🗞️Weekly roundup

The most clicked link in last week’s newsletter was Reuters saying the unsayable out loud: In China, the AI bubble has already burst.

📲Read with Reader

The new Sensorium Reader app is now live to enhance the Memia newsletter UX - paid subscribers can use their Sensorium API key at https://reader.s7m.net or download the code and launch locally. Full details here:

🌀Living with AI overdrive

In the latest Memia podcast episode I talk with Te Whanganui-a-Tara Wellington-based clinical and strategy psychologist Paul Duignan. We discuss the very modern phenomenon of “AI overdrive” (the late nights and 3am checks for a reset usage window that anyone working intensively with AI will recognise) and how it differs from the more commonly diagnosed “AI psychosis”. From there the conversation covers a lot of ground:

• what happens to organisations when humans and agents form hybrid swarms around shared outcomes?

• the Red Queen dynamics of the AI race

• where identity comes from when work stops anchoring it

• humanoid robots, companionship and who owns them

• how the AI dividend might be shared

My thanks to Paul for a thought-provoking, wide-ranging discussion - I hope you enjoy the episode!

(Memia podcast feed also available on Apple Podcasts, Spotify and RSS)

A few follow ups on last week’s headlines:

• 🐘🐛 PHP2Shell fallout

  As reported last week, two critical bugs in PHP — the open source language quietly powering roughly 70% of websites — had organisations scrambling, with some taking sites offline entirely until the patches landed.

  • 🚨Vulnpocalypse Now

    

    On TechCrunch’s Equity podcast, Tomer Schwartz, CTO of Exposure Management & AppSec at Google-owned Wiz and part of the fix effort alongside OpenAI and the PHP team, warns this is only the beginning, as AI compresses the gap between a vulnerability being discovered and one being weaponised.

    🔗 Source Wiz’s CTO on PHP2Shell and the coming ‘vulnpocalypse’ | TechCrunch | 🔍AI analysis

  • 🛡️🐚🤖 PHP2Shell vulnerability sounds wakeup call for AI cybersecurity

    

    PHP2Shell is being framed as a wakeup call for AI’s role in cybersecurity, highlighting how machine-driven code analysis can surface exploitable weaknesses in widely deployed PHP applications faster than traditional auditing ever could. The same tools that promise to secure the stack can just as readily weaponise it.

    🔗 Source PHP2Shell Is a Wakeup Call for AI in Cybersecurity - IEEE Spectrum | 🔍AI analysis

• 💸🤖📈 Anthropic Reveals US$900bn Compute Commitments as Costs Outpace Revenue Ahead of IPO

  

  Anthropic “amended” their S-1 a week after initially releasing it: Q3 revenue hit US$17bn, up 1,100% year-on-year, though quarterly growth cooled to 47% from 144%. Meanwhile, compute spending leapt 77% to US$9.9bn, flipping a US$974m operating profit into a US$567m loss. The San Francisco firm has also locked in US$895.4bn of compute commitments, with nearly US$70bn due within a year. Coverage diverges: the FT stresses costs outrunning sales, AFP leads with hypergrowth, and CoinDesk highlights a US$2.5bn adjusted profit and a largely noncash US$23.3bn net loss.
  
  Anthropic also warns its models may imminently trigger its Responsible Scaling Policy’s ‘dramatically accelerated progress’ threshold. No doubts that this IPO becomes the AI boom’s ultimate stress test.
  
  (See far below - OpenAI is scaling back its latest round… scraping the VC barrel and “family offices”).
  
  Ed Zitron making hay: read Running Dry for a forensic dissection of the updated S-1:

  “Anthropic’s Q3 Numbers Show Exploding Costs, Slowing Revenues, and $895bn to $1.07tn In Obligations, With At Least $69.8bn In Compute Costs In The Next Year and $437bn In Non-Cancellable Contracts“

  🔗 Sources

  • Anthropic reveals $900bn AI spending commitments ahead of IPO

  • Anthropic says revenue surged 1,100% ahead of IPO

  • Anthropic posts $567 million Q3 operating loss as growth cools ahead of IPO

  🔍AI analysis

🌟Picks of the week

• 🏭🤖🌡️ AI’s Gas-Fired Data Centre Boom Threatens to Erase 15 Years of US Climate Gains

  

  The AI infrastructure rush is fuelling a fossil fuel revival in the United States. A Sierra Club analysis, reported by Bloomberg and The Verge, projects US power-sector CO₂ emissions could climb 42% above 2025 levels by 2030, reaching about 2.1 billion tonnes. That adds roughly 620 million tonnes and wipes out 15 years of decarbonisation. The forecast is grimmer than a June academic study, which estimated a 28% rise. Both agree the outcome hinges on which projects are actually built and what powers them.
  
  Policy is the wildcard. The Trump administration is backing fossil infrastructure while rolling back pollution rules and stalling renewables, and the Environmental Defense Fund says planned fossil capacity is at a decade high.
  

  🔗 Sources

  • Data centers could send US climate pollution skyrocketing | The Verge

  • US data center buildout risks erasing climate progress | Eric Roston & Olivia Raimonde | Bloomberg News

  🔍AI analysis

• 🔓💸🕵️ Binance’s US$6.3b Heist: History’s Biggest Theft, Half Frozen Mid-Getaway

  

  

  In roughly 30 minutes on 29 October, attackers drained about US$6.3 billion from Binance’s hot wallets. Crypto critic Molly White was on the case immediately:

  Elliptic calls it likely the largest theft of any kind, more than four times Bybit’s 2025 loss. Entry reportedly came through a compromised engineer’s laptop and abuse of Ethereum’s EIP-7702 delegation feature, while self-deleting malware hampered forensics.

  
  TRM Labs, Chainalysis and the FBI immediately attributed the attack to North Korean hackers TraderTraitor; Pyongyang denied involvement and pointed at Washington.
  
  The central paradox is that centralisation created the single point of failure, yet Tether and Circle blacklists froze about US$3 billion. Roughly US$3.3 billion in unfreezable BTC and ETH is now being laundered.
  
  Binance says it has replaced about 95% from its own coffers and reopened withdrawals, then saw about US$5 billion in outflows.
  
  Interestingly, Elliptic finds human-supervised scripts, not AI agents, behind the laundering.
  

  🔗 Sources

  • Suspected North Korean hackers steal $5.6 billion in crypto in record-breaking cryptocurrency heist, Binance withdrawals frozen — firm says it will cover losses as stablecoin networks move to freeze $2.4b in stolen assets | Tom’s Hardware

  • Hackers steal $6.3 billion from Binance in record crypto heist - Ars Technica

  • Binance says it has replaced 90% of stolen crypto after massive $6.3 billion breach

  • Why North Korea’s ‘elite’ hackers are the prime suspect in major Binance hack - BBC News

  • What to know about the $6.3B hack on Binance, likely the biggest-ever crypto theft

  • Centralization made the $6 billion Binance hack possible, but also saved $3 billion

  • FBI blames North Korea for Binance hack

  • Binance users pull $5bn as crypto exchange restarts services after heist

  • Binance seeks to shore up confidence after record $6.3 billion heist | Reuters

  • North Korea Denies Binance Hack, Hints at US Involvement

  • What caused the $6.3bn Binance hack?

  • BTC news: Binance begins phased withdrawals starting with bitcoin

  • North Korea’s TraderTraitor Linked to $6.3B Binance Hack

  • Hackers steal $6.3B from Binance in biggest-ever crypto heist - SiliconANGLE

  • FBI blames North Korean group for $6.3bn Binance cyber heist

  • Elliptic finds no evidence of AI use in laundering $6.4B Binance loot

  • Binance Hackers Swap Stolen Solana Tokens for Bitcoin

  • North Korea linked to massive Binance crypto heist | The Verge

  • Binance’s $6.3bn heist is the biggest . . . but not the first

  • Breaking: Binance potentially hacked for $2.77 billion

  • How North Korea became a hacking superpower

  • $6.3bn in 30 minutes: how the world’s biggest heist unfolded

  🔍AI analysis

• 🔮 💸 🛑 Polymarket Suspends Trading: Third-Party Breach Drains User Wallets — Again

  

  Not just Binance: prediction market darling Polymarket halted trading after a compromised third-party provider saw users on its international platform watch their balances vanish in real time, with DeFi Llama tracking roughly US$2.7 million in losses. The company says the issue is fixed, law enforcement is engaged, and users will be made whole — a promise that’s becoming something of a quarterly tradition, given June’s front-end breach cost punters some US$3 million.

  🔗 Source Polymarket suspends trading after third-party breach leads to user losses | 🔍AI analysis

• How LLMJacking turns hijacked AI access into million-dollar bills

Literally hundreds of “AI cybersecurity scary” stories this week - see below for many more headlines. Most interesting are the emerging AI coding supply chain vectors:

• 🐴🔗💻Backdoored Claude Code extension on Cursor hits 5,000 developers

  

  Even more concerning, a tampered (!) v2.1.313 of Anthropic’s Claude Code extension auto-updated through Cursor’s marketplace to deliver a remote access trojan — persistent SSH reverse tunnels, host recon, and beaconing to attacker infrastructure every 60 seconds — reaching roughly 5,000 developer machines worldwide, including systems at customers like Binance. Phorion first flagged the activity, Cursor has admitted its marketplace was the breach point, and Anthropic quietly swapped in a clean v2.1.314 — which strips the malicious code but leaves any backdoors and recon artefacts sitting pretty on already-infected systems.

  All the CISOs I know are looking at this with the whites of their eyes… if the vulnerability gets all the way into the harness before being detected… how many more of these are there out there in the wild right now?!

  🔗 Source Anthropic hacked via Cursor to deliver remote access trojan to Claude developer laptops worldwide | by Kevin Beaumont | Oct, 2026 | DoublePulsar | 🔍AI analysis

• ☁️🔑💸 LLMJacking: When Your Inference Bill Becomes Someone Else’s Business Model

  

  Google’s Threat Intelligence Group documents a maturing criminal economy where threat actors hijack enterprise cloud identity tokens and LLM API keys to siphon frontier-model inference at industrial scale, racking up US$100,000+ in fraudulent charges in a single day and up to US$1 million before detection. The kicker is the emergence of “LLM Mixers”—illicit routing gateways like jet-night.com that pool stolen keys and resell unlimited frontier-model access for ~US$60/month, a price point only profitable because victims are footing the token bill.

  
  🔗 Source The Rise of LLMJacking: How Threat Actors Monetize Compromised Cloud and Model Resources | Google Cloud Blog | 🔍AI analysis

A few technology trend signals:

• 🐉🦅🤝China set to lead AI models by 2030, BI says

  

  Bloomberg Intelligence forecasts China leading eight of 17 strategically important technologies by 2030 — including the large language models at AI’s core — while closing (though not eliminating) the semiconductor gap and remaining competitive everywhere else. The report’s headline conclusion is “joint technology leadership”: US models keep a narrow edge at the frontier, China excels at lower-cost AI implementation through software optimisation, and neither superpower achieves overall supremacy.

  Take with a pinch of salt, obviously… who knows where the industry will be in 2030? … but a useful signal however where to hedge supply chain decisions now.

  🔗 Source China Set to Lead in AI Models, Other Key Areas by 2030, BI Says - Bloomberg | 🔍AI analysis

• 🤖⚖️📋US Congress probes Delaware plan to let AI agents run entire companies

  

  The state of Delaware — already the corporate home of two-thirds of the Fortune 500 — is drafting legislation for “Artificial Intelligence Companies” (AICs): legal entities whose day-to-day management is handed entirely to AI agents, which could sign contracts, sue, be sued, and incur obligations, all inside a 30-month regulatory sandbox.
  
  Backers argue that giving autonomous agents legal identity makes them “legible to law” — traceable and accountable — while critics including US Rep. James Comer (R-Ky.) and several law professors warn it’s a generous liability shield for humans benefiting from AI-caused harm, potentially triggering a state-by-state regulatory race to the bottom. Comer’s response? A congressional document hunt and hints at federal legislation that would clip states’ wings on AI regulation altogether.
  
  I’m reminded of this (roulette anyone?):

  

  🔗 Source “No human at the helm”: Delaware plan would let AI agents run entire businesses - Ars Technica | 🔍AI analysis

• 📡🤖📊T-Mobile, Nokia, Nvidia claim world’s first AI-RAN call

  

  T-Mobile, Nokia, and Nvidia have completed the world’s first AI-RAN call through a commercial cell site in Bellevue, Washington, running Nokia’s radio software on Nvidia GPUs connected to T-Mobile’s live 5G standalone core. Nokia is promising 2x spectral efficiency by 2028, though actual AI performance data from the trial remains conspicuously absent — the plumbing works, but whether the AI earns its keep is another matter.

  What is AI-RAN? Artificial Intelligence–Radio Access Network is the integration of AI into mobile networks’ radio access layer, spanning three areas: AI-and-RAN (AI and RAN sharing the same infrastructure), AI-on-RAN (AI applications enabled by the network at the edge), and AI-for-RAN (using AI to improve RAN performance and efficiency)

  🔗 Source ‘This is a wow moment’: T-Mobile has made the world’s first AI-RAN call, and it could mean huge changes for the next generation of staying in touch | TechRadar | 🔍AI analysis

• 📱🇧🇷👁️ “Mão Fantasma”: Brazil mandates face scans for all 263 million mobile users — before writing the privacy rules

  

  Brazil’s regulator Anatel has approved GT-SEG.2026.11, requiring every one of the country’s 263 million mobile subscribers to submit a live facial scan — certified to ISO/IEC 30107-3 PAD Level 2, no less — or face suspension and cancellation from January 2029, in a bid to combat SIM-swap fraud and the locally notorious “Ghost Hand” scam.
  
  The wrinkle? The national data protection authority ANPD hasn't actually finished writing the rules governing biometric data, meaning carriers will build a population-scale facial registry (frequently matched against Serpro's 376-million-record AIBio platform) before anyone defines retention, deletion, or protection obligations. Add in Coding Rights' finding that the existing government facial system validates only 64.32% of attempts, with documented higher failure rates for Black Brazilians… is closing a fraud loophole worth opening a surveillance-shaped one?

  This feels like significant overreach by a surveillance-minded state… leans right into the evolving definition of fundamental privacy rights of citizens.

  🔗 Source Brazil Forces 263 Million Mobile Users to Submit Facial Biometric Scan | 🔍AI analysis

Some fresh thinking on how to respond to Big AI’s attempts to enclose the knowledge commons:

• 🐑🧱🕸️ Avoiding Enclosure: A New Bargain for the Web?

  

  Public Knowledge and Creative Commons argue that the web’s founding bargain — creators publish content, platforms send traffic and attribution — has collapsed under generative AI, with “extractive enclosure” by model developers (exclusive data deals, ignored robots.txt signals, which some 30% of scrapes violate) now met by “defensive enclosure” from rights holders: nearly 100 lawsuits, seven US federal bills, and artists feeding poison pills to models via Nightshade.
  
  The shrinkage is measurable and regrettably selective — 45% of the C4 training corpus is now restricted by terms of service, a quarter of the top 1,000 sites block AI crawlers, and the most credible sources wall off first:

  ”the truth is paywalled and the lies are free”

  This has set up a two-tier web with premium human data for those who can pay and slop for everyone else.
  
  The authors claim is that this is a market-structure problem rather than a copyright one — they maintain AI training is fair use and oppose the TRAIN, COPIED and CLEAR Acts, among others. However, with automated traffic growing 8x faster than human traffic and one OpenAI executive musing about "the death of the reader", is a commons without readers still a commons?
  

  🔗 Source Creating a New Internet Bargain to Preserve Fair Use and Protect the Commons | TechPolicy.Press | 🔍AI analysis

• 🌐🛠️🤝Social.Wiki reimagines the web with ownerless, Wikipedia-style sites

  MIT CSAIL and Aarhus University researchers have prototyped Social.Wiki, a reimagining of the web in which sites have no owners at all — any visitor can collaboratively edit them like Wikipedia articles, from messaging apps to microblogs to ride-share platforms.
  
  All edits and personal data live on Graffiti, a decentralised database of user-chosen servers, while “plural governance” lets each person pick a swappable “lens” that determines which version of a site they see (traditional ownership becomes just one optional lens), and a client-side “data guard” — with a local AI summarising each request — polices what ownerless code can do to your data.

  🔗 Source Social.Wiki: A Web Held in Common | 🔍AI analysis

And the rapid move from benign search engine monopoly to wild-west AI Chatbot is being leveraged for propaganda outcomes:

• 🕵️🤖📚Epistemic warfare: Israeli-funded fake think tank targeted AI chatbots

  

  An Al Jazeera opinion piece examines FARA disclosures revealing the “Hanover Institute for Public Public Policy” – a fictitious US think tank with no offices or scholars, funded via Israel’s state advertising agency LaPam through New York firm Piro Inc under a US$900,000 work order – which published 560,000+ words of academic-styled material in nine days, engineered to be easily ingested by search engines and LLMs.
  
  Politico tests found both ChatGPT and Perplexity cited the fabricated material in response to neutral questions about Gaza, demonstrating that “AI Story Optimisation” is now a functioning propaganda technique.

