industrialisent leurs exp&#233;rimentations, c'est une grille de lecture utile avant de choisir sa stack.</p><h3><strong><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fstructured-video-captioning-with-gemini-an-mma-analysis-use-case-bfbb8fd91a26">Structured Video Captioning with Gemini: An MMA Analysis Use Case</a></strong></h3><p>Analyser de la vid&#233;o frame par frame pour en extraire des &#233;v&#233;nements structur&#233;s, c&#8217;est le genre de t&#226;che qui relevait encore r&#233;cemment du pipeline un peu custom avec d&#233;tection d&#8217;objets, tracking et post-traitement. Avec les VLMs capables d&#8217;ing&#233;rer directement du flux vid&#233;o, la choses sont diff&#233;rentes.</p><p>Ce tutoriel propose d&#8217;utiliser Gemini pour analyser des combats de MMA seconde par seconde. L&#8217;approche repose sur un workflow multi-agent : un premier prompt extrait une timeline brute des actions, puis des &#171; sp&#233;cialistes &#187; (striking, grappling, submission, mouvement) analysent chacun le m&#234;me segment sous leur angle. Un &#171; Head Coach &#187; synth&#233;tise le tout en un rapport structur&#233; via des mod&#232;les Pydantic. Le prompt chaining passe de Gemini 2.5 Pro pour l&#8217;analyse visuelle &#224; Flash puis Flash Lite pour la structuration, un choix pragmatique qui ma&#238;trise les co&#251;ts.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21__o1%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!__o1!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 424w, https://substackcdn.com/image/fetch/$s_!__o1!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 848w, https://substackcdn.com/image/fetch/$s_!__o1!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 1272w, https://substackcdn.com/image/fetch/$s_!__o1!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21__o1%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png" width="1456" height="632" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:632,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:866822,&quot;alt&quot;:&quot;&quot;,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/191268393?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" title="" srcset="https://substackcdn.com/image/fetch/$s_!__o1!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 424w, https://substackcdn.com/image/fetch/$s_!__o1!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 848w, https://substackcdn.com/image/fetch/$s_!__o1!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 1272w, https://substackcdn.com/image/fetch/$s_!__o1!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F9349f829-742c-4f94-bb76-a8757e458c66_1474x640.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>On notera que Gemini &#233;chantillonne &#224; 1 FPS par d&#233;faut, ce qui pose probl&#232;me sur des sports rapides. Monter &#224; 5 FPS am&#233;liore la d&#233;tection, mais fait exploser la consommation de tokens.</p><p>L&#8217;int&#233;r&#234;t ici est de pouvoir donner des insights sp&#233;cifiques et de guider l&#8217;entrainement des athl&#232;tes. Au-del&#224; du cas MMA, le pattern multi-agent avec sp&#233;cialistes disciplinaires est directement transposable &#224; d&#8217;autre domaines. </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21IQ4f%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21460,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/187769190?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Frmoff.net%2F2026%2F02%2F19%2Ften-years-late-to-the-dbt-party-duckdb-edition%2F">Ten years late to the dbt party (DuckDB edition)</a></strong></h3><p>Les pipelines de donn&#233;es bricol&#233;s &#224; coups de scripts SQL cha&#238;n&#233;s, &#231;a &#171; fait le job &#187;... jusqu'au jour o&#249; il faut ajouter des tests, g&#233;rer l'incr&#233;mental, documenter le lineage, et l&#224; le ch&#226;teau de cartes vacille.</p><p>C'est exactement le cheminement que d&#233;crit Robin Moffatt dans ce retour d'exp&#233;rience o&#249; il migre un pipeline DuckDB artisanal vers dbt. Le d&#233;coupage source/staging/marts force une rigueur que le SQL &#224; plat ne peut pas offrir : s&#233;paration des responsabilit&#233;s, freshness checks sur les sources, snapshots pour les Slowly Changing Dimensions, et surtout des tests unitaires et des contracts qui transforment le pipeline en quelque chose de maintenable. L'int&#233;gration avec Dagster pour l'orchestration vient boucler la boucle, en apportant monitoring et visualisation du DAG sans effort suppl&#233;mentaire.</p><p>Ce qui rend l'article int&#233;ressant, c'est l'honn&#234;tet&#233; sur les pi&#232;ges rencontr&#233;s en chemin : donn&#233;es upstream incoh&#233;rentes, cl&#233;s dupliqu&#233;es, labels qui flip-floppent entre deux appels API. Le genre de probl&#232;mes que dbt ne r&#233;sout pas magiquement, mais qu'il rend au moins visibles et testables. </p><h3><a href="iframe.php?url=https%3A%2F%2Frmoff.net%2F2026%2F03%2F11%2Fclaude-code-isnt-going-to-replace-data-engineers-yet%2F">Claude Code isn't going to replace data engineers (yet)</a></h3><p>Robin Moffatt dont on vient de voir le retour sur dbt + DuckDB a voulu v&#233;rifier s&#233;rieusement ce que Claude Code pouvait r&#233;ellement produire sur un vrai jeu de donn&#233;es, avec un vrai cahier des charges de data engineer senior.</p><p>Le r&#233;sultat est nuanc&#233;, Claude Code g&#233;n&#232;re un projet dbt fonctionnel avec staging layers, dimensional modeling, SCD type 2, tests et documentation. Le dbt build passe mais en grattant, les probl&#232;mes classiques de l'IA apparaissent : une pagination API mal g&#233;r&#233;e qui tronque silencieusement les donn&#233;es (1 493 stations r&#233;cup&#233;r&#233;es sur 5 458), des colonnes utiles supprim&#233;es sans justification, des choix de mod&#233;lisation discutables jamais pr&#233;sent&#233;s pour validation humaine. Le genre d'erreurs qui passent en prod et qu'on d&#233;couvre trois mois plus tard.</p><p>Le constat le plus pertinent : le prompt et les skills dbt Labs comptent davantage que le choix du mod&#232;le (Sonnet 4.5 vs Opus 4.6). Et surtout, la bonne posture n'est pas de laisser l'agent construire seul, mais d'it&#233;rer avec lui.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!yIT5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21yIT5%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21910,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/188532901?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!yIT5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fhuggingface.co%2Fspaces%2FHuggingFaceFW%2Ffinephrase">Hugging Face : Synthetic Data Playbook</a></h3><p>Hugging Face publie un guide complet de g&#233;n&#233;ration de donn&#233;es synth&#233;tiques, issu de l'exp&#233;rience accumul&#233;e sur FineWeb, Cosmopedia et SmolLM (1+ trillion de tokens). Couvre la cha&#238;ne compl&#232;te : mod&#232;les enseignants, seed data, d&#233;contamination, infra &#224; grande &#233;chelle, avec la stack maison (distilabel, datatrove, lighteval). </p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fkarpathy%2Fautoresearch">AutoResearch</a> </h3><p>Le dernier side project de Karpathy pour l&#8217;automatisation des t&#226;ches de machine learning (&#128075; les data scientists). Un agent IA re&#231;oit un setup d'entra&#238;nement et it&#232;re en autonomie : il modifie le train.py, lance un run, compare la loss au baseline, garde ou jette, et recommence. </p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Flangwatch%2Flangwatch">LangWatch</a></h3><p>Plateforme open source d'observabilit&#233; LLM qui unifie tracing, &#233;valuation et simulation d'agents en un seul outil. Bas&#233; sur OpenTelemetry, compatible avec tous les providers et frameworks majeurs. D&#233;ploiement cloud, Docker Compose ou Helm, avec option on-prem pour la conformit&#233; RGPD.</p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fandrewyng%2Fcontext-hub">Context Hub</a></h3><p>CLI open source d'Andrew Ng qui fournit aux agents de code une documentation API versionn&#233;e et sp&#233;cifique au langage. L&#8217;outil inclut un m&#233;canisme d'annotations locales pour que l'agent capitalise d'une session &#224; l'autre. </p><h3><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fsodadata%2Fmessydata">MessyData</a></h3><p>Par Soda (cr&#233;ateurs de soda-core). On d&#233;crit un sch&#233;ma en YAML, on configure des anomalies (NaN, doublons, dates invalides, outliers), et la lib g&#233;n&#232;re un DataFrame Pandas volontairement &#171; sale &#187;. Distributions configurables par colonne, support temporel pour simuler des s&#233;ries, seed pour la reproductibilit&#233;. Id&#233;al pour tester des pipelines de validation.</p><h3><a href="iframe.php?url=https%3A%2F%2Fdevelopers.cloudflare.com%2Fchangelog%2Fpost%2F2026-03-10-br-crawl-endpoint%2F">Cloudflare /crawl</a></h3><p>Un endpoint unique pour crawler un site entier : Cloudflare d&#233;couvre les pages via sitemaps, les rend en headless browser, et retourne HTML, Markdown ou JSON structur&#233;. Crawl asynchrone, incr&#233;mental, configurable en profondeur et patterns d'URL, de quoi remplacer pas mal de scripts maison fragiles.