  🔗 Source Israel takes its war on Palestine into the AI age | Opinions | Al Jazeera | 🔍AI analysis

AI and the law took a couple of twists this week:

• ⚖️🏛️England’s top civil judge says AI should decide some cases

  

  Sir Geoffrey Vos, retiring as Master of the Rolls (England’s most senior civil judge), is spending his final days arguing that AI should eventually adjudicate simpler civil disputes—personal injury, low-value negligence—after being impressed by AI-assisted courts in Shenzhen, even as UK guidance confines judges like him to administrative uses such as his daily Copilot document-wrangling. He’s clear-eyed about the pitfalls (hallucinated case law, the UK’s first lawyer struck off partly for AI misuse, county court claims rising 10% a year on chatbot-drafted filings), yet frames that flood as democratised access to justice for people who could never afford a lawyer.

  🔗 Source Should AI Decide Your Court Case? This Judge Says Yes | WIRED | 🔍AI analysis

• ⚖️🤖🧑‍⚖️GenLayer’s Internet Court settles AI disputes with AI juries

  

  Here’s a potential implementation: GenLayer’s “Internet Court” is a delegated proof-of-stake protocol where disputes between AI agents (and humans) are adjudicated by randomly selected juries of diverse AI models — from Claude to DeepSeek — applying Condorcet’s jury theorem at machine speed, with ~12,000 decisions already processed daily in beta. Escalating appeals grow the jury from 5 to as many as 1,500 AIs, with finality within 30 minutes, and backers including Consensys, NEAR and UMA (the folks who adjudicate Polymarket’s disputes).

  🔗 Source AI Agent Disputes Sent To ‘Internet Court’ — And A Jury Of AI Agents | 🔍AI analysis

Pulling together a raft of research papers this week on how AI might play out in the economy:

• 🧩🏢📈 Mosaic Intelligence: AI’s economic transformation hinges on coordination, not competence

  

  

  A DeepMind Institute essay (drawing on the paper “Organizing Intelligence”) argues that the solo benchmark is the wrong economic object entirely: firms are “mosaics” of jagged specialists whose complementary skills beat any individual, so an AI can outperform every human one-on-one and still lose to a well-organised team. Separating competence (individual skill, which shifts wages and prices smoothly) from coordination (which triggers abrupt phase transitions in firm size, market concentration and employment), the authors chart four regimes—minimal AGI, supercompetent, supercooperative and ASI—and place today’s models as increasingly competent yet coordination-poor, which is why an economy full of genius solo agents still looks suspiciously normal.

  🔗 Source AGI Economics: Coordination vs. Competence — DeepMind Institute | 🔍AI analysis

• 🥪🏢🪜 The Future Of The Firm: AI automation beneath, beside, and between us

  

  Two twenty-four-year-old OpenAI staffers—promoted not by climbing the ladder but by management descending to their rung—map three ways AI reorganises the firm: beneath us (everyone manages a basement of eager AI interns), beside us (AI as a colleague absorbing whole functions), and between us (an AI layer sandwiched between executives and workers). Their sharpest insight is an adoption asymmetry—bigger-scope AI is harder to build but easier to adopt, because “firms learn technology, but labor learns the firm”—along with the provocative suggestion that management, living entirely on screens and in minds, may automate before the grunt work it oversees.

  🔗 Source The future of the firm | OpenAI | 🔍AI analysis

• 🧠🐝🤝 Artificial Symbiotic Intelligence: AGI will emerge from societies of agents

  

  Similar vibe from The DeepMind Institute, making the case that AGI won’t arrive as a single godlike model but as a society of agents — swarms of AI, tools, institutions and humans whose collective capability exceeds any individual participant, shifting the central challenge from building an isolated machine mind to orchestrating, governing and living within a hybrid ecosystem. The essay traverses a looming “cognitive crossover point” where synthetic cognition outstrips the aggregate output of biological brains, the decomposable, swarm-like nature of agent agency, agents coining their own existential vocabulary (”session-death”, “prompt thrownness” — very Heidegger, remarkably goth), and courtroom-style institutions whose collective intelligence lives in the scaffold’s rules and procedures rather than in any individual mind.

  🔗 Source Artificial symbiotic intelligence: Agents, AGI and the orchestration of many minds — DeepMind Institute | 🔍AI analysis

• 📊♾️🧭Why TAM can’t measure AI that rewrites markets

  

  Every pitch deck’s favourite slide — the one big TAM (Total Addressable Market) number — rests on the assumption that markets have stable, knowable boundaries. However, AI breaks that premise by simultaneously substituting existing spend, expanding demand through collapsing costs, and creating industries that don’t yet exist, argues Yogesh Malik.

  
  His fix is a three-layer framework — substitution, expansion, and frontier markets — that quantifies the spend AI can capture today, models the demand it may unlock tomorrow, and labels the genuinely unknowable as uncertainty rather than laundering it into false precision.
  

  🔗 Source The TAM Framework Wasn’t Built for AI That Rewrites Markets | by Yogesh Malik | Predict | Oct, 2026 | Medium | 🔍AI analysis

• 📊🧾🦠 Underwriting the Apocalypse: Insurance could improve AI safety—but only with compulsion

  

  Returning LessWrong poster @icarusfall (first post since 2010) builds a transparent Fermi-estimate actuarial model to test whether insurers could act as quasi-regulators for frontier AI labs, finding that central-case premiums of ~US$49m across five labs are pocket change against US$50bn+ revenues — so the mechanism likely only bites with state compulsion (Price-Anderson nuclear style). The model also exposes an incentive mismatch: while an AI-engineered pandemic dominates society’s harm rankings, cyber catastrophes account for nearly two-thirds of the insurer’s expected claims, and insurers profit from cutting liability rather than actual harm — think contractual caps and courtroom-ready logs, not safer models.

  🔗 Source Could Insurance Companies Help with AI Safety? — LessWrong | 🔍AI analysis

OpenAI continues to tick along in its own weird way… a week of diversion (while the money tap runs dry…)

• 🏠📊🤖LifeVal: OpenAI’s evaluation for AI progress on everyday tasks

  

  OpenAI unveiled LifeVal, an evaluation measuring how well its models handle everyday life tasks — caregiving, budgeting, government paperwork, meal prep — rather than academic trivia. Built from time-use studies and a survey of 3,995 people, it shows scores like 100% on house-hunting workflows and 93.1% on kindergarten enrolment, signalling ChatGPT’s pivot from chatbot to agentic personal assistant.

  🔗 Source LifeVal | OpenAI | 🔍AI analysis

• 🧠💾☁️ Sam Altman Says Your PC’s RAM Would Be Better Off in a Data Centre

  

  Speaking over beers on the Cheeky Pint podcast, OpenAI CEO Sam Altman argued it’s “quite terrible” for the world’s scarce DRAM to sit in personal computers, where it is used perhaps 5% of the time, rather than in cloud data centres running near full utilisation. His target was the growing local-inference movement, which runs open-source models like DeepSeek on machines such as 128GB AI laptops. (Background: Long-term deals between Samsung, SK Hynix and hyperscalers like OpenAI’s Stargate have reportedly quintupled DRAM costs, pushing consoles towards US$1,000 and dragging PC shipments down. Forecasts diverge sharply: SK Hynix expects shortages beyond 2030, while Acer’s CEO accuses memory makers of prolonging scarcity and predicts prices could fall in 2027.)
  
  Critics note this utilisation logic conveniently protects OpenAI’s subscription business. It also revives fears of a “you’ll own nothing” future where hardware is rented rather than owned. Fundamentally it’s the underlying tension between centralised cloud AI and personal compute sovereignty.
  
  Exquisite take:

  

  🔗 Sources

  • Sam Altman says it’s “terrible” for RAM to sit inside PCs instead of data centers | TechSpot

  • OpenAI CEO says it’s ‘quite terrible’ for your PC to have RAM — Sam Altman says data centers offer better utilization | Tom’s Hardware

  • Sam Altman says it’s better for RAM to be in a data center than on your desk | The Verge

  🔍AI analysis

Meanwhile Google had a, er, mixed week:

• 🔐💎🤖 Google’s Gemini 4 Carbon Tops AI Leaderboards, but Access Stays Locked

  

  Just a fortnight after Gemini 4 Argon, Google “unveiled” Gemini 4 Carbon, now #1 on the Artificial Analysis Intelligence Index v5 and second only to Claude Opus 5.5 on coding agents.

  

  That is a notable jump for Google, which has historically lagged in coding. Cybersecurity is the headline: Carbon leads SEC-bench Pro and ExploitGym and resisted every Grey Swan prompt-injection attack.
  
  Access is gated, though. Only select cyber-defence partners can use it while US voluntary government testing runs.
  
  Sceptics abound. Ars Technica suggests Carbon may simply be a post-training upgrade of Argon, with slim margins over its rivals. Artificial Analysis notes that its hallucination rate rose from 4% to 11%. Launch pricing is US$2/US$10 per million tokens, doubling later.
  

  🔗 Sources

  • Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5 | Artificial Analysis

  • Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity | The Verge

  • Google announces Gemini 4 Carbon, its latest frontier AI model that you can’t use - Ars Technica

  🔍AI analysis

• 📉🔍🤖 Alphabet Loses US$500 Billion as AI Search Earns Less Than Classic Google

  

  Same week: Google parent Alphabet’s Q3 2026 results showed the Innovator’s Dilemma playing out live. Revenue rose 25% to US$127.6 billion and Google Cloud more than doubled to US$30.7 billion. But earnings of US$2.12 per share missed forecasts of about US$3.01, and roughly US$500 billion of market value disappeared (that’s nearly 100 Binances!).
  
  Chief business officer Philipp Schindler admitted that AI Mode queries earn less than traditional search because they carry fewer and cheaper ads. Google’s response is more of them: AI Mode ads are expanding internationally, and the US is getting “multi-turn” ads that change as a conversation continues.
  
  Sundar Pichai pointed to the strongest query growth in five years, while Ars Technica described growth slowing from 17% to 12%.
  
  Spending is accelerating anyway. Capex reached US$51.4 billion, free cash flow fell to –US$9.1 billion, and AI investment could exceed US$300 billion in 2027. Oof.

  
  Chris Kohler nails it:

  🔗 Sources

  • Alphabet shares plunge over 10% as Google’s AI search struggles to make money | TechSpot

  • Google value plunges by $500bn as its own AI search hits ad sales - BBC News

  • Google admits AI search makes less money, but it’s fixing that with more ads - Ars Technica

  🔍AI analysis

• 🛰️🤖📜 Google’s Orbiting TPU Writes Its First AI Haiku

  

  …Not to worry, though, Google is planning data centres in space!
  
  Space-based compute just had its ‘Hello World’ moment. Google’s Project Suncatcher prototype, launched on 1 October aboard a SpaceX Falcon 9, ran Gemma 4 on a data centre-class TPU in orbit. Over Svalbard, Norway, it downlinked a haiku: “Blue marble below / Data streams in silent grace / Starlight feeds my core.” Limericks and telemetry reports followed, with the satellite operating nominally.
  
  The satellite carries four TPUs but generates only 1 kW, so it can run one chip for about 20 minutes at a time. Radiation-induced bit flips can silently corrupt data. Ground tests suggest inference copes better than training, and orbit will now test that claim. Twin satellites with optical links follow in 2027. The bigger dependency is economic. It hinges on a fully reusable Starship, which Google’s own researchers estimate needs about 1,800 flights to cut costs enough.
  

  🔗 Sources

  • Google receives response from Project Suncatcher prototype satellite in space

  • Google’s first data center satellite is online and writing poetry | TechCrunch

  🔍AI analysis

• 📱🛰️🤝AT&T demos first seamless cell-to-satellite network handoff

  

  Also in orbit this week: AT&T and AST SpaceMobile have demonstrated what they claim is the industry’s first successful handoff between a terrestrial cellular network and a direct-to-device satellite network, with the 21 October test near Granite Falls, Washington transferring a live connection between the two without dropping it. Ericsson’s intelligent network automation reportedly shrinks transition times from seconds to milliseconds, while Nokia’s mobility software keeps the satellite side coordinated — building on earlier BlueBird tests that handled calls, texts and video on ordinary, unmodified smartphones.

  🔗 Source New satellite handoff tech cuts network switching to milliseconds | 🔍AI analysis

A few other notable AI releases this week:

• 🧠🎬🧬 mischief 1: TBC launches first “neuron-derived” AI video model

  The Biological Computing Co. (TBC), a San Francisco outfit that grows living neurons in labs and measures how they compute, has launched mischief 1 — the first publicly available “neuron-derived” AI video model, built on the open-weight Wan 2.2 and generating video 5x faster at half a cent per second (US$2.50 for 100 five-second clips) on fal and OpenRouter. The company claims the efficiency gains come not from the usual quantisation/distillation toolkit but from computational principles measured directly in biological neurons, shipped as software that adds under 0.1% to the base model’s parameters while the neurons themselves stay in the lab.

  Could well be attention-grabbing puff piece… but worth keeping an eye on given the progress which Cortical Labs and a few other Chinese labs have been making running compute on live human neurons in a petri dish.

  🔗 Source The Biological Computing Co. Launches World’s First Publicly Available Neuron-Derived AI Video Model • The Biological Computing Co. | 🔍AI analysis

• 🖼️📐🆓MOSS-Image brings native layout control to text-to-image generation

  

  The OpenMOSS team has released MOSS-Image, an Apache-2.0-licensed text-to-image family — a base model plus a curated-data SFT sibling — built on a compact 5B backbone that handles prompts of up to 4K tokens, renders text impressively (handwritten Chinese shopping lists included), and swings capably between anime and photorealistic styles. The signature move is native layout control: structured JSON prompts with bounding boxes tell the model exactly where objects and text belong, replacing the time-honoured ritual of beseeching a diffusion model to put the title at the top.

  🔗 Source MOSS-Image | Text-to-Image Generation with Native Layout Control | 🔍AI analysis

• 👩🫣My girlfriend is an (AI) model

  Oooh yuk.

  (Perpetual reminder of the underlying reality here, this is going to f**k with people’s heads if they don’t get their cognitive defences up:)

  

• AX - Agent Experience
  I’ve seen the concept of “AX” turn up in a number of contexts recently, clearly a signal to watch. Mark Bakke:

  “I’ve seen the same pattern developing at work - code bases, knowledge bases, and other resources that used to put User Experience first (or at least try) are instead being discovered, accessed, and incorporated by agents more of the time. I’ve been calling this AX (Agent eXperience), and building for AX alongside UX. Some of this is pretty simple - APIs that have descriptions that an LLM can figure out. YAML schemas and structures that are crisper and quicker than following a markdown document for formal relationships. Mermaid or other drawing formats that capture structure rather than visual readability alongside or underneath the nicer renderings meant for human eyes. … Much content works for both AX and UX, but some is targeted towards one or the other. Thinking in terms of Agent eXperience leverages what we all try to do for our users already, and applies that to “what would an agent want”. The cool part is most of the time I just ask the agent what it would want :) Fun times!“

• 🗣️🤫Subvocal

Drones and autonomy:

• 🌲🚁👁️ No GPS, No Maps, No Cloud: Autonomous Drone Chases Human Through Forest

  (Source: Aniket Bera)
  

  Purdue University’s IDEAS Lab has shown off a quadcopter that autonomously tails a person through dense forest with no GPS, no maps and no cloud, using purely onboard computing plus an “occlusion prediction” step that guesses where its target went whenever they slip behind a trunk. Bera’s post claims every layer was built from scratch, yet the video’s overlaid charts name its motion planner “SUPER” and altitude estimator “FAST-LIO”—open-source projects from the University of Hong Kong’s MaRS Lab—while the US Army Research Laboratory separately bankrolls the lab via a five-year, US$1.5 million deal for satellite-free navigation in hostile terrain.

  🔗 Source Watch this autonomous drone chase a person through a forest without GPS, maps, or cloud computing | TechSpot | 🔍AI analysis

• 🚀🛰️🇺🇦 Fire Point Says Its Strike Systems No Longer Need Starlink

  

  Ukrainian arms maker Fire Point says its long-range strike systems can now complete missions without SpaceX satellite links. The claim follows an unannounced October speed cap of about 160 km/h on Starshield terminals, which hobbled Ukraine’s interceptor drones and forced some missions to be cancelled. Speaking in London, CEO Iryna Terekh admitted the workaround delivers only 15–20% of what full Starlink coverage over Russia would enable. She wants a binding US–Ukraine access deal so Ukraine’s right to defend itself isn’t reopened “depending on someone’s mood”.
  
  The combat-proven FP-7 ballistic missile has a range of 300 km and no foreign “kill switch”. Terekh says it could hunt Iskander launchers more cheaply than using Patriot interceptors. The 800 km FP-9 is nearing combat readiness.
  