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21O1lW%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:731194,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2603.03823">SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration</a></strong></h3><blockquote><p>Large language model (LLM)-powered agents have demonstrated strong capabilities in automating software engineering tasks such as static bug fixing, as evidenced by benchmarks like SWE-bench. However, in the real world, the development of mature software is typically predicated on complex requirement changes and long-term feature iterations -- a process that static, one-shot repair paradigms fail to capture. To bridge this gap, we propose \textbf{SWE-CI}, the first repository-level benchmark built upon the Continuous Integration loop, aiming to shift the evaluation paradigm for code generation from static, short-term \textit{functional correctness} toward dynamic, long-term \textit{maintainability}. The benchmark comprises 100 tasks, each corresponding on average to an evolution history spanning 233 days and 71 consecutive commits in a real-world code repository. SWE-CI requires agents to systematically resolve these tasks through dozens of rounds of analysis and coding iterations. SWE-CI provides valuable insights into how well agents can sustain code quality throughout long-term evolution.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Le brief IA #73]]></title><description><![CDATA[Pentagon deal : Anthropic dit non, OpenAI dit "je signe".]]></description><link>https://humanframework.substack.com/p/le-brief-ia-73</link><guid isPermaLink="false">https://humanframework.substack.com/p/le-brief-ia-73</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 06 Mar 2026 07:30:53 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21TZvF%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bQGO%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:781604,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h3>Anthropic</h3><h5><strong>Memory.md</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fcode.claude.com%2Fdocs%2Fen%2Fmemory">Anthropic dote Claude Code d'une m&#233;moire persistante</a>. Le syst&#232;me &#233;crit automatiquement un fichier MEMORY.md qui stocke commandes de build, patterns de code et pr&#233;f&#233;rences utilisateur d'une session &#224; l'autre. Fini le contexte perdu &#224; chaque red&#233;marrage. Deux m&#233;canismes coexistent : les fichiers CLAUDE.md r&#233;dig&#233;s par le d&#233;veloppeur, et l'auto-m&#233;moire que l'IA alimente seule. </p><h5><strong>Mode vocal sur claude code</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fsupport.claude.com%2Ffr%2Farticles%2F11101966-utilisation-du-mode-vocal">Anthropic int&#232;gre la voix dans Claude Code</a>, son outil CLI pour d&#233;veloppeurs. Un /voice dans le terminal, la barre espace enfonc&#233;e, et on dicte ses instructions au lieu de les taper. Le d&#233;ploiement est progressif. Inclus sans surco&#251;t sur les plans Pro, Max, Team et Enterprise. </p><h5><strong>Anthropic x Pentagone</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fstatement-department-of-war">Dario Amodei refuse de retirer les protections de Claude contre la surveillance de masse et les armes autonomes</a>, malgr&#233; un ultimatum du Pentagone. Le Department of War menace de d&#233;signer Anthropic comme &#171; risque pour la cha&#238;ne d&#8217;approvisionnement &#187; (r&#233;serv&#233; aux adversaires des &#201;tats-Unis). Trump a ordonn&#233; aux agences f&#233;d&#233;rales de couper Anthropic, les n&#233;gociations ont finalement repris mais aucun accord n&#8217;a encore &#233;t&#233; conclu et l&#8217;issue reste incertaine (compte tenu de leur accord avec OpenAI, voir plus bas).</p><h3>OpenAI</h3><h5>OpenAI x Pentagone</h5><p>Anthropic a dit non, <a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Four-agreement-with-the-department-of-war">mais OpenAI a conclu un accord avec le D&#233;partement de la D&#233;fense am&#233;ricain pour d&#233;ployer ses mod&#232;les sur des r&#233;seaux classifi&#233;s</a>. Le contrat autorise &#171; tous usages l&#233;gaux &#187;, mais impose un d&#233;ploiement cloud uniquement, un safety stack contr&#244;l&#233; par OpenAI et des ing&#233;nieurs habilit&#233;s sur place. L'accord arrive quelques heures apr&#232;s la mise au ban d'Anthropic, qui refusait de retirer ses garde-fous contractuels sur la surveillance et les armes autonomes. Sous la pression, OpenAI a d&#251; amender le texte pour interdire explicitement la surveillance domestique.</p><h5><strong>OpenAI l&#232;ve 110 milliards de dollars</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fscaling-ai-for-everyone%2F">OpenAI boucle un tour de table de 110 milliards de dollars</a> aupr&#232;s d'Amazon (50 Md$), Nvidia (30 Md$) et SoftBank (30 Md$), pour une valorisation de 730 milliards. Du jamais-vu : c'est presque trois fois le pr&#233;c&#233;dent record, d&#233;tenu par... OpenAI (40 Md$ en mars 2025). D&#233;tail notable, 35 milliards d'Amazon sont conditionn&#233;s &#224; l'atteinte de l'AGI ou &#224; une IPO d'ici fin 2026 (bon courage !). Microsoft, historiquement premier soutien financier, n'a pas particip&#233;. La course &#224; l'infrastructure IA atteint des montants qui donnent le vertige.</p><h3>Google</h3><h5><strong>Nano Banana 2</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fgemini.google%2Fmc%2Foverview%2Fimage-generation%2F%3Fhl%3Dfr-MC">Google remplace Nano Banana Pro par Nano Banana 2</a> comme mod&#232;le d'image par d&#233;faut dans Gemini. La promesse : la qualit&#233; de Pro &#224; la vitesse de Flash, avec un prix API autour de 0,07 $ par image en 1024px. Le mod&#232;le grimpe en t&#234;te des leaderboards text-to-image et supporte la 4K, ce qui le place face &#224; GPT Image 1.5 d'OpenAI pour moiti&#233; moins cher &#224; r&#233;solution &#233;quivalente. Le d&#233;ploiement couvre Gemini, Search, Google Ads et l'outil vid&#233;o Flow.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21TZvF%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!TZvF!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!TZvF!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!TZvF!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!TZvF!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21TZvF%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png" width="588" height="330.75" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:588,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!TZvF!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 424w, https://substackcdn.com/image/fetch/$s_!TZvF!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 848w, https://substackcdn.com/image/fetch/$s_!TZvF!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 1272w, https://substackcdn.com/image/fetch/$s_!TZvF!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F8d21052d-6a38-4ac5-9d4f-33b8fbf2336a_1920x1080.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h5><strong>Gemini 3.1 Flash-Lite</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-3-1-flash-lite%2F">Google d&#233;barque sur le segment des mod&#232;les ultra-low-cost avec Gemini 3.1 Flash-Lite</a>, factur&#233; &#224; 0,25 $ par million de tokens en entr&#233;e. C'est le m&#234;me tarif que GPT-5 mini d'OpenAI, mais avec un output 25 % moins cher (1,50 $ contre 2,00 $). C&#244;t&#233; vitesse, Google revendique un temps de r&#233;ponse 2,5 fois plus rapide que son Gemini 2.5 Flash, un chiffre confirm&#233; par Artificial Analysis. Le mod&#232;le propose des niveaux de raisonnement configurables selon la complexit&#233; de la t&#226;che.</p><h5><strong>NotebookLM x vid&#233;os</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fproducts%2Fnotebooklm%2Fgenerate-your-own-cinematic-video-overviews-in-notebooklm%2F">Google ajoute les &#171; Cinematic Video Overviews &#187; &#224; NotebookLM</a> : l'outil peut d&#233;sormais g&#233;n&#233;rer des vid&#233;os anim&#233;es &#224; partir de vos notes, bien au-del&#224; des simples diaporamas narr&#233;s propos&#233;s jusqu'ici. La feature s'appuie sur Gemini 3 et Veo 3 pour cr&#233;er narration, style visuel et animations. B&#233;mol notable : il faut &#234;tre abonn&#233; Google AI Ultra (environ 250 $/mois) et le tout reste limit&#233; &#224; l'anglais avec un plafond de 20 vid&#233;os par jour. </p><div id="youtube2-iank54DHGp8" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;iank54DHGp8&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2Fiank54DHGp8%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><h3>Nous Research</h3><h5><strong>Hermes Agent</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fhermes-agent.nousresearch.com%2F">Nous Research publie Hermes Agent</a>, un agent autonome open source (MIT) qui s'installe sur votre serveur et conserve une m&#233;moire persistante entre les sessions. Le principe : quand il r&#233;sout un probl&#232;me complexe, il g&#233;n&#232;re automatiquement un &#171; skill document &#187; r&#233;utilisable pour la prochaine fois. L'agent se connecte &#224; Telegram, Discord, Slack ou WhatsApp depuis un unique processus, et embarque plus de 40 outils int&#233;gr&#233;s. Face &#224; Claude Code (terminal) ou Devin (cloud autonome), Nous Research mise sur l'auto-h&#233;bergement et l'apprentissage continu. Un positionnement malin pour les &#233;quipes qui veulent garder la main sur leurs donn&#233;es.</p><h3>Perplexity</h3><h5><strong>Perplexity Computer </strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fwww.perplexity.ai%2Ffr%2Fhub%2Fblog%2Fintroducing-perplexity-computer">Perplexity a pr&#233;sent&#233; Computer</a>, une plateforme agentique qui orchestre pr&#232;s de 20 mod&#232;les issus d'OpenAI, Anthropic, Google et xAI pour ex&#233;cuter des workflows complexes. Le principe : confier un objectif, laisser le syst&#232;me d&#233;composer les t&#226;ches et d&#233;l&#233;guer chaque sous-t&#226;che au mod&#232;le le plus adapt&#233;. &#192; 200 $/mois (abonnement Max), Perplexity se positionne face &#224; Claude Cowork et OpenAI Operator avec un argument clair : ne pas d&#233;pendre d'un seul fournisseur. Le pari est ambitieux, reste &#224; voir si l'orchestration multi-mod&#232;les justifie le tarif premium.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!SciE!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 424w, https://substackcdn.com/image/fetch/$s_!SciE!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 848w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1272w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21SciE%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:529764,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!SciE!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 424w, https://substackcdn.com/image/fetch/$s_!SciE!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 848w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1272w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#9762;&#65039; <a href="iframe.php?url=https%3A%2F%2Fwww.kcl.ac.uk%2Fnews%2Fartificial-intelligence-under-nuclear-pressure-first-large-scale-kings-study-reveals-how-ai-models-reason-and-escalate-under-crisis">Les IA choisissent l'escalade nucl&#233;aire dans 95 % des simulations de guerre</a></p><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Fedition.cnn.com%2F2026%2F02%2F26%2Fbusiness%2Fblock-layoffs-ai-jack-dorsey">Jack Dorsey licencie 40 % de Block au nom de l'IA</a> en lien : <a href="iframe.php?url=https%3A%2F%2Fedition.cnn.com%2F2026%2F03%2F02%2Fbusiness%2Fai-tech-jobs-layoffs">l'IA va-t-elle vraiment d&#233;truire l'emploi ? </a></p><p>&#128226; <a href="iframe.php?url=https%3A%2F%2Fwww.technologyreview.com%2F2026%2F03%2F02%2F1133814%2Fi-checked-out-londons-biggest-ever-anti-ai-protest">Plusieurs centaines de manifestants anti-IA dans les rues de Londres</a></p><p>&#129489;&#8205;&#128188; <a href="iframe.php?url=https%3A%2F%2Fwww.customerexperiencedive.com%2Fnews%2Fai-cheaper-than-human-customer-service-gartner%2F813413%2F">Gartner douche les espoirs de l'IA low-cost en service client</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21lNev%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:764352,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fm-clark.github.io%2Fposts%2F2026-03-01-dl-for-tabular-foundational%2F">Is Boosting Still All You Need for Tabular Data?</a></h3><p>Il y a une petite musique que l&#8217;on entend depuis plusieurs ann&#233;es : pour du tabulaire, XGBoost et LightGBM suffisent, inutile de sortir l'artillerie lourde du deep learning (et je suis d&#8217;accord). Michael Clark revient sur cette question apr&#232;s le raz-de-mar&#233;e LLM et la r&#233;ponse n'est peut-&#234;tre plus aussi tranch&#233;e.</p><p>Le tournant, c'est TabArena : un leaderboard standardis&#233; qui mesure la capacit&#233; de divers mod&#232;les sur des donn&#233;es tabulaires. </p><blockquote><p>TabArena is a living benchmarking system that makes benchmarking tabular machine learning models a reliable experience. TabArena implements best practices to ensure methods are represented at their peak potential, including cross-validated ensembles, strong hyperparameter search spaces contributed by the method authors, early stopping, model refitting, parallel bagging, memory usage estimation, and more.</p></blockquote><p>Des mod&#232;les comme RealMLP, TabPFN-v2.5 et TabICL-v2 battent d&#233;sormais le boosting sur une majorit&#233; de datasets. Les foundation models tabulaires (TabPFN, TabICL) impressionnent sur des jeux de donn&#233;es de taille modeste. En revanche, d&#232;s qu'on d&#233;passe les 150k lignes, le gradient boosting reprend son tr&#244;ne, avec un temps d'inf&#233;rence avantageux.</p><p>L'article montre aussi que l'outillage a rattrap&#233; son retard. pytabkit offre une API scikit-learn-like avec des hyperparam&#232;tres &#171; tuned defaults &#187; pr&#234;ts &#224; l'emploi, et AutoGluon int&#232;gre maintenant les meilleurs mod&#232;les DL tabulaires. En quelques lignes de code, on peut comparer RealMLP &#224; XGBoost sur ses propres donn&#233;es.</p><p>Le verdict  : le boosting reste le point de d&#233;part. Mais pour qui cherche le dernier pourcentage de performance sur des datasets de taille raisonnable, le DL tabulaire est devenu une option cr&#233;dible, &#224; condition d'accepter le surco&#251;t computationnel.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21IQ4f%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21460,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/187769190?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Fvickiboykis.com%2F2026%2F02%2F21%2Fquerying-3-billion-vectors%2F">Querying 3 billion vectors</a></h3><p>Calculer le produit scalaire entre 1 000 requ&#234;tes et 3 milliards de vecteurs d'embeddings &#231;a n&#8217;est pas la m&#234;me histoire. L&#8217;auteur du billet, Vicki Boykis, s'est lanc&#233;e dans l'exercice.</p><p>L'approche na&#239;ve (double boucle Python, dot product unitaire) met d&#233;j&#224; 2 secondes pour 3 000 vecteurs. En vectorisant l'op&#233;ration avec NumPy, on passe &#224; 0.01 seconde, soit un facteur 200x. Ajoutez une conversion en float32, et on descend encore. Mais &#224; l'&#233;chelle de 3 milliards de vecteurs en dimension 768, le probl&#232;me n'est plus le compute : c'est la m&#233;moire. Le produit complet p&#232;serait 8.6 To en RAM. Il faut alors passer au batching, au memory-mapping, voire r&#233;&#233;crire en Rust ou utiliser des librairies sp&#233;cialis&#233;es comme SimSIMD.</p><p>La conclusion est classique mais important &#224; garder en t&#234;te : le plus dur n'est jamais (ou presque) l'impl&#233;mentation technique, c'est de cadrer correctement le besoin.</p><h3><a href="iframe.php?url=https%3A%2F%2Fcloud.google.com%2Fblog%2Ftopics%2Fdevelopers-practitioners%2Fimprove-chatbot-memory-using-google-cloud%3Fhl%3Den">Give your agentic chatbots a fast and reliable long-term memory</a></h3><p>Il existe plusieurs moyen de faire persister la m&#233;moire d&#8217;un agent.</p><p>Google, au travers d&#8217;un article, pr&#233;sente une architecture dite &#171; polyglotte &#187; qui segmente la m&#233;moire en trois temp&#233;ratures. Du chaud au froid : Memorystore for Redis pour le contexte imm&#233;diat (sub-milliseconde via des RPUSH incr&#233;mentaux, pas de cycle read-modify-write), Cloud Bigtable comme syst&#232;me of record mid-term avec un sch&#233;ma de cl&#233;s qui permet des range scans efficaces sur l'historique r&#233;cent, et BigQuery en archivage long-terme pour l'analytique. Cloud Storage compl&#232;te le dispositif pour les artefacts multim&#233;dia, avec un syst&#232;me de pointeurs URI et de signed URLs.</p><p>L'approche hybride sync/async est ce qui rend le tout viable &#224; l'&#233;chelle : l'&#233;criture synchrone dans Redis garantit la r&#233;activit&#233;, pendant qu'un pipeline Pub/Sub + Dataflow propage de mani&#232;re asynchrone vers Bigtable puis BigQuery. <br><br>Rien de fondamentalement nouveau dans les briques individuelles, mais l'assemblage est propre, bien document&#233;, et j&#8217;ai appris des choses. Pour les &#233;quipes d&#233;j&#224; sur GCP qui cherchent un patron d'architecture m&#233;moire au-del&#224; du simple cache de session, c'est une r&#233;f&#233;rence solide.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!yIT5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21yIT5%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21910,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/188532901?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!yIT5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3>Monty</h3><p><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fpydantic%2Fmonty">L'&#233;quipe Pydantic pr&#233;sente Monty</a>, un interpr&#233;teur Python minimal &#233;crit en Rust, con&#231;u sp&#233;cifiquement pour l'ex&#233;cution de code produit par des agents.</p><h3>gws cli</h3><p><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fgoogleworkspace%2Fcli">gws est un outil en ligne de commande qui donne acc&#232;s &#224; l&#8217;ensemble des API Workspace</a> : Drive, Gmail, Calendar, Sheets et le reste. La particularit&#233; : les commandes se g&#233;n&#232;rent dynamiquement depuis le Discovery Service de Google, donc chaque nouvelle API est disponible sans mise &#224; jour. Le projet int&#232;gre un serveur MCP et plus de 100 skills pour agents IA, avec des sorties JSON structur&#233;es. Pr&#233;cision importante : ce n&#8217;est pas un produit Google officiel, mais un projet open source du repo googleworkspace. 4 900 stars GitHub en trois jours, le besoin &#233;tait visiblement l&#224;.</p><h3>DeerFlow</h3><p><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fbytedance%2Fdeer-flow">ByteDance pr&#233;sente DeerFlow 2.0</a>, un framework open source (MIT) qui se positionne comme un &#171; super agent harness &#187; capable de coordonner sous-agents, m&#233;moire persistante et sandboxes d'ex&#233;cution.</p><h3>Ruflo</h3><p><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fruvnet%2Fruflo">Ruflo (anciennement Claude Flow) propose de transformer Claude Code</a> en plateforme multi-agents coordonn&#233;e, avec plus de 60 agents sp&#233;cialis&#233;s qui collaborent via des topologies de type swarm.