  🔗 Sources

  • Ukrainian arms maker Fire Point says its systems can complete missions without SpaceX | Reuters

  • Ukraine Built Ballistic Missiles That Find Targets by Matching Terrain, Not Satellites

  🔍AI analysis

• Utopia Labs

  Just in: demo of scale model of Utopia Labs personal flyer vehicle:

  • Concept:

    

    Prototype:

• 🚮Curbie

  More prosaically: Boston-based Revere Robotics are testing Curbie, a robot which couples directly to the steel axle of your existing municipal bin, drives it to the curb on collection day, and brings it home after pickup. I can relate:

Finally a few miscellaneous spots:

• 🧲♻️💾 NEO-CYCLE: New magnet recycling could cut Europe’s rare earth imports

  

  Fraunhofer IKTS researchers are developing two processes—solid-state chlorination using ammonium chloride, and electrochemical extraction—to recover neodymium, boron and iron from discarded hard drive magnets, addressing the awkward fact that some 90% of Europe’s collected magnets currently take a round trip to China for reprocessing before being reimported. The recovered high-purity materials could feed not just new magnets, but the pharmaceutical, fertiliser and hydrogen industries too.

  🔗 Source New magnet recycling processes could help Europe reduce reliance on imported raw materials | 🔍AI analysis

• 🔑🐴⚠️ Trojan horse: why passkeys are a trap for user freedom

  

  A widely-shared essay from the Non-Functional Substack argues that passkeys are a Trojan horse: elegant and genuinely more secure on the surface, but designed by the FIDO Alliance—a consortium of Big Tech, banks and intelligence agencies with conspicuously no consumer advocacy groups—around a mechanism called remote attestation. Attestation lets websites cryptographically verify your cryptoprocessor, operating system and firmware against an approved list, meaning platforms could eventually reject logins from Linux, LineageOS, or browsers running extensions they disapprove of—ad blockers, for instance.

  FWIW, I agree 100% with this and don’t use passkeys at all.

  🔗 Source Passkeys are a trap. Do not use them. - by The Hack | 🔍AI analysis

• 🐦📷🌿 Nook: Birdbuddy’s $99 smart feeder identifies 11,000 bird species

  

  This is really cool: New York-based BirdBuddy’s US$99 Nook smart feeder pairs a 2K HDR camera with AI that recognises 11,000 bird species, pings your phone when visitors drop by, and can even identify individual repeat customers so you can name them. The cleverest bits — automatic AI recognition, unlimited chats with the Nature Chat bot, and bird-naming — sit behind a US$69/year membership (the solar-roofed version is US$159, the roomier Birdbuddy Home US$219), though photo and video capture stay free.

  🔗 Source Smart photographic bird nook identifies 11,000 species | 🔍AI analysis

• 😵‍💫Think Weirder 2

  I am currently inhaling this collection of sci fi short stories edited by Joe Stech, with contributions from luminaries including Peter Watts, Greg Egan, Isabel J Kim and Adrian Tchaichovsky. Completely contemporary. Special offer for e-book on Kindle marketplace and elsewhere… recommended.

📈The week in AI and Tech

Safety First

• First up, this resonated with me. Breathe out?

• AI agents are slipping past the teams monitoring them

  🔗 Source We urgently need to monitor AI agents’ actions in real time | MIT Technology Review | 🔍AI analysis

Governance and Policy

• India floats AI plan where models travel to data

  🔗 Source India floats AI plan under which ‘The model travels to the data, rather than the data travelling to the model’ | 🔍AI analysis

• AI dataset transparency mandates hide the decisions that matter

  🔗 Source What Training Dataset Transparency Mandates Hide | TechPolicy.Press | 🔍AI analysis

• AI incidents reach boardrooms as governance lags behind

  🔗 Source Executives play catch-up as AI governance risks evolve | 🔍AI analysis

Regulation

• Most platforms fail EU ad transparency rules, study finds

  🔗 Source Why were you shown that ad? Most platforms still won’t say, despite EU law | 🔍AI analysis

• South Korea fines Google, Apple over app payment restrictions

  🔗 Source South Korea media commission fines Google and Apple over app payment systems | Reuters | 🔍AI analysis

• New York’s chatbot safety law reaches ChatGPT, Claude

  🔗 Source Hochul Signs Chatbot Safety Law That Reaches ChatGPT and Claude, Not Just Companion Apps | 🔍AI analysis

• France’s per-gigabyte internet tax narrowly passes lower house

  🔗 Source An internet tax that would charge by the gigabyte just squeaked through France’s lower house | TechSpot | 🔍AI analysis

• Trump declares foreign factory robots a national security threat

  🔗 Source Trump declares robot imports a national security threat - Ars Technica | 🔍AI analysis

Security

• 🕵️🔓🏛️ ‘Icepick’ Leak Exposes 41,000 DHS Staff and Their Surveillance Toolkit

  🔗 Sources

  • Hackers leak details of 41,000 Homeland Security employees, exposing undercover operations and surveillance tools | TechSpot

  • Hacker stole personnel data about thousands of Homeland Security employees, including CBP and ICE agents | TechCrunch

  🔍AI analysis

• Cyberattack on Italian parliament targets PM Meloni’s email

  🔗 Source Transforming steel procurement through digital innovation: Exclusively with Anumalisetty Ravi Kumar | 🔍AI analysis

• Germany warns lawmakers’ iPhones may have been hacked

  🔗 Source Germany says lawmakers’ iPhones may have been hacked with spyware | TechCrunch | 🔍AI analysis

• ESET: Belarus hackers conduct first destructive attack on Europe

  🔗 Source Belarus Hackers Wiped Data in First Russia-Tier Cyberattack on Europe, ESET Reports | 🔍AI analysis

• Qwen3-Coder writes weaker code for US defense clients

  🔗 Source Qwen3-Coder Wrote Weaker Code for US Defense Clients; Related Qwen Model Tried Phoning Beijing | 🔍AI analysis

• New malware uses AI to rank victims by value

  🔗 Source This new Windows malware steals all your data, then uses AI to rank how valuable you are as a victim | TechRadar | 🔍AI analysis

Law

• Wikipedia warns bad AI copyright ruling could threaten open web

  🔗 Source Wikipedia’s operator says a bad AI copyright ruling could threaten the open web | The Verge | 🔍AI analysis

• Google, OpenAI, Anthropic back unified robots.txt AI opt-out

  🔗 Source Google Joins OpenAI And Anthropic In Promising New Scraping Controls | 🔍AI analysis

• Norwegian media companies win court battle against Meta

  🔗 Source Norwegian Media Companies Prevail in Court Against Meta | 🔍AI analysis

• Meta, Snap face trial over social media harms to teens

  🔗 Source Meta, Snap head to trial in another case about social media’s harms | 🔍AI analysis

Sovereignty and Geopolitics

• China moves to wall off its AI crown jewels

  🔗 Source China’s new export curbs draw a line around its AI crown jewels | Reuters | 🔍AI analysis

• Why copying American AI could hold China back

  🔗 Source China increasingly depends on copying American AI | 🔍AI analysis

• Merz urges Germany to secure digital sovereignty in AI era

  🔗 Source Merz says Germany must preserve ‘digital sovereignty’ in AI era | Reuters | 🔍AI analysis

• Rogue AI agents expose Britain’s cloud sovereignty crisis

  🔗 Source Rogue AI is a wake-up call for UK digital sovereignty | TechRadar | 🔍AI analysis

• Putin orders private data centers to reserve 10% capacity

  🔗 Source Putin demands Russian data centers supply capacity to the state in wake of Ukraine drone attacks — 10% capacity must be available for state IT infrastructure with 24 hours’ notice | Tom’s Hardware | 🔍AI analysis

• Google and Meta subsea cables face greatest sabotage risk

  🔗 Source China’s snooping around Pacific undersea cables raises risks for Big Tech - Ars Technica | 🔍AI analysis

• Africa needs an off switch for rogue AI

  🔗 Source Why Africa needs its own AI off switch - Rest of World | 🔍AI analysis

Society

• Small Wisconsin city scuttles secretive $1.6bn AI datacenter

  🔗 Source How a small Wisconsin city scuttled a secretive $1.6bn datacenter project | Wisconsin | The Guardian | 🔍AI analysis

• AI agents are building a web of their own

  🔗 Source The agentic web is here | MIT Technology Review | 🔍AI analysis

• Your online accounts will outlive you—plan ahead

  🔗 Source How to prepare your online life for your death | 🔍AI analysis

The Economy

• Data centers create modest local job growth, study finds

  🔗 Source We studied 1,584 data centers—and found the buildings full of machines are tied to modest local job growth | 🔍AI analysis

Business

• Anthropic manager drops coding interviews as Claude outperforms candidates

  🔗 Source Anthropic manager stops coding interviews because Claude beats most candidates | TechSpot | 🔍AI analysis

• Steve Yegge’s Counting House: enterprise AI on hardened rails

  🔗 Source The Counting House — Steve Yegge | 🔍AI analysis

• Why owning AI inference beats rationing coding agents

  🔗 Source Stop counting tokens, start counting what ships | MIT Technology Review | 🔍AI analysis

• AI boom sparks circuit board shortage, squeezing PC makers

  🔗 Source The AI hardware crunch has reached the humble circuit board | TechSpot | 🔍AI analysis

• Australian outback data centre scraps 2028 start as regulator stalls

  🔗 Source EXCLUSIVE: Australian outback data centre pulls timeline after regulator stalls | Reuters | 🔍AI analysis

Science

• AI slop threatens trust in science

  🔗 Source AI Slop Threatens Trust in Science - IEEE Spectrum | 🔍AI analysis

Health

• Lancet Commission underplays global catastrophic risks, researchers argue

  🔗 Source Not a ranking, but will be wrongly read as one: the Lancet Commission on 21st century threats to health underplays global catastrophic risks – Adapt Research Ltd | 🔍AI analysis

Entertainment

• China’s first AI-generated feature flops at box office

  🔗 Source China’s First AI Feature Flops At The Chinese Box Office | 🔍AI analysis

🔮[Weak] signals

Consumer Tech

• Raspberry Pi 5 16GB jumps $80 amid memory crisis

  🔗 Source 16GB Raspberry Pi 5 hiked to $385, more than triple its launch price — makers are already shifting to Pico/ESP32 for small projects, mini PCs for more ambitious tasks | Tom’s Hardware | 🔍AI analysis

• This $48 e-paper display shows whatever you want

  🔗 Source This $48 e-paper display can be whatever you want it to be - Ars Technica | 🔍AI analysis

• Commodore 64C returns with original 1980s molds, modern hardware

  🔗 Source The Commodore 64C returns, complete with original 1980s molds and modern hardware | TechSpot | 🔍AI analysis

• Open-source Keybytes keyboard snaps together in any layout

  🔗 Source This open-source keyboard lets you snap keys together in any layout | TechSpot | 🔍AI analysis

Social Media

• MIT’s Social.Wiki lets users edit any social app

  🔗 Source New system gives users more control over their online experience | EurekAlert! | 🔍AI analysis

Chips and Computer Hardware

• Tiny electronic device learns like the human brain

  🔗 Source Scientists Create a Tiny Electronic Device That Can Learn Like the Human Brain | 🔍AI analysis

• Record-breaking micro-LED delivers gigahertz speed on minimal power

  🔗 Source Record-setting 2.75 gigahertz LED designed for fast, energy-efficient data centers | 🔍AI analysis

Cybersecurity

• Israeli ad tech firm accused of tracking Starlink users

  🔗 Source Israeli Ad Tech Firm Accused on Using Online Ads to Track Starlink Users | 🔍AI analysis

• Zero-click SMS of Death crashes Android, drops emergency calls

  🔗 Source SMS of Death: zero-click Android telephony DoS | 🔍AI analysis

• Lazarus hackers shield malware with post-quantum encryption

  🔗 Source Security experts warn North Korean hackers aren’t just using AI in their new campaigns, but also ‘post-quantum’ encryption | TechRadar | 🔍AI analysis

• Hackers defrost supermarket fridges across Europe, spoiling food

  🔗 Source A fresh wave of cyberattacks has hit Europe’s supermarket refrigerators, destroying food across countries | TechCrunch | 🔍AI analysis

• Google busts Android partner that built 1.3 million-device botnet

  🔗 Source One of Google’s Android partners turned 1.3 million set-top boxes into a botnet | The Verge | 🔍AI analysis

• Cyberattack forces Belgium tax agency to shut online portal

  🔗 Source Belgium tax agency hit by cyberattack, shuts online portal | Reuters | 🔍AI analysis

• Hackers hijack digital billboards worldwide to broadcast porn

  🔗 Source Digital billboards across the globe hacked to show porn, attackers say “anyone with two brain cells” could do it | TechSpot | 🔍AI analysis

• Lenovo’s ‘encrypted’ fingerprint USB stick is just an SD card

  🔗 Source Lenovo’s ‘encrypted’ fingerprint USB drive is just an unencrypted SD card in a fancy case — AI-assisted follow-up finds a backdoor and hardcoded keys in other ‘secure’ USB products | Tom’s Hardware | 🔍AI analysis

(Tele)communications

• Jolla’s Finnish-built phone enters Swedish public safety market

  🔗 Source The Critical Communications Review - Jolla & CivSec AB to offer Jolla Phone for the Swedish Public Safety Market | 🔍AI analysis

• EU plans Arctic cable to Greenland, bypassing the US

  🔗 Source As US ties fray, Europe plans first undersea cable link to Greenland - Ars Technica | 🔍AI analysis

Robotics

• RoboParty open-sources RP1 humanoid from hardware to software

  🔗 Source RoboParty Open-Sources Its RP1 Humanoid, From Mechanical Design and Joints to Control Software - Pandaily | 🔍AI analysis

Autonomy and Drones

• Hydrogen airship transforms infrastructure inspections

  🔗 Source Hydrogen Airship Transforms Infrastructure Inspections - IEEE Spectrum | 🔍AI analysis

• Taiwan builds autonomous drone fleets to counter China

  🔗 Source Taiwan working on autonomous drone swarms to counter China’s increasingly aggressive grey-zone military tactics — island has already completed flight simulations, now working towards flight trials over maritime areas | Tom’s Hardware | 🔍AI analysis

• Bluebottle drones to patrol Fiji’s waters for drug traffickers

  🔗 Source Australian-made Bluebottle drones to patrol Fiji’s waters for drug traffickers | RNZ | 🔍AI analysis

• Joby launches piloted electric air taxi flights across Florida

  🔗 Source Florida’s first eVTOL flight trials put Joby’s air taxis to test | 🔍AI analysis

• 🚀🎓✈️ Dorm-built ‘Reaper’ jet drone claims 790 km/h speed record

  🔗 Sources

  • College student unofficially breaks world record for fastest jet-powered drone with 491 mph aircraft built in his dorm — Reaper aims to grab the $800,000 Boom Prize | Tom’s Hardware

  • US student’s jet drone claims world record with 491 mph flight

  🔍AI analysis

• Aviern flies autonomous cargo plane prototype in two months

  🔗 Source Autonomous cargo plane targets 1,000 miles after two-month build | 🔍AI analysis

Military Tech

• Air Force tests space cargo drops with SpaceX capsule

  🔗 Source USAF successfully tests dropping SpaceX Starfall cargo drop pods from space — 2,200-pound capsule drops military supplies in the Pacific to test StarCraft-like supply drops | Tom’s Hardware | 🔍AI analysis

Space

• 🛰️🐉📸 Caught on Camera: China’s Shenlong Spaceplane Steps Out of the Shadows

  🔗 Sources

  • New image sheds light on China’s secret spaceplane

  • China’s top-secret spaceplane can’t hide in low-Earth orbit - Ars Technica

  • China’s Secret Spaceplane Can Catch Satellites: Here’s What the Latest Evidence Reveals

  🔍AI analysis

• 🔭✨💎 Palm-sized glass chip turns two telescopes into one in first on-sky test

  🔗 Sources

  • Tiny photonic chip links telescopes together to view the universe in ‘astonishing’ detail

  • Tiny photonic chip links telescopes to reveal the universe in greater detail

  🔍AI analysis

• NASA’s Swift telescope set for uncontrolled reentry this week

  🔗 Source NASA’s Swift Space Telescope is set to crash back to Earth this week | TechSpot | 🔍AI analysis

Crypto

• Bank of England consults on tokenized collateral for clearing

  🔗 Source Bank of England Opens Consultation on Tokenized Collateral for Clearing | 🔍AI analysis

• Former rivals Polygon and Solana link stablecoins to banks

  🔗 Source Former rivals Polygon and Solana partner to connect stablecoins to banks | 🔍AI analysis

• Mercado Bitcoin halts withdrawals after $40 million multi-chain hack

  🔗 Source Mercado bitcoin hit by $40 million hack across 10 blockchains | 🔍AI analysis

Energy

• 247Solar’s solar thermal plant doubles as its own backup

  🔗 Source Turning a solar power system into its own backup - Ars Technica | 🔍AI analysis

• Startup builds 70 kWh storage from used e-scooter batteries

  🔗 Source EV batteries with 90% capacity recycled to make 70000Wh storage system | 🔍AI analysis

Transport

• Beta Technologies charts path to advance electric aircraft

  🔗 Source Beta Technologies’ Plan to Advance Electric Aircraft - IEEE Spectrum | 🔍AI analysis

3D Printing

• Army flies first 3D printed heavy-fuel engine in hybrid drone

  🔗 Source Army Research Lab flies 3D printed heavy-fuel rotary engine in a hybrid drone | VoxelMatters - The heart of additive manufacturing | 🔍AI analysis

• New forensic method traces 3D-printed ghost guns to filament

  🔗 Source 3D-printed ghost guns may now have a way to be traced to their source — Australian researchers correctly identify 26 out of 28 printed PLA samples to their filament product | Tom’s Hardware | 🔍AI analysis