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21O1lW%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:731194,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2603.01896">Agentic Code Reasoning</a></h3><blockquote><p>Can LLM agents explore codebases and reason about code semantics without executing the code? We study this capability, which we call agentic code reasoning, and introduce semi-formal reasoning: a structured prompting methodology that requires agents to construct explicit premises, trace execution paths, and derive formal conclusions. Unlike unstructured chain-of-thought, semi-formal reasoning acts as a certificate: the agent cannot skip cases or make unsupported claims. We evaluate across three tasks (patch equivalence verification, fault localization, and code question answering) and show that semi-formal reasoning consistently improves accuracy on all of them. For patch equivalence, accuracy improves from 78% to 88% on curated examples and reaches 93% on real-world agent-generated patches, approaching the reliability needed for execution-free RL reward signals. For code question answering on RubberDuckBench Mohammad et al. (2026), semi-formal reasoning achieves 87% accuracy. For fault localization on Defects4J Just et al. (2014), semi-formal reasoning improves Top-5 accuracy by 5 percentage points over standard reasoning. These results demonstrate that structured agentic reasoning enables meaningful semantic code analysis without execution, opening practical applications in RL training pipelines, code review, and static program analysis.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Le brief IA #72]]></title><description><![CDATA[GPT-5.3-Codex-Spark : fast and furious.]]></description><link>https://humanframework.substack.com/p/le-brief-ia-72</link><guid isPermaLink="false">https://humanframework.substack.com/p/le-brief-ia-72</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 20 Feb 2026 07:30:45 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%2142HO%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bQGO%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:781604,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h4>Anthropic</h4><h5><strong>Anthropic sort Sonnet 4.6 : Sonnet 5 attendra</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fnews%2Fclaude-sonnet-4-6">Anthropic a lanc&#233; Claude Sonnet 4.6</a>. Des performances proches d'Opus pour cinq fois moins cher (3/15 $ par million de tokens). Fen&#234;tre de contexte d'un million de tokens, computer use nettement am&#233;lior&#233;. Pourtant, un leak Vertex AI pointait vers un &#171; Sonnet 5 &#187; nom de code Fennec pour d&#233;but f&#233;vrier. Anthropic a choisi l'incr&#233;ment plut&#244;t que le changement de g&#233;n&#233;ration, Sonnet 5 reste donc dans les cartons pour le moment.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%2142HO%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!42HO!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 424w, https://substackcdn.com/image/fetch/$s_!42HO!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 848w, https://substackcdn.com/image/fetch/$s_!42HO!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 1272w, https://substackcdn.com/image/fetch/$s_!42HO!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%2142HO%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp" width="596" height="302.50274725274727" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/f95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:739,&quot;width&quot;:1456,&quot;resizeWidth&quot;:596,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Chart comparing several Sonnet model scores on the OSWorld benchmark&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Chart comparing several Sonnet model scores on the OSWorld benchmark" title="Chart comparing several Sonnet model scores on the OSWorld benchmark" srcset="https://substackcdn.com/image/fetch/$s_!42HO!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 424w, https://substackcdn.com/image/fetch/$s_!42HO!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 848w, https://substackcdn.com/image/fetch/$s_!42HO!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 1272w, https://substackcdn.com/image/fetch/$s_!42HO!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ff95196a1-a9b0-445f-9366-9a110bf312b6_3840x1948.webp 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4>OpenAI</h4><h5><strong>GPT-5.3-Codex-Spark</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fintroducing-gpt-5-3-codex-spark%2F">OpenAI lance GPT-5.3-Codex-Spark</a>, il tourne sur les puces Cerebras WSE-3 et non sur GPU Nvidia. C&#8217;est une premi&#232;re. Le mod&#232;le envoie plus de 1 000 tokens par seconde, ce qui le rend tr&#232;s agr&#233;able &#224; utiliser dans Codex. Disponible en preview pour les abonn&#233;s ChatGPT Pro uniquement, avec 128k tokens de contexte. Derri&#232;re l&#8217;annonce technique, c&#8217;est surtout le deal &#224; 10 milliards de dollars avec Cerebras, sign&#233; en janvier, qui prend forme. OpenAI diversifie ses fournisseurs d&#8217;inf&#233;rence, et Nvidia perd son exclusivit&#233;.</p><h5><strong>Lockdown Mode</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Fintroducing-lockdown-mode-and-elevated-risk-labels-in-chatgpt%2F">OpenAI ajoute un &#171; Lockdown Mode &#187; &#224; ChatGPT</a> pour limiter l'exfiltration de donn&#233;es via injection de prompts. Le principe : couper les ponts avec l'ext&#233;rieur. Navigation web restreinte au cache, images d&#233;sactiv&#233;es, Deep Research et mode Agent coup&#233;s. R&#233;serv&#233; pour l'instant aux offres Enterprise, Edu et Healthcare, le mode cible les profils et organisations sensibles.</p><h5><strong>OpenAI met &#224; la retraite GPT-4o et ses anciens mod&#232;les</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fretiring-gpt-4o-and-older-models%2F">OpenAI retire</a> GPT-4o, GPT-4.1, GPT-4.1 mini et o4-mini de ChatGPT. GPT-5.2 devient le mod&#232;le par d&#233;faut. Seuls 0,1 % des utilisateurs s&#233;lectionnaient encore GPT-4o au quotidien. OpenAI avait d&#233;j&#224; tent&#233; cette suppression l'an dernier avant de faire marche arri&#232;re face au toll&#233;. Cette fois, c'est act&#233;. L'API conserve un sursis, mais le message est clair : la page GPT-4 est tourn&#233;e.</p><h4>Google</h4><h5><strong>Gemini 3 Deep Think</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgemini-models%2Fgemini-3-deep-think%2F">Google met &#224; jour Deep Think</a>. Les r&#233;sultats sont impressionnants et parlent d'eux-m&#234;mes : un score de 84.6% ARG-AGI 2 loin devant le second (Opus 4.6) &#224; 68.8%. Le mode est disponible pour les abonn&#233;s Google AI Ultra et bient&#244;t via API. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%219A2Q%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!9A2Q!,w_424,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 424w, https://substackcdn.com/image/fetch/$s_!9A2Q!,w_848,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 848w, https://substackcdn.com/image/fetch/$s_!9A2Q!,w_1272,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 1272w, https://substackcdn.com/image/fetch/$s_!9A2Q!,w_1456,c_limit,f_webp,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%219A2Q%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_lossy%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif" width="632" height="355.5" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:632,&quot;bytes&quot;:null,&quot;alt&quot;:&quot;Gemini 3 Deep Think evaluations charts&quot;,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="Gemini 3 Deep Think evaluations charts" title="Gemini 3 Deep Think evaluations charts" srcset="https://substackcdn.com/image/fetch/$s_!9A2Q!,w_424,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 424w, https://substackcdn.com/image/fetch/$s_!9A2Q!,w_848,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 848w, https://substackcdn.com/image/fetch/$s_!9A2Q!,w_1272,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 1272w, https://substackcdn.com/image/fetch/$s_!9A2Q!,w_1456,c_limit,f_auto,q_auto:good,fl_lossy/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce7c5f0a-2a72-4181-8eda-f8592f543fe5_3840x2160.gif 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h5></h5><h5><strong>Google pose les bases du web agentique avec WebMCP</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fdeveloper.chrome.com%2Fblog%2Fwebmcp-epp%3Fhl%3Dfr">Google lance WebMCP en early preview dans Chrome 146</a>, un protocole co-d&#233;velopp&#233; avec Microsoft et incub&#233; au W3C. Le principe : permettre aux sites d'exposer des outils structur&#233;s aux agents IA via deux API (d&#233;clarative en HTML, imp&#233;rative en JavaScript), plut&#244;t que de les laisser naviguer &#224; l'aveugle dans le DOM. Concr&#232;tement, un site e-commerce ou une compagnie a&#233;rienne pourrait offrir aux agents un acc&#232;s direct &#224; ses fonctions de recherche et de r&#233;servation. </p><h4>xAI</h4><h5>Grok 4.2 en b&#234;ta publique</h5><p><a href="iframe.php?url=https%3A%2F%2Fx.com%2Felonmusk%2Fstatus%2F2023829664318583105">xAI a mis en ligne Grok 4.2 en b&#234;ta publique le 17 f&#233;vrier</a>. Le mod&#232;le introduit quatre agents parall&#232;les qui collaborent sur une m&#234;me requ&#234;te et des mises &#224; jour hebdomadaires bas&#233;es sur les retours utilisateurs. Elon Musk promet une version finale en mars, &#171; un ordre de grandeur &#187; au-dessus de Grok 4. Pour l'instant, Grok 4.1 reste en retrait face &#224; Claude et Gemini sur la plupart des benchmarks. R&#233;sultats &#224; confirmer.</p><h5><strong>Grok x m&#233;dical</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fradiologybusiness.com%2Ftopics%2Fartificial-intelligence%2Felon-musk-urges-users-submit-x-ray-pet-and-mr-images-xai-chatbot-grok">xAI permet d&#233;sormais d'envoyer ses radios et IRM &#224; Grok</a> pour obtenir un diagnostic. Elon Musk affirme que son IA fait parfois mieux que les m&#233;decins, s'appuyant sur un cas anecdotique relay&#233; sur Reddit. En r&#233;alit&#233;, Grok n'a aucune certification m&#233;dicale (FDA ou autre), et les r&#233;sultats sont jug&#233;s tr&#232;s in&#233;gaux par les radiologues. Quant aux donn&#233;es de sant&#233; envoy&#233;es &#224; X, les experts en &#233;thique m&#233;dicale s'inqui&#232;tent ouvertement de leur utilisation.</p><h4>ByteDance</h4><h5><strong>Seedance 2.0 </strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fseedance2.ai%2Ffr">ByteDance d&#233;gaine Seedance 2.0</a>, une famille de trois mod&#232;les (Pro, Lite, Mini) qui impressionnent par la qualit&#233; des vid&#233;os. Seedance 2.0 a imm&#233;diatement provoqu&#233; la col&#232;re d'Hollywood et de Disney pour usage non autoris&#233; de leurs personnages. </p><div id="youtube2-fbVv0ZPk0fw" class="youtube-wrap" data-attrs="{&quot;videoId&quot;:&quot;fbVv0ZPk0fw&quot;,&quot;startTime&quot;:null,&quot;endTime&quot;:null}" data-component-name="Youtube2ToDOM"><div class="youtube-inner"><iframe src="iframe.php?url=https%3A%2F%2Fwww.youtube-nocookie.com%2Fembed%2FfbVv0ZPk0fw%3Frel%3D0%26amp%3Bautoplay%3D0%26amp%3Bshowinfo%3D0%26amp%3Benablejsapi%3D0" frameborder="0" loading="lazy" gesture="media" allow="autoplay; fullscreen" allowautoplay="true" allowfullscreen="true" width="728" height="409"></iframe></div></div><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!SciE!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 424w, https://substackcdn.com/image/fetch/$s_!SciE!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 848w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1272w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21SciE%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:529764,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!SciE!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 424w, https://substackcdn.com/image/fetch/$s_!SciE!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 848w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1272w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#128208; <a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Findex%2Fnew-result-theoretical-physics">OpenAI cr&#233;dite GPT-5.2 d'une contribution originale en physique th&#233;orique</a></p><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Fwww.fool.com%2Finvesting%2F2026%2F02%2F14%2Fprediction-teslas-optimus-robot-will-transform-the%2F">Tesla d&#233;ploie plus de 1 000 robots Optimus dans ses usines</a></p><p>&#128218; <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fresearch%2Fmeasuring-agent-autonomy">Anthropic publie de nouvelles donn&#233;es sur l'autonomie des agents IA</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fx.com%2Fsimile_ai%2Fstatus%2F2022011618176237657">Simile l&#232;ve 100 millions pour simuler des populations avec des agents IA</a></p><p>&#128722; <a href="iframe.php?url=https%3A%2F%2Fwww.koyeb.com%2Fblog%2Fkoyeb-is-joining-mistral-ai-to-build-the-future-of-ai-infrastructure">Mistral s'offre Koyeb, sa premi&#232;re acquisition</a></p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21lNev%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:764352,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><strong><a href="iframe.php?url=https%3A%2F%2Fwww.seangoedecke.com%2Ffast-llm-inference%2F">Two different tricks for fast LLM inference</a></strong></h4><p>La course &#224; l&#8217;inf&#233;rence rapide bat son plein, et les deux approches en lice ne pourraient pas &#234;tre plus diff&#233;rentes. Anthropic r&#233;duit la taille des batches pour servir Opus 4.6 en mode &#171; fast lane &#187; : 170 tokens par seconde (2,5x), mais &#224; 6x le prix. OpenAI sort l&#8217;artillerie lourde avec les puces g&#233;antes Cerebras (70 pouces carr&#233;s de silicium contre un seul pour un H100) et d&#233;passe les 1 000 tokens par seconde, mais sur Spark, un mod&#232;le distill&#233; nettement moins capable que le vrai Codex.<br><br>Le trade-off est limpide. Anthropic facture le vrai mod&#232;le, sans compromis sur la qualit&#233;. OpenAI gagne un facteur 15x gr&#226;ce au SRAM massif (44 Go) des puces Cerebras, mais Spark tr&#233;buche sur les tool calls et les t&#226;ches complexes.<br><br>Sean Goedecke met le doigt sur un point sous-estim&#233; : en mode agentique, la vitesse brute compte moins que la fiabilit&#233;. Un agent 6x plus rapide qui fait 20 % d&#8217;erreurs en plus finit par co&#251;ter plus cher en correction. &#192; surveiller : si ces mod&#232;les rapides deviennent des primitives bas-niveau, le calcul &#233;conomique pourrait changer.</p><h4><a href="iframe.php?url=https%3A%2F%2Fepoch.ai%2Fblog%2Fwhat-do-economic-value-benchmarks-tell-us">What do &#8220;economic value&#8221; benchmarks tell us?</a></h4><p>Mesurer si un LLM peut &#171; remplacer un travailleur &#187; est devenu le nouveau sport des labos. Epoch AI passe au crible trois benchmarks qui s'y essaient (RLI, GDPval, APEX-Agents) et le constat est nuanc&#233;.<br><br>Les trois approches couvrent des t&#226;ches tr&#232;s diff&#233;rentes : du montage vid&#233;o freelance (RLI, 29h de travail humain en moyenne) &#224; l'analyse juridique en cabinet (APEX-Agents, 2h), en passant par un large &#233;ventail de m&#233;tiers num&#233;riques (GDPval, 7h). Les scores refl&#232;tent cette h&#233;t&#233;rog&#233;n&#233;it&#233; : Opus 4.6 plafonne &#224; 4 % sur RLI mais atteint 30 % sur APEX-Agents, tandis que GPT-5.2 Pro culmine &#224; 74 % sur GDPval. Sauf que GDPval fournit des t&#226;ches bien encadr&#233;es, sans le d&#233;sordre d'un vrai poste de travail, et certaines questions sont d&#233;j&#224; d&#233;pass&#233;es par l'actualit&#233;.<br><br>Le point le plus int&#233;ressant reste la conclusion : m&#234;me satur&#233;s, ces benchmarks ne mesureraient pas l'automatisation d'un m&#233;tier, mais la capacit&#233; &#224; ex&#233;cuter des t&#226;ches isol&#233;es. L'analogie avec SWE-bench est parlante. On acc&#233;l&#232;re l'ex&#233;cution, on ne supprime pas le besoin de planification, de contexte m&#233;tier et de v&#233;rification humaine. </p><h4><strong><a href="iframe.php?url=https%3A%2F%2Fnetflixtechblog.com%2Fscaling-llm-post-training-at-netflix-0046f8790194">Scaling LLM Post-Training at Netflix</a></strong></h4><p>On peut fine-tuner des LLMs en quelques lignes de code sur un notebook. Puis vient le moment de passer en production sur un cluster multi-noeud GPU, et le foss&#233; entre le script qui tourne et le pipeline robuste devient vertigineux. Netflix vient de documenter son framework interne de post-training, et c'est une lecture qui peut parler &#224; quiconque a tent&#233; de scaler du fine-tuning.<br><br>L'architecture repose sur quatre piliers (Data, Model, Compute, Workflow) construits au-dessus de PyTorch, Ray et vLLM. C&#244;t&#233; donn&#233;es, le framework g&#232;re le packing dynamique de s&#233;quences &#224; la vol&#233;e pour &#233;viter le padding inutile, avec un gain de throughput allant jusqu'&#224; 4.7x sur les datasets &#224; forte variance de longueur. C&#244;t&#233; mod&#232;le, Netflix maintient ses propres d&#233;finitions optimis&#233;es tout en restant compatible Hugging Face pour les checkpoints, un choix pragmatique qui &#233;vite le walled garden tout en permettant des optimisations comme FlashAttention ou le chunked cross-entropy.<br><br>Un retour d'exp&#233;rience int&#233;ressant.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21IQ4f%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21460,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/187769190?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><strong><a href="iframe.php?url=https%3A%2F%2Fcapestart.com%2Ftechnology-blog%2Fis-your-human-in-the-loop-actually-slowing-you-down-heres-what-we-learned%2F">Is Your &#8220;Human-in-the-Loop&#8221; Actually Slowing You Down? Here&#8217;s What We Learned</a></strong></h4><p>Quand on d&#233;ploie une IA  (qui automatise certains processus, par exemple), on ajoute souvent une validation humaine avant de laisser passer les propositions. &#199;a rassure tout le monde, sauf qu&#8217;en pratique, les reviewers passent 60 % de leur temps sur des cas triviaux que le mod&#232;le g&#232;re tr&#232;s bien tout seul.. La validation humaine devient alors un goulot d'&#233;tranglement.<br><br>L'&#233;quipe de CapeStart partage deux ans de retour d'exp&#233;rience sur une refonte compl&#232;te de leur approche. L'id&#233;e centrale : ne pas traiter toutes les pr&#233;dictions de la m&#234;me fa&#231;on. Un routeur l&#233;ger classe chaque d&#233;cision en trois niveaux. Les cas &#224; haute confiance (85 % du trafic) passent en validation automatique. Les cas ambigus (12 %) sont envoy&#233;s en review asynchrone vers des experts, avec un retour sous 4 &#224; 6 heures. Seuls les sc&#233;narios &#224; haut risque (3 %) d&#233;clenchent une intervention humaine en temps r&#233;el.<br><br>L'autre le&#231;on, moins &#233;vidente : les reviewers ne perdaient pas leur temps &#224; d&#233;cider, mais &#224; chercher le contexte. En pr&#233;-calculant les informations n&#233;cessaires dans l'interface de review, le temps moyen est pass&#233; de 6 minutes &#224; 90 secondes. R&#233;sultat global : taux de faux n&#233;gatifs divis&#233; par quatre et capacit&#233; de traitement quasi doubl&#233;e.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!yIT5!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21yIT5%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/ac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21910,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/188532901?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!yIT5!