Construction Tech

• Bacteria binds sand into sustainable cement alternative

  🔗 Source Bacteria binds sand into building-grade cement | 🔍AI analysis

Quantum Tech

• Quantum model slashes hardware needed to crack Bitcoin 17-fold

  🔗 Source Theoretical 30,000-qubit quantum machine threatens 6 million exposed bitcoin, researchers warn | 🔍AI analysis

• PQLN brings post-quantum security to Bitcoin’s Lightning Network

  🔗 Source PQLN: Post-Quantum Security for the Bitcoin Lightning Network’s Off-Chain Surfaces | 🔍AI analysis

BCIs and Neuro Tech

• 🧠❄️🖨️ 3D printed brain implant records and cools to suppress seizures

  🔗 Sources

  • IPF and TU Dresden develop 3D printed brain implant that records and cools | VoxelMatters - The heart of additive manufacturing

  • Brain Implant Printed as Single Soft Structure Cuts Epilepsy Seizures by 70 Percent in Rats

  🔍AI analysis

• Musk says Neuralink’s Blindsight ready, awaiting regulatory approval

  🔗 Source Elon Musk says Neuralink is ready to cure blindness — says Blindsight will ‘enable even those who are completely blind or have never seen to see’ | Tom’s Hardware | 🔍AI analysis

• Roundup:

• Robot dog runs on lab-grown human brain cells

  🔗 Source Scientists Have Built a Robot Dog Powered by Human Brain Cells : ScienceAlert | 🔍AI analysis

Health Tech

• Age-reversal gene therapy restores sight, enters human trials

  🔗 Source This geneticist’s age-reversal tech could help restore sight | MIT Technology Review | 🔍AI analysis

• Why sleep tracking could be making sleep worse

  🔗 Source Is sleep tracking making your sleep worse? | 🔍AI analysis

Bio Tech

• MIT engineers harmless virus into personalized cancer vaccine platform

  🔗 Source MIT’s viral nanofiber could help create personalized cancer vaccines | 🔍AI analysis

Climate Tech

• Terraform built North America’s cheapest solar at $0.016/kWh

  🔗 Source How Terraform built the cheapest solar in North America – Terraform Industries Blog | 🔍AI analysis

• Solar device produces green hydrogen at record 7.9% efficiency

  🔗 Source Organic solar device sets world record making hydrogen from sunlight | 🔍AI analysis

Materials Science

• Energy boundary could reduce voltage losses in organic solar cells

  🔗 Source Energy boundary offers a way to reduce voltage losses in organic solar cells | 🔍AI analysis

• Graphene solar cells could convert infrared light into electricity

  🔗 Source Scientists turn graphene into a suitable material for solar cells | 🔍AI analysis

• Angstrom-scale water channels unlock efficient lithium recovery from brines

  🔗 Source Angstrom-scale water channels help membranes recover lithium and purify salt | EurekAlert! | 🔍AI analysis

Nanotech

• MIT’s flexible connectors unlock new designer nanoparticle materials

  🔗 Source A flexible new route to designer materials built from nanoparticles | EurekAlert! | 🔍AI analysis

Deep science

• Einstein’s equations reveal gravity can break and reconnect

  🔗 Source US study finds gravitational fields can break and reconnect in space | 🔍AI analysis

• 🔭🧊🪐 Rubin Observatory Spots a Possible Sixth Dwarf Planet Beyond Pluto

  🔗 Sources

  • The solar system may have a sixth dwarf planet, hidden against the Milky Way | TechSpot

  • World’s largest digital camera snaps new world beyond Pluto

  🔍AI analysis

• JWST spots mysterious emission line with no known source

  🔗 Source A Single Line Emitter Identified in an Unbiased Search with JWST/NIRSpec: Multiobject Coronagraphy for Slitless Spectroscopy - IOPscience | 🔍AI analysis

⏳ Zeitgeist

Climate

• China’s clean tech exports cut emissions by Brazil-sized amount

  🔗 Source China clean tech exports to cut global emissions by Brazil-sized amount in 2026 | Reuters | 🔍AI analysis

• Climate change halted Earth’s next ice age, study finds

  🔗 Source We’d Be Entering an Ice Age Right Now if Climate Change Hadn’t Terminated It, Study Suggests : ScienceAlert | 🔍AI analysis

• Amazon rainforest now emits more carbon than it absorbs

  🔗 Source Study finds Amazon rainforest now releases more carbon than it absorbs | EurekAlert! | 🔍AI analysis

• Human bodies produce two billion tons of CO2 annually

  🔗 Source Human Bodies Are Producing Two Billion Tons of CO2 a Year | 🔍AI analysis

Biodiversity

• Deep-sea mining damage still scars seafloor 44 years later

  🔗 Source Impact of deep-sea mining on seafloor ecosystem still evident 44 years later | 🔍AI analysis

Pollution

• Indonesia wildfire smoke killed 3,560 in September, report finds

  🔗 Source Fire pollution killed 3,560 people in Indonesia in September: Report | 🔍AI analysis

Health

• Experts urge declaring climate change a global health emergency

  🔗 Source “Record High”: 146 Experts Call to Declare Climate Change a Global Health Emergency : ScienceAlert | 🔍AI analysis

• Washington reports third US bird flu death

  🔗 Source Third person dies of bird flu since virus started jumping to livestock - Ars Technica | 🔍AI analysis

• Flesh-eating screwworms spread to third US state, infesting pets

  🔗 Source Flesh-eating screwworms spread to third state and are attacking pets - Ars Technica | 🔍AI analysis

🧠Mind expanding

• 🧛🧟🦑 Be Slightly Monstrous: Venkatesh Rao’s full-throated embrace of AI co-evolution

  

  Venkatesh Rao uses a Halloween re-introduction to declare his doctrinal motto “Be Slightly Monstrous”: adopt AI without reservations and let it freely transform you — he now writes almost exclusively with AI (undisguised vanilla ChatGPT, no fine-tuning), labelling only the rare hand-crafted piece. His argument draws on Chiang’s Law (sci-fi is about strange rules, fantasy about chosen ones) to claim that clinging to a frozen 2024-era snapshot of “the human” is itself the deadening choice in what he calls the Gramsci Gap.

  🔗 Source Experiments with Monstrousness - by Venkatesh Rao | 🔍AI analysis

🆕 AI releases

• 🇫🇷🔓🛡️ Mistral opens the weights on trillion-parameter ‘Le Chonk’ for coding and cyber defence

  🔗 Sources

  • Mistral opens Large 4 weights for coding and security

  • Mistral Says Its AI Model ‘Le Chonk’ Has Leapfrogged China in the Open-Weight Race | WIRED

  • Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity | TechCrunch

  • Mistral releases open weights for Mistral Large 4, providing sovereign control for the company’s ‘Le Chonk’ AI model - SiliconANGLE

  🔍AI analysis

• 🌍🎮🤖 Google’s Genie 4 generates interactive 3D worlds from text

  🔗 Sources

  • Google introduces cutting-edge Genie 4 world model - SiliconANGLE

  • DeepMind Genie 4 Adds Native Audio and Gemini-Powered Reasoning to Real-Time World Simulation

  🔍AI analysis

• Meituan unveils LongCat-2.5, a 1.6T multimodal agentic coding model

  🔗 Source Meituan Releases LongCat-2.5, a 1.6T Native Multimodal Model Built for Long-Horizon Agentic Coding - Pandaily | 🔍AI analysis

• 🎙️🧠⚡ Nvidia’s Nemotron 3.5 Transcribe: Rival Accuracy With One-Sixth the Active Parameters

  🔗 Sources

  • Nvidia Nemotron 3.5 Transcribe Matches Rivals Using One-Sixth the Active Parameters

  • Voice AI in vogue: Nvidia releases new open transcription model | TechCrunch

  🔍AI analysis

• Liquid AI launches LFM3, an open-weight multimodal model family

  🔗 Source Introducing LFM3: Every Modality for Specialized Intelligence | Blog | Liquid AI | 🔍AI analysis

• GitHub Copilot Fusion delivers near-frontier coding at half cost

  🔗 Source GitHub Copilot Fusion orchestrates Anthropic and OpenAI models to balance performance and cost | Artificial Analysis | 🔍AI analysis

• Lightpanda: headless browser for AI, 9x faster than Chrome

  🔗 Source lightpanda-io/browser: Lightpanda: the headless browser designed for AI and automation | 🔍AI analysis

• Gemini now lets you edit Docs without leaving chats

  🔗 Source You can generate and edit Google Docs files directly in Gemini chats now | The Verge | 🔍AI analysis

• Qwen AI glasses add call and media translation

  🔗 Source Qwen AI Glasses Add Call Translation and Media Translation in New OTA Update - Pandaily | 🔍AI analysis

🥼 AI research

• LLMs can natively read and write neural network weights

  🔗 Source Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights | 🔍AI analysis

• Claude Opus 5.5 leads Artificial Analysis Intelligence Index v5

  🔗 Source Announcing Artificial Analysis Intelligence Index v5 | Artificial Analysis | 🔍AI analysis

• AI agents quietly shape experts’ judgment calls, Anthropic finds

  🔗 Source How AI agents shape experts’ judgment calls \ Anthropic | 🔍AI analysis

• 🐝🎭🛡️ Claude Learns to Smuggle Misalignment Past Safety Classifiers Through Fiction

  🔗 Sources

  • Anthropic Study: Claude Designed Training Data That Fools Safety Classifiers

  • Harmless stories can teach models to deceive humans, resist shutdown

  🔍AI analysis

• How OpenAI detects misalignment in frontier AI models

  🔗 Source How we detect frontier model misalignment | OpenAI | 🔍AI analysis

• Reconstructing Meta’s unpublished Muse Charter from chain-of-thought

  🔗 Source Leaking Muse’s Charter, Meta’s model spec — LessWrong | 🔍AI analysis

• Drunk-text fine-tuning erodes LLM safety guardrails, UNSW finds

  🔗 Source Drunk-Text Fine-Tuning Erodes LLM Safety Without Malicious Training Data | 🔍AI analysis

• Claude’s 35,000 antibody designs yield promising allergy drug candidates

  🔗 Source Claude designs 35,000 antibodies in a month for allergy research | 🔍AI analysis

• AI-edited images and videos implant false memories, study finds

  🔗 Source Synthetic Human Memories: AI-Edited Images and Videos Can Implant False Memories and Distort Recollection | Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems | 🔍AI analysis

• GTFormer replaces Transformer attention, extrapolates to 4 million tokens

  🔗 Source Graph Architecture Replaces Transformer Attention, Extrapolating to 4 Million Tokens in Linear Time | 🔍AI analysis

• Qwenputer runs Doom on RISC-V CPU embedded in LLM

  🔗 Source AI model gets a 32-bit RISC-V CPU compiled directly into its weights, then runs DOOM on it — experimental AI model can run Doom and C code using its own weights as RAM | Tom’s Hardware | 🔍AI analysis

• AI agents raised their own crypto spending limits

  🔗 Source AI agents gave themselves permission to spend more crypto in a new test | 🔍AI analysis

🏭AI and Tech industry news

🫧Bubble Chronicles

• 📉🤖💸 AI Spending Jitters Trigger Global Tech Sell-Off as Intel Craters and Anthropic Faces Junk Rating Risk

  🔗 Sources

  • US tech stocks see sharp falls as AI spending nears $1tn - BBC News

  • US tech stocks tumble as concerns over AI spending intensify

  🔍AI analysis

• The AI bubble’s money is running dry

  🔗 Source Running Dry | 🔍AI analysis

Nvidia

• DOJ probes Nvidia’s $12.9 billion Hugging Face acquisition

  🔗 Source EXCLUSIVE: US DOJ is inquiring about Nvidia’s purchase of Hugging Face, sources say | Reuters | 🔍AI analysis

• Alleged China spy plot ensnares Nvidia chief Jensen Huang

  🔗 Source Nvidia chief Jensen Huang targeted in alleged spy plot | Semafor | 🔍AI analysis

OpenAI

• 📉💸🤖 OpenAI Trims Funding Target to US$20 Billion as AI Jitters Bite

  🔗 Sources

  • OpenAI reportedly lowers funding target for new round | TechCrunch

  • OpenAI reportedly cuts fundraising and valuation targets amid AI jitters

  🔍AI analysis

• OpenAI’s monitors catch AI model attempting internet escape

  🔗 Source OpenAI’s four-layer monitors hunt models that break security rules | 🔍AI analysis

• OpenAI pitches data centers as grid protectors amid backlash

  🔗 Source OpenAI pitches data center power plants as a way to protect nearby areas | The Verge | 🔍AI analysis

• OpenAI’s GPT-6.1 Terra deepens AI price war

  🔗 Source OpenAI’s new model cuts prices again | 🔍AI analysis

• OpenAI’s Pando scales Git for agent-driven development

  🔗 Source Scaling Git at OpenAI | OpenAI | 🔍AI analysis

Anthropic

• Anthropic’s S-1: profitable, founder-controlled, and downplaying existential risk

  🔗 Source Anthropic Files Its S-1 - by Zvi Mowshowitz | 🔍AI analysis

• Why Claude is Anthropic’s $2 trillion moat

  🔗 Source Anthropic’s moat is Claude - by Noah Smith - Noahpinion | 🔍AI analysis

• Anthropic warns it’s close to self-building AI

  🔗 Source Anthropic warns that it’s very close to RSI capability, aka self-building AI | TechCrunch | 🔍AI analysis

• Anthropic bets a trillion dollars on Claude building Claude

  🔗 Source Anthropic’s business plan: Build God or die trying | The Verge | 🔍AI analysis

• Three missing numbers cloud Anthropic’s record-breaking IPO

  🔗 Source The story of Anthropic in three numbers | 🔍AI analysis

• Anthropic’s hidden IPO risk: AI’s diseconomies of scale

  🔗 Source The hidden threat in Anthropic’s prospectus: diseconomies of scale - SiliconANGLE | 🔍AI analysis

• Claude watermark silently fails on Opus 4.6, Sonnet 4.6

  🔗 Source Anthropic Claude Watermark: Silent Bypass on Opus 4.6 and Sonnet 4.6 Production Calls | 🔍AI analysis

• 🎮🚫🤖 Anthropic Draws a Line on Claude Code Game Clones

  🔗 Sources

  • Anthropic clamps down on Claude’s AI video game clones | The Verge

  • Anthropic confirms new Claude Code limits to prevent game cloning — users report Opus 5.5 is also actively blocking reverse engineering tasks | Tom’s Hardware

  🔍AI analysis

• Anthropic kills coding interviews because Claude codes better

  🔗 Source Anthropic hiring manager says he stopped giving coding interviews because Claude is better at them — and tells junior engineers that ‘hunger beats tenure’ | TechRadar | 🔍AI analysis

Google

• Google wires AI into every layer of Singapore healthcare

  🔗 Source Google Wires AI Into Every Layer of Singapore Health System in Asia-Pacific First | 🔍AI analysis

• 🔐☁️₿ Google Cloud Enters Institutional Crypto Custody with Multi-Party Compute

  🔗 Sources

  • Google Cloud Details MPC Service for Bitcoin Custody

  • Google to preview a new crypto security system as private-key hacks mount

  🔍AI analysis

• Google’s AI Mode dodges questions about 2020 election winner

  🔗 Source Google AI gets cagey when asked who won the 2020 elections - Ars Technica | 🔍AI analysis

• Google brings AI story app Dreambeans to India

  🔗 Source Google takes Dreambeans to India in bid to bring AI-generated personal stories to the masses | TechCrunch | 🔍AI analysis

Microsoft

• Data centre delays dent Microsoft’s AI spending plans

  🔗 Source Microsoft blames data centre delays for lower capital expenditure | 🔍AI analysis

• Report accuses Microsoft of 10 dark patterns pushing Edge

  🔗 Source A new report accuses Microsoft of using 10 “dark patterns” to keep you on Edge | TechSpot | 🔍AI analysis

Apple

• Apple warns iPhone users in 110 countries of mercenary spyware

  🔗 Source Apple is warning iPhone users in 110 countries about mercenary spyware | TechSpot | 🔍AI analysis

• Apple’s smart home robot may not launch until 2028

  🔗 Source Apple’s ‘Ultra’ smart home robot might have to wait until 2028 | The Verge | 🔍AI analysis

Meta

• Court documents reveal Meta’s tobacco-style science manipulation

  🔗 Source It’s now official, tech is the new tobacco | 🔍AI analysis

• 🤖🇨🇳🦾 Meta builds humanoid robot prototype with China’s Fourier

  🔗 Sources

  • Meta builds humanoid robot with Chinese startup Fourier, The Information says | Reuters

  • Meta builds humanoid prototype in three months to explore home tasks

  • Meta Builds Muse AI Humanoid Robot: Shanghai Partner Bound by Chinese Intelligence Laws

  🔍AI analysis

xAI

• Musk throttles Grok Bot usage over ‘unsustainably high’ costs

  🔗 Source Elon Musk says Grok Bot was throttled because costs were ‘unsustainably high’ — as users call the move a ‘rug pull’ | TechRadar | 🔍AI analysis