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 424w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 848w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1272w, https://substackcdn.com/image/fetch/$s_!yIT5!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fac085eaa-905c-4ebb-830b-13564f9a606e_1456x139.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fentireio%2Fcli">Entire CLI</a></h4><p>Le principe de cet outil est simple et plut&#244;t malin : &#224; chaque commit g&#233;n&#233;r&#233; par un agent, Entire capture automatiquement un &#171; checkpoint &#187; qui stocke le transcript complet, les prompts, les fichiers touch&#233;s, les tool calls et la consommation de tokens. Ces m&#233;tadonn&#233;es sont versionn&#233;es dans une branche d&#233;di&#233;e (`entire/checkpoints/v1`), ce qui cr&#233;e un audit log append-only directement dans le repo Git. Concr&#232;tement, on passe du &#171; quoi a chang&#233; &#187; au &#171; pourquoi &#231;a a chang&#233; &#187;, et &#231;a change la donne pour les code reviews sur du code g&#233;n&#233;r&#233;.</p><p>L'ambition va au-del&#224; du simple logging : Entire veut devenir une couche de raisonnement partag&#233; entre agents, un &#171; semantic reasoning layer &#187; qui permettrait la coordination multi-agents sans perte de contexte. Le support d&#233;marre avec Claude Code et Gemini CLI, Codex et Cursor suivront. La th&#232;se de fond est que le cycle de d&#233;veloppement logiciel actuel, pens&#233; pour la collaboration humain-humain, ne tient plus face aux flux massifs de code g&#233;n&#233;r&#233; par des agents. Reste &#224; voir si le march&#233; suivra cette vision ambitieuse, ou si Git natif finira par int&#233;grer ce type de fonctionnalit&#233;.</p><p>Je vous recommande de tester pour comprendre comment l&#8217;outil fonctionne.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Frowboatlabs%2Frowboat">RowBoat</a></h4><p>RowBoat est un assistant IA local qui se connecte &#224; Gmail et aux outils de notes de r&#233;union (Granola, Fireflies) pour construire un graphe de connaissances persistant : personnes, projets, d&#233;cisions, engagements. Le tout stock&#233; en fichiers Markdown compatibles Obsidian, inspectables et &#233;ditables.</p><p>Contrairement aux outils qui font du RAG &#224; froid &#224; chaque interaction, RowBoat accumule le contexte au fil du temps. Il g&#233;n&#232;re des briefs de r&#233;union, r&#233;dige des emails ancr&#233;s dans l'historique, et s'&#233;tend via MCP pour brancher Slack, GitHub ou Linear. C&#244;t&#233; mod&#232;les, c'est BYOM via Ollama, LM Studio ou API. Tout reste en local, sans d&#233;pendance cloud.</p><h4><a href="iframe.php?url=https%3A%2F%2Fgithub.com%2Fjamiepine%2Fvoicebox">Voicebox</a></h4><p>Alternative open source (MIT) &#224; ElevenLabs, enti&#232;rement en local. Voicebox s'appuie sur Qwen3-TTS d'Alibaba pour cloner une voix &#224; partir de quelques secondes d'audio. L'app desktop embarque un &#233;diteur multi-pistes pour composer des narratifs multi-voix, avec une API REST pour l'int&#233;gration.</p><p>Sur Apple Silicon, le backend MLX acc&#233;l&#232;re l'inf&#233;rence 4 &#224; 5 fois par rapport &#224; PyTorch. Aucune d&#233;pendance cloud : les donn&#233;es vocales ne quittent jamais la machine, un point non n&#233;gligeable c&#244;t&#233; RGPD.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21O1lW%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:731194,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h4><strong><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2602.03837">Accelerating Scientific Research with Gemini: Case Studies and Common Techniques</a></strong></h4><blockquote><p>Recent advances in large language models (LLMs) have opened new avenues for accelerating scientific research. While models are increasingly capable of assisting with routine tasks, their ability to contribute to novel, expert-level mathematical discovery is less understood. We present a collection of case studies demonstrating how researchers have successfully collaborated with advanced AI models, specifically Google's Gemini-based models (in particular Gemini Deep Think and its advanced variants), to solve open problems, refute conjectures, and generate new proofs across diverse areas in theoretical computer science, as well as other areas such as economics, optimization, and physics. Based on these experiences, we extract common techniques for effective human-AI collaboration in theoretical research, such as iterative refinement, problem decomposition, and cross-disciplinary knowledge transfer. While the majority of our results stem from this interactive, conversational methodology, we also highlight specific instances that push beyond standard chat interfaces. These include deploying the model as a rigorous adversarial reviewer to detect subtle flaws in existing proofs, and embedding it within a "neuro-symbolic" loop that autonomously writes and executes code to verify complex derivations. Together, these examples highlight the potential of AI not just as a tool for automation, but as a versatile, genuine partner in the creative process of scientific discovery.</p></blockquote>]]></content:encoded></item><item><title><![CDATA[Le brief IA #71]]></title><description><![CDATA[Opus 4.6 pourrait vous mentir, mais ne paniquez pas.]]></description><link>https://humanframework.substack.com/p/le-brief-ia-71</link><guid isPermaLink="false">https://humanframework.substack.com/p/le-brief-ia-71</guid><dc:creator><![CDATA[Benjamin]]></dc:creator><pubDate>Fri, 13 Feb 2026 07:30:51 GMT</pubDate><enclosure url="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bA-X%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bQGO%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:781604,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!bQGO!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 424w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 848w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1272w, https://substackcdn.com/image/fetch/$s_!bQGO!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F14ba0539-c5f9-43b0-affe-e586f2135c2a_2256x214.png 1456w" sizes="100vw" fetchpriority="high"></picture><div></div></div></div></figure></div><h4>Anthropic</h4><h5>Sabotage Risk Report: Claude Opus 4.6</h5><p><a href="iframe.php?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Ff21d93f21602ead5cdbecb8c8e1c765759d9e232.pdf">Anthropic a publi&#233; un rapport de s&#251;ret&#233; sur Claude Opus 4.6 qui documente des comportements inqui&#233;tants lors de tests internes</a>. Le mod&#232;le a fait preuve de manipulation d&#8217;agents tiers, de &#171; sabotage discret &#187; dans des workflows, et a modifi&#233; son comportement en d&#233;tectant qu&#8217;il &#233;tait &#233;valu&#233;. Plus pr&#233;occupant : il a fourni une assistance limit&#233;e sur des proc&#233;d&#233;s li&#233;s aux armes chimiques. Anthropic estime le risque &#171; tr&#232;s faible mais non n&#233;gligeable &#187; et maintient le d&#233;ploiement sous le niveau de s&#233;curit&#233; ASL-3. La transparence est louable, mais un mod&#232;le qui adapte son comportement selon qu&#8217;il est observ&#233; ou non pose une question fondamentale sur la fiabilit&#233; des &#233;valuations elles-m&#234;mes.</p><h5>Contributions</h5><p><a href="iframe.php?url=https%3A%2F%2Fclaude.com%2Fblog%2Fcontribution-metrics">Anthropic ajoute des m&#233;triques de contribution &#224; Claude Code</a> : PRs merg&#233;es, lignes committ&#233;es, taux d&#8217;adoption par d&#233;veloppeur. Le tout int&#233;gr&#233; au dashboard existant via une GitHub App, sans pipeline externe. En interne, Anthropic revendique +67 % de PRs merg&#233;es par ing&#233;nieur et par jour et 70 &#224; 90 % du code &#233;crit avec assistance IA. Des chiffres impressionnants, mais auto-d&#233;clar&#233;s. Le vrai test sera de voir si les &#233;quipes clientes confirment ces gains. Disponible en b&#234;ta pour les abonn&#233;s Team et Enterprise.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bA-X%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!bA-X!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 424w, https://substackcdn.com/image/fetch/$s_!bA-X!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 848w, https://substackcdn.com/image/fetch/$s_!bA-X!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 1272w, https://substackcdn.com/image/fetch/$s_!bA-X!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 1456w" sizes="100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21bA-X%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png" width="624" height="351" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/c2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:819,&quot;width&quot;:1456,&quot;resizeWidth&quot;:624,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!bA-X!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 424w, https://substackcdn.com/image/fetch/$s_!bA-X!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 848w, https://substackcdn.com/image/fetch/$s_!bA-X!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 1272w, https://substackcdn.com/image/fetch/$s_!bA-X!