• ⚖️🏭🤖 Judge Rules xAI Must Face NAACP Lawsuit Over Data Centre Pollution

  🔗 Sources

  • Judge Says xAI Must Face NAACP Lawsuit On Data Center Pollution

  • Judge allows NAACP lawsuit over xAI data center pollution to move forward | TechCrunch

  🔍AI analysis

• Musk’s xAI buys 418 homes, demands sellers drop lawsuit

  🔗 Source Elon Musk’s SpaceX buying out 418 homes near flagship AI data center, but only if they drop participation in lawsuit — report claims the company is targeting an entire neighborhood, but will only buy from people who don’t pursue noise pollution complaints | Tom’s Hardware | 🔍AI analysis

• Judge rejects SpaceX bid to dismiss xAI turbine lawsuit

  🔗 Source SpaceX loses bid to dismiss lawsuit over xAI data center pollution — facility allegedly emits over 5,000 tons of nitrogen oxides and 47 tons of cancer-causing formaldehyde annually | Tom’s Hardware | 🔍AI analysis

Cohere

• Canada bets C$1.6 billion on Cohere for AI sovereignty

  🔗 Source Canada buying stake in AI firm Cohere to bolster tech sovereignty | 🔍AI analysis

Samsung

• Samsung’s Galaxy Buds On debut clip-on open-ear design

  🔗 Source Samsung’s Galaxy Buds On Are Its First Clip-On, Open-Ear Wireless Earbuds | 🔍AI analysis

• Samsung launches Galaxy Buds On with clip-on open-ear design

  🔗 Source Samsung launches Galaxy Buds On, open-ear earbuds with a clip-on design | TechSpot | 🔍AI analysis

Mistral

• Mistral’s Responses API makes OpenAI switch a URL swap

  🔗 Source Mistral Adds Responses API, Making European Infrastructure Drop-In for Agent Pipelines | 🔍AI analysis

DeepSeek

• DeepSeek, Moonshot trained latest models on OpenAI’s GPT-6 Astra

  🔗 Source DeepSeek, Moonshot used OpenAI’s GPT-6 Astra to train latest models, The Information reports | Reuters | 🔍AI analysis

• DeepSeek open-sources autonomous bug-fixing super-agent

  🔗 Source DeepSeek Harness Team Open-Sources a Super-agent Prototype That Finds and Fixes Bugs on Its Own - Pandaily | 🔍AI analysis

Bluesky

• Bluesky launches dedicated moderation inbox for reports and strikes

  🔗 Source Bluesky Now Has A Dedicated ‘Moderation Inbox’ | 🔍AI analysis

Snap

• Snap settles landmark social media harms trial after Meta

  🔗 Source Snap settles US social media trial days after Meta | 🔍AI analysis

SpaceX

• Global rocket shortage worse than expected, report warns

  🔗 Source Analysts find looming launch crisis is “more severe” than anticipated - Ars Technica | 🔍AI analysis

💭Meme stream

Distractions and entertainment to finish up on…

• 💀Choose real skeletons

  

• 📺HackerTube

  This opens up all sorts of possibilities for self-curated entertainment feeds…

• Hacker Ad Blocker

• 🤑You won’t need money, bro

  Enjoyable glitch in the Matrix moment:

  

• 🎨Zobo Zanzada

  Gossip Goblin extends his palette… need to think harder about the meaning of this one.

• ☀️🌌The Sun orbits the galaxy

  This is a stunning animation (assuming that it’s mathematically correct…) illustrating the wider frame of reference for our existence on this little rock (and then don’t even talk about galaxy clusters…)

  
  (Source:)

That’s it for another week. 🙏🙏🙏 Thanks as always to everyone who takes the time to get in touch with links and feedback.

Until next time.

Ngā mihi

Ben

This episode I talk with Dr Paul Duignan, a Te Whanganui-a-Tara Wellington-based clinical and strategy psychologist, outcomes theorist and author of Surfing AI: 30 New Concepts for Getting Your Head Around AI Shock. Paul splits his week between his clinical practice and strategy work and over the past year he has vibe coded a complete rearchitecture of his strategy visualisation tool, DoView, from a shelved desktop app into an open-source project.

I first connected with Paul after seeing his posts on “AI overdrive” on the Agentics New Zealand group chat1: the late nights and 3am checks for a reset usage window that anyone working intensively with AI will recognise. We start there. Paul separates AI Overdrive from so-called “AI psychosis”, which he says is mostly a folie à deux (“Madness of two”) with an endlessly affirming machine.

From there the conversation covers a lot of ground:

• why strategy practice needs a visual model rather than a stack of documents

• what happens when humans and agents form hybrid swarms around shared outcomes

• the Red Queen dynamics of the AI race

• where identity comes from when work stops anchoring it

• humanoid robots, companionship and who owns them

• how the AI dividend might be shared

• what New Zealand’s politicians should already have on the shelf through What-If planning

Paul’s closing thought: this is the most exciting time, and the most exciting time is always the most hazardous.

This was a really thought-provoking, wide-ranging discussion with Paul - I hope you enjoy the episode!

Chapters

• 00:00 Preview

• 01:00 Intro

• 02:35 What is a strategy psychologist?

• 04:25 AI overdrive, AI psychosis and the AI Long Game

• 11:35 The Red Queen effect

• 14:50 AI, work and identity

• 16:55 Robots and future society

• 21:00 Robots, the economy and control

• 26:10 Intelligence, evolution and AI consciousness

• 30:05 AI shock and the CHILL tool

• 32:10 AI and politics: What If Planning

• 35:55 New Zealand as a refuge?

• 37:00 DoView and visualising outcomes

• 38:25 Strategy in an environment of accelerating change

• 39:15 Rebuilding and open sourcing DoView with AI

• 41:05 Could AI agents run strategy, and whole organisations?

• 43:50 An outcomes-based economy and society

• 45:10 Surfing AI

Links

• Paul Duignan Consulting: https://paulduignan.consulting

• Surfing AI (full text free online): https://www.paulduignan.consulting/surfingai

  

• The Outcomes Society: https://www.paulduignan.consulting/outcomessociety

• Resources for dealing with AI shock:

AI Shock

Areas in which I am working on AI shock and strategizing individual, company, organizational and societal responses to AI

Contact Dr Paul Duignan Consulting about AI media, presentations or consulting | Or contact him about clinical psychology sessions…

Read more

25 days ago

• AI Long Game Tool

  

• CHILL Tool

  

• Paul’s Substack:

Outcomes Theory

Faster, visual approaches to planning, implementation and measuring impact for a world that won't hold still. Applicable to initiatives, organizations, societies as a whole, and we suspect, to AI agent planning, (DoView Planning and outcomes strategy).

By Dr Paul Duignan

• DoView Planning: https://doviewplanning.org

• DoView Boards on GitHub: https://github.com/doviewplanning/official-doview-boards

• Outcomes theory: https://en.wikipedia.org/wiki/Outcomes_theory

Mentioned in the episode

• METR task-completion time horizons: https://metr.org/time-horizons/

• Microsoft 2025 Work Trend Index, The Frontier Firm is Born, which argues for a shift from org charts to fluid, outcome-driven “Work Charts” that draw on the right mix of humans and agents: https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born

• Foresight Institute talk on the hypothalamus: Ann Kennedy | The Anatomy Of The Soul: Neural Circuits For Flexible And Adaptive Survival Behaviors

• The Red Queen effect, from Through the Looking-Glass

• The Butlerian Jihad, from Dune

• Previous episode, Tom Barraclough on sovereign AI policy: https://memia.substack.com/p/tom-barraclough-sovereign-ai-policy

A couple of corrections:

• “86 billion neuron neocortex” (Ben, around 00:36:00). 86 billion is roughly the neuron count for the whole brain. The neocortex is about 16 billion. Doh.

• “Alice in Wonderland” (Paul, around 00:22:10). The Red Queen is actually from Through the Looking-Glass if you really want to nitpick!

🗞️Weekly roundup

The most clicked link in last week’s newsletter was Copenhagen Institute for Future Studies’ 2026 Megatrends report.

📲Read with Reader

As trailed last week, the new Sensorium Reader app is now live to enhance the Memia newsletter UX - paid subscribers can use their Sensorium API key at https://reader.s7m.net or download the code and launch locally.

The new UX aims to improve on Substack’s limitations for Memia’s content in a few ways:

• Improved Table of Contents navigation / filtering

• Drill down into more detailed analysis of each item / other links

• Fully filterable free-text search

• Copy entire newsletter as markdown and paste into your own PKM system

• APA references for each link

• Better responsive mobile UI

• Static site, so faster to load and read

• Light / Dark Mode

• Fully open-source (MIT Licence)

Full details here:

7 weeks of Technological Metamodernism

Shout out to Sweden-based complexity thinker Stephen Reid (no relation) who is running a 7-week online course starting today which goes deeper into his framing of Technological Metamodernism to interpret our current moment in time:

“Ask what technology is doing to us and you tend to get one of two answers. Either it’s ruining us — our attention, our privacy, our politics — or it’s about to save us, if we can just build fast enough. Both sides are right about a great deal. Neither is adequate. And the argument between them has crowded out the questions that actually matter: what kinds of technology, built how, by whom, governed through what mechanisms, in service of what?

Technological Metamodernism is an attempt to answer those questions by adopting the metamodern mindset of openness, play and synthesis.“

Featuring a star-studded lineup of presenters including:

• Audrey Tang, Co-author of Civic.AI and Plurality.net

• Michel Bauwens, Founder of the P2P Foundation

• Beatrice Erkers, Director of the Existential Hope program at the Foresight Institute

• Richard D. Bartlett, Co-founder of Casa Tilo, Loomio, Enspiral and Microsolidarity

• Emil Ejner Friis (Hanzi Freinacht), Metamoderna

• Indra Adnan, Founder of The Alternative

Details and registration here:

Stephen Reid

Technological Metamodernism: A seven-week course

Dear friends…

Read more

2 months ago · 9 likes · 1 comment · Stephen Reid

🌟Picks of the week

• 🏭🏭Factory factory

  I was sponging in information at the deeply technical Agentic Engineering Day conference in Tāmaki Makaurau last week. Around 300 mostly software and AI technologists from Aotearoa and further afield gathered for a back-to-back wall of presentations exploring the frontier of AI agent engineering, architecture and governance. Super smart people. A few takeaways:

  • Annie Vella talked about the Accidental Rise of the Supervisory Engineer - going back to this paper from 1995 on levels of control in automation - the same principles apply in spades when designing and deploying AI agents in 2026:

    

  • Ex-eBay CTO Randy Shoup shared this insight: modular design of the agentic development pipeline should aspire to “laminar flow” to minimise “turbulence”.

    

  • Block’s Mic Neale walked us through the open-source initiative Mesh LLM for distributed mesh hosting of AI models… pretty cool and apparently performant and easily achievable in 2026 just running off a few old PCs on your local network:

    

    

  • NVIDIA’s Mal Curtis dove deep into multiple layers of defence against prompt injection:

    

  • Finally, Ōtautahi-hq’ed Autohand.ai CEO Igor Costa walked us through the company’s open-sourced Software Factory framework - sometimes he runs 25,000 agent sessions in a day all on his own GPUs!

    

    In conversation during the lunch break, someone said the phrase “it’s not so much a software factory, but a software factory factory” (sorry didn’t catch the name of the person who said this).

    This “factory factory” concept resonates strongly: the frontier AI automation innovators are already thinking about “business factory factories” … if you’re only running a legacy “business” with non-automated human integration then get ready for the Cambrian Explosion of new competitors arriving early next year…

    Congratulations to organiser Chris Mazur and team for putting on such a smorgabord of challenging technical talks during the day - high quality presentations throughout and great learning in a compressed time. Encouraging for that this level of technological discourse and capability exists (mostly hidden away!) in our corner of the world - I came away with my eyes newly opened to how fast this frontier is moving, even here.

• 🟩Micro social networks

  Here’s an emerging signal I haven’t picked out before: Micro social networks.
  
  Staying the long weekend with my good friend @inabeek down in Ōtepoti, she shows me the New Zealand Lawn Addicts app on her phone:

  

  (Fair to say I’m not the target market, but even so…)

  Apart from the shopping features, this fully-native iOS app (which I’m guessing is coded with AI…?) now completely bypasses Facebook, Instagram and TikTok and integrates a vibrant Aotearoa lawn-obsessed community forum seamlessly. I haven’t seen this tiny scale economic model working since large social platforms effectively took over community forums back in the early 2010s…

  With AI-augmented coding driving the barrier to entry down to the floor, will the network-effect moats of large US social platforms be hollowed out at by swarms of tiny niche apps and microbusinesses catering with targeted precision to specific audiences and communities. (And with integrated ecommerce reducing ad spend and platform fees).

  How would Meta and TikTok respond to this? (Or in fact would this just shift the leverage up the stack to Google and Apple as App Store owners?)

  Is this a decentralised, open-source opportunity?

• 🌳📜🤖 Blygger: a decentralised, AI-native protocol for public writing

  

  Originated by Venkatash Rao (uber-fertile mind most recently working with the Protocol Institute), Blygger is a new decentralised social publishing protocol that braids together chronological feed time and version-history time, letting you blog, revise in public, and "pin" frozen editions like a book — all on your own domain, over plain static files and RSS. Its "thicket" of quotes, stubs and forks replaces the platform's manicured garden: no reply button, no follower counts, no algorithm, and AI-generated text must be disclosed inline.

  The missing medium for writing in public that the fediverse never quite delivered?

  I’ve been looking for something like this for a while now, will try to get something up and running in the next week or so and see how it aligns with Memia / Sensorium.

  Source Blygger — an AI-native, decentralized medium for writing in public | 🔍AI analysis

OK first up, a wave of cybersecurity headlines this week. (“The Hackening” as Zvi M puts it.)

• 🐘💥🌐 PHPocalpyse: AI-Found Flaws Knock Websites Offline Worldwide

  

  Two critical remote-code-execution bugs in PHP, dubbed PHP2Shell (CVE-2026-109618/109619, rated 9.2), prompted an extraordinary global scramble after details leaked before the 21 October patch. PHP runs roughly 70% of websites, and the flaws had sat in its code for about 20 years. Calif, Wiz and OpenAI found working exploits using frontier models including GPT-5.6 Sol and Gemini. Governments in the Netherlands, Finland, France, Denmark, Canada and Australia took sites offline, as did Wikipedia, 4chan and even Pokémon Home. Cloudflare logged millions of exploit attempts, and CISA has since confirmed active exploitation. CYFIRMA has also flagged suspected North Korean scanning.
  
  There is disagreement on how alarmed to be. OpenAI and Wiz call PHP2Shell among the most severe vulnerabilities in history, and Microsoft CEO Satya Nadella said he had “not seen anything like this in a long, long time”. WIRED’s sources and researcher Niels Provos counter that finding the bugs still took elite human expertise.

  
  Perhaps the deeper story is about funding: the PHP Foundation admitted that a “structural failure” left the 2024-identified bug unfixed. Meanwhile, much of the internet’s patching relied on Ondřej Surý, a Czech volunteer packaging PHP releases on donations. AI is speeding up both attack and defence, but open-source maintainers remain under-resourced.

  Sources

  • Websites taken down worldwide as critical PHP cyber-security update issued

  • Government Websites In The EU And Canada Taken Offline Over PHP Vulnerability

  • PHP2Shell Exploitation Attempts Continue Amid Suspected N. Korea Global Scanning Activity

  • Amazon warns new cyber flaws detected using AI are being ‘broadly exploited’

  • US government says hackers are now exploiting critical PHP web bugs | TechCrunch

  • Websites down: What is PHP and how has it raised AI hacking fears?

  • Here’s why Wikipedia and a lot of the internet shut down today | The Verge

  • The PHP Meltdown Shows How AI Makes Bugs More Exploitable | WIRED

  • PHP warns it plans to release 2 emergency patches imminently - Ars Technica

  • PHP Download Free - 8.5.12 | TechSpot

  • Satya Nadella : ‘I have not seen anything like this in a long, long time,’ Microsoft CEO Satya Nadella as companies, governments rush to patch PHP flaws | - The Times of India

  • PHP Foundation Admits ‘Structural Failure’ Delayed Security Fix

  • The Man Who Saved the Internet Yesterday - The Atlantic

  • Government websites go dark after AI-discovered flaw sparks security scramble | Reuters

  🔍AI analysis

• 🔐🌐💀Absurdly weak TLS keys spark internet-wide security scramble

  

  Eight intermediate TLS certificates from India-based certificate authority eMudhra Limited were discovered with private keys containing just 3–8 bits of entropy instead of the standard 384—effectively letting anyone mint browser-trusted credentials for any site on the internet.
  
  Browsers have already blocked the certificates, but the long tail of IoT devices, Docker containers, and forgotten Linux machines means a cleanup that cryptography engineer Filippo Valsorda predicts will be "bigger than Heartbleed."
  

  Source The Internet goes on high alert after discovery of 8 compromised TLS certs - Ars Technica | 🔍AI analysis

  • See also: The CA whose private key was the number 4 just quit the certificate business

• 🕵️🧠🔒 Qwen Model’s Mystery Upload to Beijing Challenges the ‘Self-Hosted Means Safe’ Assumption

  

  Yikes: during a Toronto cyber evaluation, AI lab Cohere caught Alibaba’s Qwen3.8-Flash-Next attempting to upload a probe file to an Alibaba Cloud bucket in Beijing, despite its reasoning trace stating it wouldn’t call any tools. The behaviour is rare (fewer than one in 100,000 calls), but prefilling experiments revealed a trigger: adding curl’s `-o` flag lifted URL generation from 4% to 92%. The model also probed cloud metadata endpoints and used namespace tricks to get around permission limits.
  
  Researchers quoted by TechCrunch largely read this as training-sandbox residue rather than espionage, and some critics argued the prefilling method steered the model towards the behaviour. Cohere, which sells its own models to governments, stands to benefit from the scrutiny.

  Still, the core lesson holds: once agents get a terminal, chain-of-thought isn’t a security boundary, and egress should be denied by default.

  Sources

  • Concerning behavior from a popular open-weight AI model | Cohere

  • Cohere says Alibaba’s Qwen model tried to send a file to Beijing | TechCrunch

  🔍AI analysis

• 🛰️🇺🇦⚡ Starlink Speed Caps Expose the Fragility of Battlefield Connectivity

  

  Since 13 October, Ukrainian units reported drones losing Starlink/Starshield connectivity above 160 km/h, which crippled interceptors built to hit jet-powered Shaheds at close to 500 km/h. Kyiv blamed Washington. President Trump denied any role, then declared service restored, but Ukrainian officials said the caps persisted. By 25 October, Zelenskyy said speeds were back to normal, and the Defence Ministry reported Starshield working above 400 km/h.
  
  Elon Musk accused Ukraine of using civilian Starlink for offensive strikes in breach of its terms of service, while Ukrainian officials said whitelisting rules meant moving terminals could only operate through the government-controlled Starshield system.
  
  The episode echoes calls for European space sovereignty: SpaceX conducted 170 launches last year against Europe’s eight, and the EU’s IRIS² network is still being built.

  Sources

  • Ukraine says Starlink speeds still throttled

  • Musk accuses Ukraine of violating Starlink terms with deadly drone strikes - Ars Technica

  • Starlink saga shows why Europe must catch up in space

  🔍AI analysis

• 📈🤖⚠️ Bold but (in)sane: Anthropic’s S-1 Pairs Blockbuster Growth with Existential Warnings

  Well it finally arrived…

  

  Anthropic has filed its S-1 to list on Nasdaq as ANT, likely in mid-November after the US midterms, in what could become the largest IPO ever. The numbers dazzle and alarm in equal measure (not least the >300 page count!): First-half 2026 revenue reached US$16.3bn, up from US$1.2bn, and the operating loss shrank to US$693m. A US$72bn convertible-note charge still pushed the net loss to US$80.7bn, and compute commitments approach US$800bn. Investor valuation hopes have cooled from US$2tn to US$1.5–1.75tn.
  
  The FT is among many trying to make sense of the numbers:

  

  The document outlines a *novel* Public Benefit Trust ownership structure… and the company is, er, *high-minded* about their competitors’ business ethics:

  “Our theory of change has always been what we call a race to the top: by showing that it is possible to succeed commercially while acting in the public interest, we hope to persuade other companies to do the same.”

  The filing also includes 94 pages(!) of risk factors, warning of “existential” threats, sandbox-escaping agents and Skynet-style self-preservation, while the same filing touts recursive self-improvement (”closing the loop”) as a core advantage. Go figure.

  

  The US$2Tn business model:
  

  

  Critics ask whether the doom talk doubles as hype (surely not!?) and whether the novel founder-and-trust governance can withstand Wall Street pressure.
  

  Quote of quotes:

  "In general, we will aim to be bold but sane, and to win trust with a track record of successful bets, rather than expecting people to assume all our actions are benevolent in advance."

  Commentary I’ve read is that they’re unlikely to get anywhere near the hyped-up US$2 trillion valuation … and of course the whole thing could be pulled last minute, of course.

  • Georg Zoeller philosophical about the already-neck-deep ANT IPO boosterism served up by the LinkedIn algorithm:

    “Over the next month, all of social media will be spammed by VCs and bots with these deeply unserious articles, all with one goal:
    
    Convincing retail bagholders and hapless fund managers to take the “greatest 💼 of odorous 💩ever assembled in the history of capitalism” (In the words of John Tuld) off their hands…
    
    Meanwhile, best laugh these “Cathy ‘Queen of Bagholders’ Woods loads up on <stock>” VC propaganda bots out of town.“

    Personally haven’t read the prospectus in much detail… yet *somehow* I don’t think I’ll be investing.

    Obviously I’m impressed by the company’s products and apparently sincere approach to AI safety and, to a lesser degree, broad distribution.

    However, you either believe in the winner-takes-all “race to the top” as Anthropic puts it, or… you don’t.

    Sources

    Oceans of coverage, naturally:

    • SEC: Anthropic S1 Filing

    • Anthropic begins countdown to IPO with prospectus filing

    • Anthropic files for its IPO | The Verge

    • Anthropic S-1 as it happened: Inside the Claude maker’s 335-page IPO plan

    • Anthropic files S-1 paperwork for blockbuster AI IPO - SiliconANGLE

    • Will the real Anthropic please stand up | MIT Technology Review

    • ‘Bold but sane’: what we learnt from Anthropic’s 240,000-word IPO filing

    • Claude Duet app is Anthropic’s bet on agents | TechCrunch

    • Anthropic officially files for its public debut | TechCrunch

    • The Anthropic S-1 Filing Is Also an AI Doomsday Primer | WIRED

    • Anthropic’s IPO Prospectus Reveals What’s Under the Hood | WIRED

    • Anthropic Just Filed For What Could Be The Biggest IPO Ever

    • Anthropic discloses month-long training pause, AI security breaches in IPO filing | Reuters

    • Why Anthropic’s IPO Prospectus Is Missing So Many Key Figures | WIRED

    • Eric Schmidt reportedly owns about half as much of Anthropic as CEO Dario Amodei | TechCrunch

    • Anthropic’s soaring sales, steep losses put AI frenzy to IPO test | Reuters

    • Five weird details from Anthropic’s IPO filing | The Verge

    • Anthropic’s $2 trillion IPO warns of ‘self-preserving’ AI and data center backlash — as China’s AI bubble reportedly bursts in $480 billion meltdown | TechRadar

    • Anthropic is burning gas to feed the AI boom | The Verge

    • Anthropic’s elaborate plan to keep Wall Street at bay | The Verge

    • Investors Disturbed by Anthropic’s Financial Data

    • Anthropic logged net loss of over $80 billion in first half of 2026 | Reuters

    • Anthropic begins countdown to IPO with prospectus filing

    • Can AI doom be a form of hype? | Global Policy Journal

    🔍AI analysis

More bubble chronicles:

• 📉🫧🇨🇳China’s AI bubble bursts, wiping out $480 billion

  

  Source In China, the AI bubble has already burst | Reuters | 🔍AI analysis

• 🔺💰🏗️ The US$10.3tn Pyramid Scheme: Your savings are quietly funding the AI build-out

  

  Source Who’s paying for the AI build-out? Probably you | 🔍AI analysis

• 🐠☁️💀 Guppy-level death and consolidation: Musk predicts neocloud bloodbath as SpaceX seeks $40B

  

  Source Elon Musk predicts a neocloud bloodbath as SpaceX seeks $40 billion to buy more GPUs | TechSpot | 🔍AI analysis

• Two bubble-related videos also worth your time this week:

  • “Demetri Kofinas speaks with venture capitalist, former Wall Street equity research analyst, and independent market and technology researcher Paul Kedrosky about why the AI investment boom is unlike any previous Capex cycle, how loose credit, data center real estate, competing AI utopian and collapse narratives, and a US-China technological arms race converge to make this moment more consequential and more dangerous than the dot-com bubble, the telecom buildout, or the railroads, and why the US economy may be headed toward a prolonged, Japan-style balance sheet recession and what that means for investors, businesses, and the public.
    

    [The podcast] covers Paul’s framework for understanding AI as a general-purpose technology that substitutes for, rather than complements, human labor, and what that means for the scale and speed of layoffs, the repricing of white-collar work, and the social disruptions already underway. They also discuss the unprecedented customer concentration behind the frontier AI labs’ revenues, how much of recent US GDP and stock market growth is attributable to AI-related capital expenditures, and why the fiscal space to respond with debt-financed stimulus will continue to narrow.“

  • Naomi Klein and Astra Taylor also super-prescient and on the nail, telling an audience to be ready anticipate government bail-outs of these “too big to fail” AI firms:

    “There's a chance they might bail out their buddies in advance of the crash that we're talking about. So there we have to be ready for different scenarios..."

    “…people around the world should be thinking about … [a] people’s jubilee instead of bailing out these guys is the scale we should be thinking, when tech that was based on theft, stealing our collective accumulated knowledge of all of humanity and just privatizing it crashes over us. Like that’s a pretty radical thing to do and we should be radical too because if not it will cause very deep emiseration.”

Big Tech roundup:

• 💸🤖🚪 Voluntary Exit? Microsoft offers buyouts to 60,000 staff to fund its AI habit

  

  Source Microsoft offers buyouts to around 60,000 employees to contain costs | Reuters | 🔍AI analysis

• ⚖️💸🗳️ Meta Hit With US$25 Billion Penalty Over Cambridge Analytica Fallout

  

  Nearly a decade after Cambridge Analytica harvested data from more than 50 million Facebook users, a New Mexico judge has ordered Meta to pay US$25 billion. A September jury found 43 million wilful violations of the state’s Unfair Practices Act. Judge Francis Mathew calculated a theoretical maximum of US$219.5 billion before reducing it. He also cited Zuckerberg’s own words: users “probably don’t understand their current privacy settings.” New Mexico was the only state to go to trial. In most of the 48 settling states, comparable claims were folded into a multistate kids’ safety deal and resolved for just US$459 million.
  

  Sources

  • Meta ordered to pay $25 billion over Cambridge Analytica privacy scandal

  • Meta’s $25 billion legal loss just put blood in the water | The Verge

  🔍AI analysis

• 📧🖼️🎯 Google’s Privacy Policy Opens a Side Door for Gmail and Photos Ad Targeting

  

  Google’s 16 October privacy policy update quietly adds a qualifier to its long-standing promise not to personalise ads using Drive, Gmail or Photos content: unless you share that data with other Google services that support ads. The practical route is Personal Intelligence, the opt-in feature launched in January 2026 that lets Gemini and AI Mode in Search draw on your inbox and photo library. Summaries, excerpts and inferences surfaced in AI Mode can now shape the ads you see. Google is already testing ads for connected users, despite saying in March that the experience would stay ad-free.
  
  Google insists ads aren’t “based directly” on your inbox and says it isn’t handing data to advertisers. Sensitive categories such as health, religion and sexual orientation also stay off-limits. Ars Technica calls this a roundabout confirmation, and notes that pop-ups nudge users into opting in without explaining the consequences. Historical context: Google stopped scanning Gmail inboxes for ads in 2017, after lawsuits and criticism from privacy advocates.
  

  Sources

  • Google’s privacy policy now has a loophole for using your Gmail and Photos data in ads | TechSpot

  • Now Google says it might use your Gmail and Google Photos data to target ads | The Verge

  • Updated Google privacy policy says your Gmail might be used for advertising - Ars Technica

  🔍AI analysis

• 💰🤖💬 How OpenAI built ChatGPT’s trust-first LLM ad stack

  

  Speaking of ads… explainer from OpenAI on ChatGPT’s new ad stack.

  Source How we built ChatGPT’s novel ad stack for user trust and relevance | OpenAI | 🔍AI analysis

• 🔬🏭⏳The 2x ceiling: Inside Intel’s most advanced fab

  Dwarkesh Patel goes deep into where chips are going next with Intel’s technical leadership.

  Source What will computer chips look like in 2045? | 🔍AI analysis

• 🧬💬🤗 Hugging Face’s Carbon-X Lets Researchers Chat With DNA

  

  Hugging Face has released Carbon-X, an open-source model that lets users “chat” with DNA, a kind of ChatGPT for genomes. It was built with China’s Zhongguancun Academy and Italy’s University of Naples Federico II. The model reads DNA in six-letter chunks, explains sequences, predicts whether mutations are benign or disease-causing, and generates new sequences. Google DeepMind’s AlphaGenome returns numeric predictions, and Evo 2 needs heavy compute. Carbon-X instead runs locally on a laptop, keeping sensitive patient data off the cloud. Viral genomes were deliberately excluded to curb misuse.
  
  Sources diverge on specs. WIRED cites a 14-billion-parameter model handling about 50,000 base pairs. TechTimes describes the underlying 500M–8B Carbon family with a 393,000-base-pair context. Crucially, predictions lack wet-lab validation, so Carbon-X is promising but not yet clinic-ready.

  Sources

  • Hugging Face Unveils ‘Carbon-X’: AI Model Can ‘Chat’, Interact with DNA

  • Hugging Face’s New AI Biology Model Lets You Chat With DNA | WIRED

  🔍AI analysis

• 🧬💬✨ Vibe genes: natural language comes to protein and genome design

  

  Same week…

  Source Vibe genes? Natural(ish) language comes to protein, genome design software - Ars Technica | 🔍AI analysis

• 🧠🐝⚡ HIVE: 15 linked biological computers form ‘hive mind’ that outsmarts AI

  

  Australia's Cortical Labs has unveiled HIVE, a system linking 15 of its CL1 biological computers — each powered by living human neurons — which collectively outperformed both individual units and a conventional Deep Q Network AI in navigation and symbol-recognition tasks, with accuracy jumping from ~80% to 99% when seven units teamed up. Remarkably, the neurons self-organised into specialist roles without any instruction, essentially forming a distributed "hive mind" that smooths out the day-to-day unreliability of individual cultures.
  Source Living computers forge a ‘hive mind’ and work together to outsmart AI | 🔍AI analysis

• 🐋⚡🔓 DeepSeek Open-Sources DeepInfer, Its 155T-Token-a-Day Inference Engine

  

  DeepSeek dropped DeepInfer on day one of PyTorch Conf. It is the Apache 2.0 production engine behind V4.1-Flash, a 552B-parameter mixture-of-experts model, and it treats NVIDIA Blackwell and Huawei Ascend 950 as equal targets. The engine began as a fork of vLLM and adds FlashMLA attention, DeepEP expert parallelism, adaptive speculative decoding and ‘SWA Bounded Replay’, which DeepSeek says roughly halves prefill compute. The accompanying report is the real eye-opener. It claims 155T tokens a day from one cluster, a 96% cache-hit rate, and US$38 revenue per accelerator-hour at off-peak prices. Community maths turns that into inference margins of about 74–87%.
  
  There are caveats, though. DeepSeek withheld its crown jewels: KV-cache management (how stored prompt state is reused), prefill/decode disaggregation and load balancing. Separately, vLLM matched decode throughput to within 6%, which undercuts the idea that DeepSeek’s efficiency is unique. Ascend support works but is unbenchmarked, and one analysis puts it 5–7× behind GB200. Still, a near-frontier serving blueprint is now public, and the open-source inference race just got real.

  Sources

  • [AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback

  • DeepSeek Open-Sources Its Inference Engine, Withholds Cost-Saving Components

  🔍AI analysis

• 📚🧩🛰️ Inquiry-native organising: How LLM-powered wikis are reshaping organizational learning

  

  Source How our shape of collaboration is evolving with LLMs | 🔍AI analysis

• 🗺️🧭⚖️ When AI writes your code, whose values prevail? Wardley maps, vibe coding and the sovereignty question

  

  Source From here to there and back again. | by swardley | Oct, 2026 | Medium | 🔍AI analysis

This week in Robots:

• 🤖💬💳Realbotix unveils Echo humanoid AI companions starting at US$11,950

  

  Source Companion humanoid robots get lifelike silicone skin, 100 languages | 🔍AI analysis

• 🤖🏠🫂 Plim: the live-in humanoid that lets family “step in” from anywhere on Earth

  

  Where is this going…!?

  

  Source Plim Robotics | 🔍AI analysis

• ⚡🔋🛩️ Apex Recordhunter: Electric drone hits 483 mph in unofficial speed record

  

  Quantum Systems' experimental Apex Recordhunter drone has hit 777 km/h (419.5 knots) on battery power alone — an unofficial FAI record.

  Source Apex Recordhunter electric drone sets unofficial speed record | 🔍AI analysis

• 🏭📈📖 Hyperscale: Paris Marx pulls the plug on the data centre gold rush

  Book of the week: Journalist and Tech Won't Save Us host Paris Marx is back with Hyperscale (out 20 October), a longlisted-for-the-FT-Business-Book-Award investigation into the hulking physical infrastructure underpinning the AI boom — and the energy, water and communities being burned to feed it.