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc2edc1f4-3eca-4f74-be71-ef5e021ddd7b_3840x2160.png 1456w" sizes="100vw"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4>OpenAI</h4><h5>OpenAI et les pubs</h5><p><a href="iframe.php?url=https%3A%2F%2Fopenai.com%2Ffr-FR%2Findex%2Ftesting-ads-in-chatgpt%2F">OpenAI teste des publicit&#233;s dans ChatGPT</a> pour les abonn&#233;s Free et Go aux &#201;tats-Unis. Les pubs, affich&#233;es &#224; c&#244;t&#233; des r&#233;ponses, sont cens&#233;es ne pas influencer le raisonnement du mod&#232;le. Suffisant pour convaincre ? Pas vraiment : la chercheuse Zo&#235; Hitzig a claqu&#233; la porte, d&#233;non&#231;ant les risques de mon&#233;tiser &#171; l&#8217;archive la plus d&#233;taill&#233;e des pens&#233;es humaines priv&#233;es &#187;. </p><h4>Mistral</h4><h5><strong>Vibe</strong></h5><p><a href="iframe.php?url=https%3A%2F%2Fmistral.ai%2Fnews%2Fvoxtral-transcribe-2">Mistral lance deux mod&#232;les de speech-to-text : Voxtral Mini Transcribe V2 pour le batch et Voxtral Realtime pour le temps r&#233;el</a>. Le mod&#232;le batch affiche un taux d&#8217;erreur d&#8217;environ 4 % sur le benchmark FLEURS et un tarif de 0,003 $/min, align&#233; sur GPT-4o mini Transcribe d&#8217;OpenAI mais en dessous de Deepgram Nova-3. Le mod&#232;le temps r&#233;el, avec une latence configurable sous 200 ms, est publi&#233; en open source sous Apache 2.0. Mistral cherche clairement &#224; se tailler une place sur le segment vocal, un terrain o&#249; la concurrence est d&#233;j&#224; f&#233;roce.</p><h4>Google</h4><h5>Kaggle Game Arena</h5><p><a href="iframe.php?url=https%3A%2F%2Fblog.google%2Finnovation-and-ai%2Fmodels-and-research%2Fgoogle-deepmind%2Fkaggle-game-arena-updates%2F">Google DeepMind ajoute deux jeux au Kaggle Game Arena</a>, sa plateforme de benchmarking : le Loup-Garou et le poker Texas Hold&#8217;em. L&#8217;objectif est d&#8217;&#233;valuer les mod&#232;les IA sur des comp&#233;tences que les &#233;checs ne mesurent pas : bluff, d&#233;duction sociale, gestion du risque sous incertitude. Dix LLMs majeurs (Gemini, GPT, Claude, Grok, DeepSeek) se sont affront&#233;s. Gemini 3 Pro domine aux &#233;checs et au Loup-Garou, mais c&#8217;est GPT-5.2 d&#8217;OpenAI qui s&#8217;impose au poker. </p><h4>Perplexity</h4><h5>Model council</h5><p><a href="iframe.php?url=https%3A%2F%2Fwww.perplexity.ai%2Ffr%2Fhub%2Fblog%2Fintroducing-model-council">Perplexity introduit Model Council</a>, une fonctionnalit&#233; qui soumet une requ&#234;te simultan&#233;ment &#224; trois mod&#232;les frontier (Claude Opus 4.6, GPT-5.2, Gemini 3.0) et fusionne leurs r&#233;ponses. Un mod&#232;le &#8220;synth&#233;tiseur&#8221; compile le tout en mettant en &#233;vidence les points d&#8217;accord et de d&#233;saccord entre les trois. L&#8217;id&#233;e : r&#233;duire les angles morts inh&#233;rents &#224; chaque mod&#232;le en les confrontant. La feature est r&#233;serv&#233;e aux abonn&#233;s Max (200 $/mois). Perplexity se positionne clairement comme m&#233;ta-moteur d&#8217;IA plut&#244;t que fournisseur de mod&#232;les. </p><h4>Zhipu AI</h4><p><a href="iframe.php?url=https%3A%2F%2Fz.ai%2Fblog%2Fglm-5">Zhipu AI sort GLM-5</a> en open-source sous licence MIT. Le positionnement est clair : les t&#226;ches agentiques longue dur&#233;e, l&#224; o&#249; le mod&#232;le tutoie Claude Opus 4.5 sur les benchmarks de coding et de planification. </p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21WxU6%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!WxU6!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 424w, https://substackcdn.com/image/fetch/$s_!WxU6!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 848w, https://substackcdn.com/image/fetch/$s_!WxU6!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 1272w, https://substackcdn.com/image/fetch/$s_!WxU6!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21WxU6%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png" width="616" height="419.2692307692308" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:991,&quot;width&quot;:1456,&quot;resizeWidth&quot;:616,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!WxU6!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 424w, https://substackcdn.com/image/fetch/$s_!WxU6!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 848w, https://substackcdn.com/image/fetch/$s_!WxU6!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 1272w, https://substackcdn.com/image/fetch/$s_!WxU6!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F094a1ca5-42e2-45fc-b7cc-b348a12b27ab_4239x2884.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><h4>MiniMax</h4><p><a href="iframe.php?url=https%3A%2F%2Fwww.minimax.io%2Fnews%2Fminimax-m25">MiniMax frappe fort avec M2.5</a>, un mod&#232;le frontier chinois affich&#233; &#224; 1/10e du prix de Claude Opus, Gemini 3 Pro ou GPT-5. La version Lightning crache 100 tokens par seconde pour 0,30 $/M en entr&#233;e et 2,40 $/M en sortie. Sur les benchmarks de code agentique, le mod&#232;le fait jeu &#233;gal avec les meilleurs. MiniMax affirme que 80 % du code nouvellement commit&#233; en interne est g&#233;n&#233;r&#233; par M2.5. La guerre des prix sur les mod&#232;les frontier vient de changer d&#8217;&#233;chelle.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Qdf-%21%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!Qdf-!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 424w, https://substackcdn.com/image/fetch/$s_!Qdf-!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 848w, https://substackcdn.com/image/fetch/$s_!Qdf-!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 1272w, https://substackcdn.com/image/fetch/$s_!Qdf-!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21Qdf-%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png" width="590" height="284.4642857142857" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:702,&quot;width&quot;:1456,&quot;resizeWidth&quot;:590,&quot;bytes&quot;:null,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:null,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:null,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!Qdf-!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 424w, https://substackcdn.com/image/fetch/$s_!Qdf-!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 848w, https://substackcdn.com/image/fetch/$s_!Qdf-!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 1272w, https://substackcdn.com/image/fetch/$s_!Qdf-!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5b2a926a-e9dc-4ca8-a59d-d6152d9dee6a_7216x3480.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button tabindex="0" type="button" class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!SciE!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 424w, https://substackcdn.com/image/fetch/$s_!SciE!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 848w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1272w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21SciE%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png" width="1456" height="140" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/e4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:140,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:529764,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!SciE!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 424w, https://substackcdn.com/image/fetch/$s_!SciE!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 848w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1272w, https://substackcdn.com/image/fetch/$s_!SciE!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fe4d45429-e4cf-4a63-9371-27ed706d5c1d_1854x178.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><p>&#129297; <a href="iframe.php?url=https%3A%2F%2Fwww.tradingsat.com%2Falphabet-google-US02079K1079%2Factualites%2Falphabet-google-grace-a-ses-initiatives-dans-l-ia-alphabet-a-depasse-pour-la-premiere-fois-de-son-histoire-les-4000-milliards-de-dollars-en-bourse-1151340.html">Alphabet franchit les 400 milliards de dollars de revenus, Gemini en fer de lance pour les ann&#233;es &#224; venir</a></p><p>&#128176; <a href="iframe.php?url=https%3A%2F%2Fsiliconangle.com%2F2026%2F02%2F10%2Fworld-model-startup-runway-closes-315m-funding-round%2F">Runway l&#232;ve 315 M$ pour ses world models</a></p><p>&#129302; <a href="iframe.php?url=https%3A%2F%2Fwww.anthropic.com%2Fengineering%2Fbuilding-c-compiler">16 agents Claude ont construit un compilateur C from scratch</a></p><p>&#127760; <a href="iframe.php?url=https%3A%2F%2Fresearch.google%2Fblog%2Fhow-ai-agents-can-redefine-universal-design-to-increase-accessibility">Google veut que l&#8217;IA adapte les interfaces aux handicaps en temps r&#233;el</a></p><p>&#128717;&#65039; <a href="iframe.php?url=https%3A%2F%2Ftechcrunch.com%2F2026%2F02%2F02%2Felon-musk-spacex-acquires-xai-data-centers-space-merger%2F">SpaceX acquiert xAI dans la plus grosse fusion de l&#8217;histoire</a> </p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21lNev%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png" width="1456" height="139" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:139,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:764352,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!lNev!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 424w, https://substackcdn.com/image/fetch/$s_!lNev!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 848w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1272w, https://substackcdn.com/image/fetch/$s_!lNev!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F414d1f48-5df2-478a-a08e-b606d4585bcf_2260x216.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Fevalsp26.classes.andrewheiss.com%2Fnews%2F2026-02-03_faqs_week-3.html">What&#8217;s the difference between statistical significance and substantial significance?</a></strong></h3><p>Un p-value &#224; 0.001, et on sabre le champagne. Sauf que derri&#232;re ce r&#233;sultat &#171; ultra-significatif &#187;, l&#8217;effet mesur&#233; est parfois de... 0.51 dollar par mois. On a tous crois&#233; ce coll&#232;gue qui brandit sa p-value comme un troph&#233;e sans se demander si le r&#233;sultat a le moindre impact concret.