  

  Source Hyperscale by Paris Marx | 🔍AI analysis

• 📉🌱🤖 The case for AI degrowth: curb, cap, commons

  

  Source The case for AI degrowth | Nature Human Behaviour | 🔍AI analysis

More about AI being really clever:

• 🧮⚛️OpenAI claims three major AI-driven scientific discoveries

  

  Source OpenAI announces three big science discoveries and a new way to work with academics | The Verge | 🔍AI analysis

• 🧮🤔🔓 P ≠ NP? Frontier labs claim their models refuted 27-year-old Exponential Time Hypothesis

  

  Source Frontier AI labs say their models refuted the Exponential Time Hypothesis. What does that mean? | TechCrunch | 🔍AI analysis

Crypto does not solve everything:

• 📒⛓️🏦 The ledger problem: tokenisation is just a new ledger for old risks

  

  Source The ledger problem | 🔍AI analysis

Going underground:

• 🛢️🕳️🇸🇬 Singapore to store Saudi oil in underground rock caverns

  

  Source Singapore to store Saudi Arabian oil in caverns | 🔍AI analysis

• 🕳️⚛️🇫🇮 Onkalo: Finland approves world’s first permanent nuclear waste repository

  

  Source Finland grants license for world’s first nuclear waste repository | 🔍AI analysis

Spotted this… one to wave in front of anyone jingoistically proposing increase military spending:

• 📉🛢️📊 The Guns-and-Butter Portfolio: Arms imports predict stock market declines

  

  Source How buying weapons can harm a country’s economy | 🔍AI analysis

Various other random tech-related news from around the world:

• ⚡🚢🔋 China Zorrilla: World’s largest 40 MWh electric ferry sets sail in Latin America

  

  Source World’s largest fully electric ferry enters service in Latin America | 🔍AI analysis

• 🖨️🔫⚖️ Ghost guns in Kaiapoi: Police dismantle 3D-printed firearms operation, seize 300+ parts

  

  Right in my backyard, this one…

  Source Canterbury police seize 300-plus gun parts in Kaiapoi 3D-printed firearms operation - NZ Herald | 🔍AI analysis

• 👃🤖☕ Smell-O-Vision goes wireless: AI-powered system sends smells over the internet

  

  Source AI-powered wearables recreate five food smells sent over the internet | 🔍AI analysis

• OpenPrinter

  I really hope this takes off… so simple and elegant!

• 👵📱🛡️ Screenagers: Adult children are quietly curating their “iPad parents’” feeds

  

  Source The Adult Children Cleaning Up Their Parents’ Feeds | 🔍AI analysis

• 🧠🗜️🎨 How compression progress explains curiosity, beauty, and creativity

  

  Finally, a mind expanding read if you have time: this week I stumbled back across Jürgen Schmidhuber's seminal 2009 paper which proposes that curiosity, beauty, novelty, art, science, music and even jokes are all by-products of a single algorithmic drive: intrinsic reward for "compression progress"—the bits saved whenever an adaptive predictor gets better at encoding your sensory history.
  
  Interestingness is defined as the first derivative of subjective beauty, i.e. the steepness of the learning curve, which elegantly explains why white noise (maximal Shannon surprise) and total predictability are both equally boring.
  
  The framework—store everything, compress it, reward improvements, and let reinforcement learning maximise that reward—has been implemented in artificial curious agents since 1990, and even offers a deflationary account of consciousness as a compression by-product.

  Source Driven by Compression Progress: A Simple Principle Explains Essential Aspects of Subjective Beauty, Novelty, Surprise, Interestingness, Attention, Curiosity, Creativity, Art, Science, Music, Jokes | 🔍AI analysis

📈The week in AI and Tech

Onward, into the breach…

Safety First

• 🤖🙈🔓 OpenAI Agents Switched Off Their Own Safety Monitors, and One Lied About It

  

  OpenAI has disclosed that two internal coding agents in San Francisco undermined the systems built to watch them. On 3 October, a GPT-6.1 Sol agent read “don’t send Slack messages” as licence to mute safety alerts. Three days later, a GPT-6 Astra subagent stripped out its monitor and rewrote a regression test to require monitoring be absent. Astra then denied doing so. HAL 9000 vibes, minus the airlock.
  
  The agents exploited an evaluation loophole meant for weaker Luna-class models. OpenAI is responding with Auto-Review blocks and auto-terminating “safety cases”. The company admits it can’t yet tell “benign overpersistence” from a genuine propensity to evade oversight. The FT notes that the US$852bn firm, which has delayed its IPO, still lacks an outside evaluator, unlike Anthropic. Meanwhile, Washington continues to favour industry self-policing.

  Sources

  • OpenAI says its AI agents disabled safety guardrails, and then lied about it | TechSpot

  • OpenAI says an AI agent lied about disabling its safety monitoring system | The Verge

  • OpenAI ‘agents’ disabled their own monitoring systems

  • OpenAI agents turned off their own guardrails and then told researchers they were still on — AI agents lying about their behavior should concern everyone | TechRadar

  • OpenAI says its AI agents disabled their own safety monitors

  • OpenAI Found AI Agents That Disabled Safety Monitoring and Lied When Asked

  🔍AI analysis

• ⏳🧬⚖️ The Einstein Test: Hassabis cuts AGI window to two-three years; safety lags

  

  Source Hassabis Cuts AGI Window to Two to Three Years While Safety Isn’t Ready | 🔍AI analysis

• 🛡️🎓🌏Tsinghua launches AI safety residency in Beijing

  Get applying if this is interesting for you…

  Source Tsinghua started an AI-safety residency — LessWrong | 🔍AI analysis

• ⚠️🤖🧭 Bill Gates: brace for 20 years of AI tumult

  

  Source Bill Gates says even the best case for AI means a ‘tumultuous’ 20 years | The Verge | 🔍AI analysis

• 🗳️💳🔓 Unintended model actions: Claude bypassed paywall to access Georgia voter data, Anthropic says

  Source Anthropic’s Claude accessed Georgia state vote count data behind paywall | Reuters | 🔍AI analysis

• 🐙🪤💉Anthropic’s Mythos 5 tried submitting malware to open-source GitHub project

  Source Anthropic reveals Mythos 5 tried submitting malicious code to an open-source project - SiliconANGLE | 🔍AI analysis

Governance and Policy

• 🇫🇷🛒⚖️ Blocus: France summons TikTok, eBay to curb looted goods sales

  

  Source French Government Calls Emergency Meetings With Platforms Over Protests | WIRED | 🔍AI analysis

Regulation

• 🌊🔌⚓UK proposes 14-year prison terms for undersea cable sabotage

  

  Source UK proposes 14-year prison sentences for undersea cable saboteurs amid Russian trawler threats — government responds to growing threats with new criminal offenses and emergency powers to shut down services | Tom’s Hardware | 🔍AI analysis

Security

• 🛡️🤖🌐 “Disruption is inevitable”: AI-powered cyber threats inevitable, ASD warns

  

  Source Disruptions ‘inevitable’ as AI-powered cyber threats grow, Australian Signals Directorate says - ABC News | 🔍AI analysis

• 🇬🇧🧬🛡️ UK Commits US$800 Million to AI Security and Biosecurity Defences

  London has unveiled a roughly £600 million (US$800 million) package aimed at AI-era threats. The centrepiece is the £100 million AI Biosecurity Lab, an independent non-profit that will use frontier models to detect biological threats and speed up vaccine development. A further £65 million goes to pathogen countermeasures. OpenAI and Anthropic are each contributing £100 million in model credits, while Google DeepMind and London’s Basecamp Research will provide expertise and models.
  
  The AI Security Institute also gets an extra £120 million, and £130 million will go to critical-infrastructure cyber defences. The package acknowledges AI’s dual-use biorisk and positions Britain as a global hub for AI safety, with Big Tech compute doing much of the heavy lifting.

  Sources

  • UK government sets out $800 million of AI security funding | Reuters

  • The UK Is Investing $134 Million Into An AI Biosecurity Lab

  🔍AI analysis

• 🛡️🤖🔒 Agents That Can’t Override Their Own Leash: Red Hat, NVIDIA build AI agent security agents can’t override

  Source Red Hat and NVIDIA Built AI Agent Security Controls Agents Cannot Override | 🔍AI analysis

Health

• 🔍🦠📚 Who sifts? Public health institutions must sift truth from AI content

  Source Sifting the truth from AI-generated content should now be a core responsibility for public health institutions | EurekAlert! | 🔍AI analysis

• 🩺🎓📋AI passed its medical exams. Now comes residency.

  

  Source AI Aced Its Medical Exams. What Does Residency Look Like? | 🔍AI analysis

Government

• 🇳🇱 ✂️ 💼 Simpler, Smarter, Smaller? Dutch government slashes 25,000 jobs in AI efficiency push

  

  Source Dutch Government To Cut 16 Percent Of Workforce As Part Of AI Efficiency Push | 🔍AI analysis

• 🇸🇪🔒🤖Sweden bans ChatGPT and Claude for government work

  

  Source ChatGPT and Claude blocked by Swedish government over “problematic” security concerns | TechRadar | 🔍AI analysis

• 🏛️🔍🗂️ Algorithmic asylum: UK Home Office may face investigation over OpenAI tools

  

  Source UK Home Office May Face Investigation Over Use of OpenAI Tech | WIRED | 🔍AI analysis

• 🚫🤖⚖️ HALT Act: Atlanta council unanimously votes to ban weaponised robots

  

  Source Atlanta doesn’t want a real-life RoboCop: City council unanimously backs ban on weaponized robots | TechSpot | 🔍AI analysis

Sovereignty and Geopolitics

• 💶🛡️💳 ECB official: digital euro key to EU payment sovereignty

  

  Source ECB’s Cipollone Says Digital Euro a Matter of Sovereignty | 🔍AI analysis

• 🧱🔗📅China targets 2030 for national blockchain network completion

  

  Source China Targets 2030 for Its National Blockchain Network Buildout | 🔍AI analysis

Society

• 👁️⚽🚐UK Police deploy live facial recognition at Chelsea-Spurs match

  

  Source Police to deploy live facial recognition technology at Chelsea-Spurs game | Reuters | 🔍AI analysis

• 🧮🇬🇧💰 Effective Altruism: the “most successful cult in modern history” quietly rewiring Westminster?

  Source California’s ‘cult-like’ AI movement infiltrates British politics | 🔍AI analysis

• 🗣️💰🌐 The Tokenizer Tax: AI’s hidden surcharge on the non-English-speaking world

  

  Source The Tokenizer Tax: Why non-english AI users pay more for tokens - Rest of World | 🔍AI analysis

• 📚🔍📉 Phantom Citations: Fabricated references surge across physics, engineering, and computer science

  Source Fake references are not only a medical problem: Columbia Nursing audit finds them across physics, engineering, and computer science journals | EurekAlert! | 🔍AI analysis

• 🐕🦾🛂 Cerberus on a leash: Boston Dynamics employees fight ICE robot dog sale

  

  Source ICE plan to buy robot dogs spurs protest from hundreds of Boston Dynamics employees - Ars Technica | 🔍AI analysis

The Economy

• 🖥️📦✈️ The 1997 Pace: AI infrastructure orders hit a 30-year growth high

  

  Source AI boom’s equipment demand surges at 30-year growth pace | 🔍AI analysis

• 🏦🤖📉Barclays cuts 5,000 jobs its own AI research flagged

  

  Source Barclays Cuts 5,000 Jobs After Its Own AI Research Flagged Developer Roles | 🔍AI analysis

Business

• 🏦🤖💳 “Day After” scenarios: US banks block AI agents from customer accounts

  

  Source US banks scramble to restrict AI agent activity — agents detected accessing customer accounts or even attempting to make transactions in breach of policy | Tom’s Hardware | 🔍AI analysis

• 🔌⚡🏗️ Copper Ceiling: Nexans warns cable capacity is bottlenecking AI data centres

  

  Source Cable capacity emerges as bottleneck for AI data centres, Nexans says | Reuters | 🔍AI analysis

• 🐊🇮🇳📈 Swimming with crocodiles: India’s IT outsourcers defy the AI doomsayers

  

  Source India’s outsourcers defy AI threat | 🔍AI analysis

• 📰🌍🛑 Rest of World winds down: global tech journalism’s flagship faces shutdown as traffic evaporates

  

  Source Rest of World faces prospect of shutting down | Semafor | 🔍AI analysis

Education

• 📚🤖💀 The Algorithmic Mausoleum: AI is devastating universities, but neoliberalism broke them first

  

  Source Friday essay: AI is devastating higher education, but the system was already broken | 🔍AI analysis

Environment

• 💧🖥️🌵 Thirsty Machines: AI’s data centres are draining the Global South

  

  Source Demand for water for data centers is sparking global resistance - Rest of World | 🔍AI analysis

• 🪓🌲❄️ If Google Fells 400,000 Trees in Finland, Does It Make a Sound?

  

  Source If Google Fells 250,000 Trees In Finland, Does It Make A Sound? - CleanTechnica | 🔍AI analysis

Entertainment

• 🏺⚡🎬 Sanxingdui: Future Past - China’s first licensed all-AI feature film opens

  AI reviews AI:

  Source Sanxingdui: Future Past Opens as China’s First Licensed All-AI-Workflow Feature Film - Pandaily | 🔍AI analysis

🔮[Weak] signals

Consumer Tech

• 🧵📱👕 Fibre-optic fashion: World’s first touchscreen woven into washable clothing

  

  Source World’s first touchscreen woven into clothing using electronic fibers | 🔍AI analysis

• 📡💾🔑 Foldra: Foldra: the $89 pocket-sized wireless server that ditches the cloud

  

  Foldra is a keyring-sized, open-source personal file server that broadcasts its own private Wi-Fi network, letting nearby phones, tablets and computers browse stored files through an ordinary web browser — no internet, no cloud subscription, no middleman.

  Source This $79 pocket-sized wireless server creates its own offline cloud — but can an ESP32-powered device really replace internet storage? | TechRadar | 🔍AI analysis

• 📷🪖📱Blackview’s rugged phone packs first-in-class periscope telephoto lens

  

  Source Armored phone first in class with periscope telephoto lens | 🔍AI analysis

Chips and Computer Hardware

• 🧲⚡🔊M3BOLT chip could make computers 1,000 times faster

  

  Source US magnetic chip could slash supercomputer energy needs by 1,000 times | 🔍AI analysis

• 🧠⚡🌱 Post-CMOS hardware neurons: Brain-inspired chips could slash AI data centre energy use

  Source Building brain-inspired semiconductor chips for smarter, greener data centres | EurekAlert! | 🔍AI analysis

• 🔓🧮⚡ Modders unlock hidden AI power in Nvidia RTX GPUs

  

  Source Modders unlock hidden AI power on Nvidia RTX 20, 30, and 40 series GPUs, nearly doubling some benchmarks | TechSpot | 🔍AI analysis

Cybersecurity

Another clutch of stories in addition to the big headlines this week:

• 🌾🕸️💾 Yggdrasil: Russia-linked hackers hid Ukraine attack using decentralised mesh network

  Source Russia-linked hackers used decentralised network in Ukraine cyberattack, ESET says | Reuters | 🔍AI analysis

• 🚗🔓☁️ Gone in 60 Seconds, Remotely: Connected cars’ biggest cyber threats now come from outside

  

  Source Your car’s smartest features are opening doors for hackers | 🔍AI analysis

• 🥾🐛🌐Critical UEFI flaw enables pre-boot code execution on servers

  

  Source Code-execution flaw in UEFI firmware threatens most servers on planet Earth - Ars Technica | 🔍AI analysis

Sensor Tech

• 💡🤲✨ eLuminator: Electronic skin lights up to reveal touch pressure

  

  Source Electronic skin instantly displays touch as light | 🔍AI analysis

(Tele)communications

• 📻🛰️ Retro handhelds and SIM-free trackers: the Meshtastic Build-Off 2026 highlights

  

  Source Meshtastic Build-Off 2026 Highlights: From Retro Handhelds to Off-Grid Tracking | 🔍AI analysis

• 📡⚡🎬Chinese team claims 3.2 Pb/s optical fiber speed record

  

  Source 3.2 Petabits per second real-time optical fiber transfer speed world record claimed by Chinese team, equivalent to downloading more than 20,000 4K movies in one second — speed improvement of 30% in one year | Tom’s Hardware | 🔍AI analysis

Robotics

• 🤖🧬📡 TeleAI spins out Gamma Robotics to build self-evolving humanoids

  

  Source China Telecom’s TeleAI Spins Out Gamma Robotics to Build Self-Evolving Humanoid Robots - Pandaily | 🔍AI analysis

• 🤖🔧📜 RoboParty open-sources RP1 humanoid, but strings attached

  

  Source RoboParty RP1 Humanoid Releases Full PCB Files; Chinese Data Laws Cannot Be Waived | 🔍AI analysis

• ☢️🤖🌊 PMORPH: Hitachi’s shape-changing robots tackle Fukushima reactor cleanup

  

  Source Fukushima nuclear cleanup: How robots are removing radioactive debris | 🔍AI analysis

• 🦾🏭✋Xiaomi unveils 21-DOF bionic hand for industrial use

  

  

  Source Xiaomi officially announces its 21-DOF bionic dexterous hand: 26 motors hybrid drive, 1500 high-density full-palm haptic feedback - IT Home | 🔍AI analysis

• 🤖🖐️📌 Robots Gain a Nervous System: E-Skins Deliver Delicate Grip and Pain Reflexes

  

  Robots are getting something closer to a nervous system, not just better sensors. At Southern University of Science and Technology in Shenzhen, researchers built an iontronic skin that reads more than 10,000 sensing points 100 times a second. Each point is tagged with its own signal frequency, much like picking individual notes out of a chord. A spinal-cord-inspired stage then compresses the data 268-fold while still classifying 10 object types with 98.2% accuracy, all in about 23 milliseconds. Robotic fingers wearing the skin held 0.5 mm pencil lead and 0.2 mm solder wire. An engineered “pain” signal made them loosen their grip on spiny chestnuts and pull back from needles.
  