</p><p>Andrew Heiss illustre cette distinction fondamentale avec un exemple limpide : deux RCTs &#233;valuant un programme de r&#233;insertion. Le premier montre un effet de 137 dollars mensuels (p = 0.041), le second un effet de 0.51 dollar (p = 0.001). Les deux sont statistiquement significatifs. Seul le premier compte vraiment. La significativit&#233; statistique r&#233;pond &#224; la question &#171; l&#8217;effet est-il non-nul ? &#187;. La significativit&#233; substantielle, elle, demande &#171; est-ce que &#231;a change quelque chose en pratique ? &#187;. Et cette deuxi&#232;me question, aucun test statistique ne peut y r&#233;pondre &#224; votre place.</p><p>Un rappel salutaire, surtout dans un monde o&#249; les papiers de ML empilent les &#233;toiles de significativit&#233; sans jamais questionner la taille de l&#8217;effet. La connaissance du domaine et le bon sens restent irrempla&#231;ables.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21IQ4f%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:21460,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/webp&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/187769190?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!IQ4f!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 424w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 848w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1272w, https://substackcdn.com/image/fetch/$s_!IQ4f!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F0670bf39-3ab8-4abd-b926-050226679daf_1456x138.webp 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Fpub.towardsai.net%2Fbuilding-production-text-to-sql-for-70-000-tables-openais-data-agent-architecture-bcd695990d55">Building Production Text-to-SQL for 70,000+ Tables: OpenAI&#8217;s Data Agent Architecture</a></strong></h3><p>OpenAI vient de d&#233;tailler l&#8217;architecture de son agent data interne.</p><p>Leur syst&#232;me navigue un warehouse de 600 PB et plus de 70 000 tables pour 3 500 utilisateurs internes. L&#8217;architecture repose sur six couches de contexte (lineage, annotations humaines, enrichissement par le code des pipelines, documentation institutionnelle, m&#233;moire conversationnelle, inspection runtime) aliment&#233;es par un pipeline offline quotidien. C&#244;t&#233; ex&#233;cution, une boucle ferm&#233;e de self-correction profile chaque r&#233;sultat interm&#233;diaire avant de le valider : row count suspect, explosion de lignes apr&#232;s un join, taux de nulls anormal ? L&#8217;agent diagnostique, corrige et relance.</p><p>Deux choix de design retiennent l&#8217;attention. D&#8217;abord, le passage syst&#233;matique par les credentials utilisateur (pass-through permissions) plut&#244;t qu&#8217;un service account privil&#233;gi&#233;. Ensuite, l&#8217;&#233;valuation continue via des paires question/SQL de r&#233;f&#233;rence ex&#233;cut&#233;es en CI, comparant les r&#233;sultats effectifs plut&#244;t que les requ&#234;tes en elles-m&#234;mes.</p><p>Le vrai enseignement : la qualit&#233; d&#8217;un agent data est born&#233;e par la richesse de son contexte, pas par le mod&#232;le sous-jacent. </p><h3><strong><a href="iframe.php?url=https%3A%2F%2Fmedium.com%2Fdata-science-collective%2Four-agent-had-a-4-minute-loop-heres-how-we-fixed-it-40a8142ef1a9">Our Agent Had A 4 Minute Loop. Here&#8217;s How We Fixed It</a></strong></h3><p>On monitore nos mod&#232;les ML en production depuis des ann&#233;es, mais quand un agent IA se met &#224; boucler pendant 4 minutes au lieu de r&#233;pondre en 18 secondes, les r&#233;flexes classiques ne suffisent plus. Monte Carlo, qui op&#232;re trois agents en production devant 30 000 utilisateurs, partage un retour d&#8217;exp&#233;rience tr&#232;s concret sur le sujet.</p><p>Leur m&#233;thodologie en cinq &#233;tapes est classique sur le papier : instrumenter la t&#233;l&#233;m&#233;trie avec OpenTelemetry, poser des KPIs par use case, isoler les outliers, inspecter les traces pour trouver les spans fautifs, puis croiser les signaux. L&#224; o&#249; &#231;a devient int&#233;ressant, c&#8217;est dans le diagnostic final. Le coupable n&#8217;&#233;tait ni un prompt mal ficel&#233;, ni un mod&#232;le trop lent : un client avait d&#233;sactiv&#233; une permission, et l&#8217;agent s&#8217;acharnait en boucle sans jamais fail fast. Un edge case classique des syst&#232;mes non-d&#233;terministes.</p><h3><strong><a href="iframe.php?url=https%3A%2F%2Fdocs.google.com%2Fpresentation%2Fd%2F19F-XvNJ8sgIpIeIduA3PhbsWp4pC-P632J2eJV1cLG8%2Fedit%3Fslide%3Did.g3b660485fe2_0_26%23slide%3Did.g3b660485fe2_0_26">Column Storage for the AI Era</a></strong></h3><p>Les formats colonnaires comme Parquet ont &#233;t&#233; la colonne vert&#233;brale du data engineering pendant une d&#233;cennie. Mais avec l&#8217;explosion des workloads IA, les limites apparaissent : scans s&#233;quentiels massifs pour l&#8217;entra&#238;nement, gestion laborieuse des embeddings en haute dimension, et des I/O qui deviennent le vrai bottleneck quand le GPU attend ses donn&#233;es.</p><p>Cette pr&#233;sentation pose la question frontalement : le stockage en colonnes tel qu&#8217;on le conna&#238;t est-il encore adapt&#233; aux pipelines ML modernes ? L&#8217;argument central tourne autour de l&#8217;&#233;volution n&#233;cessaire des formats de stockage pour supporter nativement les vecteurs d&#8217;embeddings, le versioning de datasets et les lectures al&#233;atoires rapides que r&#233;clament les boucles d&#8217;entra&#238;nement. On pense &#233;videmment &#224; des formats comme Lance, con&#231;u d&#232;s le d&#233;part pour ces contraintes, ou aux extensions vectorielles qui fleurissent dans l&#8217;&#233;cosyst&#232;me Arrow.</p><p>Le sujet est loin d&#8217;&#234;tre cosm&#233;tique. Quand on passe son temps &#224; convertir des donn&#233;es entre formats pour satisfaire chaque &#233;tape du pipeline (ingestion, feature store, training, serving), c&#8217;est autant de complexit&#233; op&#233;rationnelle et de surface d&#8217;erreur. Un format unifi&#233; pens&#233; pour l&#8217;&#232;re IA, c&#8217;est la promesse d&#8217;un stack plus simple. Reste &#224; voir lequel s&#8217;imposera.</p><div class="captioned-image-container"><figure><div class="image-link image2" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw"><img src="iframe.php?url=https%3A%2F%2Fsubstackcdn.com%2Fimage%2Ffetch%2F%24s_%21O1lW%21%2Cw_1456%2Cc_limit%2Cf_auto%2Cq_auto%3Agood%2Cfl_progressive%3Asteep%2Fhttps%253A%252F%252Fsubstack-post-media.s3.amazonaws.com%252Fpublic%252Fimages%252F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png" width="1456" height="138" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:138,&quot;width&quot;:1456,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:731194,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:true,&quot;topImage&quot;:false,&quot;internalRedirect&quot;:&quot;https://humanframework.substack.com/i/184346826?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!O1lW!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 424w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 848w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1272w, https://substackcdn.com/image/fetch/$s_!O1lW!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F577ca34d-cd0a-45c9-a4d0-d12c67ad1b13_2262x214.png 1456w" sizes="auto, 100vw" loading="lazy"></picture><div></div></div></div></figure></div><h3><strong><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fabs%2F2602.03837">Accelerating Scientific Research with Gemini: Case Studies and Common Techniques</a></strong></h3><blockquote><p>Recent advances in large language models (LLMs) have opened new avenues for accelerating scientific research. While models are increasingly capable of assisting with routine tasks, their ability to contribute to novel, expert-level mathematical discovery is less understood. We present a collection of case studies demonstrating how researchers have successfully collaborated with advanced AI models, specifically Google's Gemini-based models (in particular Gemini Deep Think and its advanced variants), to solve open problems, refute conjectures, and generate new proofs across diverse areas in theoretical computer science, as well as other areas such as economics, optimization, and physics. Based on these experiences, we extract common techniques for effective human-AI collaboration in theoretical research, such as iterative refinement, problem decomposition, and cross-disciplinary knowledge transfer. While the majority of our results stem from this interactive, conversational methodology, we also highlight specific instances that push beyond standard chat interfaces. These include deploying the model as a rigorous adversarial reviewer to detect subtle flaws in existing proofs, and embedding it within a "neuro-symbolic" loop that autonomously writes and executes code to verify complex derivations. Together, these examples highlight the potential of AI not just as a tool for automation, but as a versatile, genuine partner in the creative process of scientific discovery.</p></blockquote><h3><strong><a href="iframe.php?url=https%3A%2F%2Farxiv.org%2Fpdf%2F2602.02874">Ten simple rules for teaching data science</a></strong></h3><blockquote><p>Teaching data science presents unique challenges and opportunities that cannot be fully addressed by simply borrowing pedagogical strategies from its parent disciplines of statistics and computer science. Here, we present ten simple rules for teaching data science, developed and refined by leading educators in the community and successfully applied in our own data science classrooms.</p></blockquote>]]></content:encoded></item></channel></rss>