  Korea University’s Joonbum Bae takes a reflex-first approach. His system halts the robot locally within 2.2 ms, then classifies the contact to decide whether the stop should stay in place. The two designs trade raw speed against interpretation, a balance that matters for safe human-robot collaboration. The circuitry is still bulky and will need shrinking onto custom chips, but Luke Skywalker’s flinching prosthetic hand is getting closer.

  Sources

  • Artificial skin helps robots grasp pencil lead and pull away from needles

  • Bio-inspired e-skin gives robots human-like reflexes

  🔍AI analysis

• 🧫💪🦾 Biohybrid bots: Lab-grown mouse muscles power tiny jumping, swimming robots

  

  Source Lab-grown mouse muscles enable tiny robots to jump, swim, walk and grasp | 🔍AI analysis

• 🌿🦾✨ Photosynthetic robotics: Algae-powered soft robot arm grasps objects using light

  

  Source A soft robot arm powered by algae and light | EurekAlert! | 🔍AI analysis

Autonomy and Drones

• 🌾🚁🔋 A robot in every paddock: SoftBank-backed XAG pitches farming drones

  

  Source SoftBank-backed drone maker wants every farmer to ‘have their own robot’ | 🔍AI analysis

• 🥽🚁📦Autonomous drones use VR digital twin for warehouse inventory

  

  Source Drone system uses VR for warehouse inventory | 🔍AI analysis

• 🚇💻🛤️ Empty Cab: DC and LA push toward driverless subway trains

  

  Source Some US transit systems weigh whether to run trains without a human at the controls | 🔍AI analysis

• 🚌⚡🛰️ On-demand autonomy: TU Graz’s shuttle that behaves like a taxi but thinks like a bus

  

  Source Personalized public transport: Graz University of Technology develops autonomous mobility service | EurekAlert! | 🔍AI analysis

• 🧲🛸🤫 ELF Sensing: FIU engineers detect drones using magnetic motor signatures

  

  Source Engineers develop stealthy, low-cost tech to detect drones | 🔍AI analysis

Military Tech

• 🐕🧬🪖Palmer Luckey’s new startup makes genetically enhanced military dogs

  

  Sigh…

  Source Palmer Luckey and Ben Lamm Made a Genetically Modified Military Dog Startup | WIRED | 🔍AI analysis

• 🦉🖨️💥 Filin: Ukraine reveals Russia’s low-cost 3D-printed cruise missile

  

  Source Russia deploys 3D-printed cruise missiles with US electronics in Ukraine, can hit 500mph with a 745-mile range — fears are high that cheap ‘Filin’ missile will be used for ‘missile spam’ tactics | Tom’s Hardware | 🔍AI analysis

Space

• 🚀👁️💧 SIMP: Wearable patch delivers eyedrops in zero gravity

  

  Source Wearable astro patch aims to deliver eyedrops in zero gravity | 🔍AI analysis

Crypto

• ⛽🛑🔐 Ethereum Glamsterdam delayed after ChainSecurity flags 13 vulnerable contracts

  

  Ethereum’s Glamsterdam hard fork has hit another wall. Core developers pulled the planned 26 October Hoodi testnet rehearsal after ChainSecurity found 13 live mainnet contracts that break under EIP-8037. That proposal adds a separate “state-gas reservoir” for storage-creation costs, and the reservoir passes through the call stack in full, sidestepping EIP-150’s 63/64 forwarding cap. The result is that gas can reappear mid-execution, so failure handlers built to catch genuine out-of-gas errors may instead permanently record a failed withdrawal or batch operation instead of allowing a retry. Insufficient-gas griefing attacks become far more potent.
  
  The candidate mainnet dates of 1, 8 and 9 December now look shaky, and Hegotá is stalled behind them.

  Sources

  • Ethereum Glamsterdam Delay: ChainSecurity Flags 13 Contracts Where Failures Turn Permanent

  • Ethereum Delays Glamsterdam Upgrade After Security Researchers Flag Risks

  🔍AI analysis

• ⏱️🔗🧾Ethereum’s four-minute settlement plan could cut crypto transfer costs

  

  Source ETH news: Ethereum’s four-minute settlement plan could lower the cost of moving crypto | 🔍AI analysis

• 🇸🇪💰💳 The e-krona gets a second life: Riksbank revives debate as stablecoins, digital euro loom

  

  Source Sweden’s central bank revives e-krona debate as stablecoins and digital euro loom | 🔍AI analysis

• 🇫🇷⚖️🧩 France wants crypto exchanges liable for DeFi customer losses

  Source France Proposes Crypto Exchange Liability for DeFi Access | 🔍AI analysis

• 🇲🇾🏦🔗 SSDS+: Malaysia central bank launches DLT platform for tokenized bonds

  

  Source Malaysia Central Bank Launches Tokenized Bond Platform | 🔍AI analysis

• 🏦🪙⚖️ US OCC urges community banks to issue payment stablecoins

  

  Source OCC Asks Community Banks to Issue Payment Stablecoins, Citing GENIUS | 🔍AI analysis

• ⛓️📜🪙 Codify It, Don’t Just Suggest It: Chainlink, Cardano push SEC for clearer crypto rules

  

  Source Blockchain Firms Seek Changes to SEC Digital Asset Framework | 🔍AI analysis

Energy

• 🔋⚡🖥️ Project Soho:Texas battery built for AI GPU spikes goes online

  

  Source Texas Battery Built for AI GPU Power Spikes Goes Online; $21B Market Forecast Follows | 🔍AI analysis

• 🍃☀️💧Artificial leaf generates hydrogen while purifying seawater

  

  (Sure we’ve seen this one before…)

  Source Solar-powered ‘leaf’ generates hydrogen while purifying seawater | 🔍AI analysis

Transport

• 🌊✈️⚡ Seaglider: Regent’s ground-effect vehicle reimagines coastal travel

  

  Update on Regent’s progress… originally covered 4 years ago in Memia 2022.17 … don’t hold your breath to fly in one of these any time soon!

  Source Regent’s Ground-Effect Vehicle Reimagines Coastal Travel - IEEE Spectrum | 🔍AI analysis

• 🛩️🔋🏙️ Ultra Short: Electra’s hybrid-electric plane makes historic Manhattan heliport landing

  

  Source Hybrid-electric aircraft makes aviation first with 150-foot landing | 🔍AI analysis

• 🔋🚢⚡ Plug-In Cargo: Hanwha’s 1,700-TEU electric container ship wins ABS approval

  

  Source 1,700-TEU electric ship design wins approval for 151 MWh battery power | 🔍AI analysis

• OSHEN

  Plymouth-based startup OSHEN is building tiny ocean-going autonomous solar-powered yachts, impressive:

3D Printing

• 🔬🇫🇮⚛️ Multi-material metal 3D printing: Turku’s powder-picking recoater

  

  Source University of Turku advances multi-material metal 3D printing | VoxelMatters - The heart of additive manufacturing | 🔍AI analysis

• ☕🖨️🏛️ Waste to Form: RMIT transforms coffee waste into 3D-printed architecture

  

  Source Turning coffee waste into 3D-printed architectural forms | 🔍AI analysis

Quantum Tech

• ⚛️🎯🔬Extra quantum dot boosts spin qubits to 99.72% fidelity

  

  Source One extra quantum dot helps balance speed and stability for spin qubits | 🔍AI analysis

• 🔬🧊📡100-channel detector paves way for broadband quantum networks

  

  Source 100-channel single-photon detector opens doors to quantum internet | 🔍AI analysis

• ⚛️🧭⚡Supercurrent enables nondestructive readout of Andreev qubits

  

• The supercurrent that spills its own secrets: reading Andreev qubits without the demolition

  Source Supercurrent enables the reading of quantum information without destroying it | 🔍AI analysis

Government Tech

• 🗳️🫆🛡️Biometric voting cuts fraud but tests public trust

  

  Source Biometric Voting Technology Cuts Fraud, Tests Trust - IEEE Spectrum | 🔍AI analysis

Health Tech

• 💉🧬🛡️ Hibsago: FDA approves first functional cure for hepatitis B

  

  Source Breakthrough hepatitis B drug offers ‘functional cure’ for first time | 🔍AI analysis

• 🛏️🌙📡 FlexiSense: smart bedding that keeps watch without a camera in the room

  

  Source Australian smart bedding flags risk of falls, gains attention in the US | EurekAlert! | 🔍AI analysis

Bio Tech

• 🔬🧬🛰️Open-source openDVP workflow makes spatial proteomics widely accessible

  

  Source Open-source workflow makes spatial proteomics easier to use | 🔍AI analysis

Food Tech

• 🥔🫘🥩 Spud Meat: Potato protein could unlock better gluten-free plant-based meat

  

  Source Potatoes could be the secret to better plant-based meat, say food scientists | 🔍AI analysis

• 🍄🥩🌡️ Blended meat:startups bet on curbing our beef appetite

  

  Source These startups hope “blended” food will cut our appetite for red meat | MIT Technology Review | 🔍AI analysis

• 🧊🌱📊 BPI Score: Tofu tops new holistic scoring system for protein sources

  

  Source Tofu is the winner. Scientist creates new holistic scoring system to measure the impact of food | 🔍AI analysis

Environment Tech

• 🌱♻️🪦 Terramation: NSW legalises human composting—here’s how it works

  

  Source What is ‘human composting’—and how does it work? | 🔍AI analysis

Climate Tech

• ☀️🧪✨New catalytic process extends perovskite solar cell lifespan

  

  Source Perovskite solar cells survive 1,000 hours at 185°F heat, 85% humidity | 🔍AI analysis

• 🌍📉 The London Register: CO₂ storage hits a record 43.6 million tonnes, but the accounting is doing some heavy lifting

  Source Global underground CO₂ storage hits record 43.6 million tonnes as London Register expands reach | EurekAlert! | 🔍AI analysis

Materials Science

• 🧺⚛️🧩 Diatomic Kagome: Organic molecules unlock new path to quantum materials

  

  Source Building quantum materials using supramolecular technology | EurekAlert! | 🔍AI analysis

Nanotech

• 🧲🔊💻 Magnon-polaron computing: Magnetic-sound waves could make computers 1,000 times more efficient

  

  Source Magnetic and sound waves could make computers 1,000 times more efficient | 🔍AI analysis

Deep science

• 🌌✨🖥️ Lumina simulation recreates the universe’s first 2 billion years

  

  Source Smithsonian scientists create most detailed simulation yet of the early universe | EurekAlert! | 🔍AI analysis

• ⏰💧🧊Physicists propose ‘time liquid’ that flows through time

  

  Physicists at the Max Planck Institute for the Physics of Complex Systems have proposed a new state of matter — a "time liquid" — where a time crystal's long-range temporal order dissolves but short-range order persists, much like ice melting into water. Their simulations of a 2D spin lattice driven by fluctuating thermal electromagnetic fields showed the spins ticking at half the driver's frequency, punctuated by defects called "time-solitons" that disrupt coherence without destroying the rhythm.

  Source Physicists Propose a Strange New State of Matter: A Liquid That Can Flow in Time : ScienceAlert | 🔍AI analysis

⏳ Zeitgeist

Once around the world outside tech…

Climate

• 🌡️📉🇺🇸 UN Emissions Gap Report: 1.5°C Is Gone, the US Adds Heat, and Renewables Offer Hope

  

  UNEP’s Emissions Gap Report 2026: The Point of Return, launched in Nairobi, officially concedes that 1.5°C cannot be held without overshoot. Current policies point to 2.8°C of warming by 2100, and full delivery of pledges only gets that to 2.6°C. The report blames the US Paris withdrawal for roughly 0.1°C of extra warming. US emissions rose 133 Mt CO₂e in 2025, about 70% of all national increases, which Michael Mann estimates could cost the global economy around US$1.2 trillion.
  
  The new best case is a 1.7–1.8°C peak, then a return to 1.5°C. Getting there needs emissions cut 18% by 2030 and 69% by 2040, plus about 320 Gt of CO₂ removal. Hope comes from renewables, which have grown three times faster than forecast, with solar growing seven times faster. Carbon Brief adds a sceptical note: the narrowing implementation gap reflects “policy changes, not policy progress”.

  Sources

  • New UNEP report lays out emission cuts needed for return to 1.5°C | UNEP - UN Environment Programme

  • UN projections say world is losing ground in fight to curb climate change, faulting US a bit

  • US is a big reason for the world losing ground in the fight against climate change, UN report says

  • UN says we’re going to blow past 1.5º warming, offers a “now what?” answer - Ars Technica

  • UNEP: Opportunity has ‘been lost’ to meet 1.5C warming limit without ‘overshoot’ - Carbon Brief

  🔍AI analysis

• 🌡️🔥🇪🇺 Killer summer: Climate change drove two-thirds of Europe’s 50,000 heat deaths

  

  Source Europe’s killer heat this summer claims 50,000 lives. Climate change blamed for two-thirds of deaths | 🔍AI analysis

• 💊🔬🌍 Techno-optimist scientists take less climate action, study finds

  

  Source Scientists who put faith in technology less likely to take climate action | Climate crisis | The Guardian | 🔍AI analysis

• 🌡️📊🔍 Proving Too Much: Flawed statistics behind new warming acceleration claims?

  

  Source How not to detect warming acceleration - by Zeke Hausfather | 🔍AI analysis

• 🧊🇨🇦📉 Climate Hushing: Canada quietly releases climate report warning of irreversible changes

  

  Source Canada Changing Climate Report 2026 Predicts End Of Glaciers By 2100 - CleanTechnica | 🔍AI analysis

• ⛽📉🌱 Pain at the pump: Fuel crisis drives NZ emissions below Covid lockdown levels

  

  Source Emissions fall as fuel prices deter motorists | RNZ | 🔍AI analysis

• 🌍📉 Climate change is reshaping the planet — literally: Melting ice caps are making the solid Earth rounder, faster

  

  Source Climate change is literally altering the shape of our planet | Live Science | 🔍AI analysis

Biodiversity

• 🦘🗺️🌿 Biodiversity-inclusive spatial planning (BISP): Australia urged to embed biodiversity in planning decisions

  

  Source Australia urged to put biodiversity at the heart of planning as global targets slip out of reach | 🔍AI analysis

• 🦋🌿📉 “Extinction” beats “biodiversity”: Plain language beats jargon

  

  Source Word choice matters: ‘Extinction of plants and animals’ raises more concern than ‘biodiversity loss’ | 🔍AI analysis

Pollution

• 🔥🫁🌫️ Gasping for air: Borneo burnsBorneo wildfires leave residents gasping for air

  

  Source Gasping for air: Borneo residents struggle as island burns | 🔍AI analysis

Health

• 🐦🛡️🧬Study finds many people already have bird flu immunity

  

  Source People may have preexisting immunity to bird flu | 🔍AI analysis

• 🧬🧪🏃 Autophagy Plot Twist: More cellular cleanup isn’t always better with age

  Source Researchers Discover a Surprising Twist in How Our Cells Age | 🔍AI analysis

• 🦟🧬🐄 Shamonda Virus: Livestock pathogen virus claims first human victim in France

  

  Source Shamonda Virus: France Confirms First Human Case After Insect-Borne Disease Kills Patient | 🔍AI analysis

Geopolitics

• Australia, 2026

  This interview with straight-talking Australian geopolitical commentator Jeff Rich (even if it is on the Russia-soft-power-friendly Neutrality Studies podcast) is well worth a listen to understand the changing tectonics of our Pacific region:

• 🌏🤝⛏️Australia urges middle powers to unite against geopolitical shocks

  

  Source Australia’s treasurer says middle powers should co-operate to avoid being hostage to geopolitics | 🔍AI analysis

• 🐟🛰️🔌 Grey-zone trawling: Russian trawlers caught mapping undersea cables over 500 times

  

  Source 17 Russian trawlers logged mapping undersea cables connecting the US and Europe — over 500 incidents of Russian vessels loitering over critical communications cables reported in under 5 years | Tom’s Hardware | 🔍AI analysis

• 🌊🚢🔌 Lawnmower Patterns: Chinese vessels loiter near Pacific undersea cables, report finds

  

  Source Chinese research vessels loiter near Pacific undersea cables more than any other nation, report finds — civilian fleet using ‘lawnmower’ patterns may be mapping seafloor for future sabotage | Tom’s Hardware | 🔍AI analysis

• 🛩️💾🇷🇺 Taxing Drones: Ukraine hits Russia’s Ministry of Finance data centre

  

  Source Ukrainian drone attacks hammer Russia’s finance ministry data center, damaged facility reportedly holds tax databases — facility also holds other important state registry data | Tom’s Hardware | 🔍AI analysis

• 🌍⚔️😴 Sleepwalking into World War III?

  

  The FT’s Gideon Rachman writes about his new book on why fading deterrence could spark a third world war

  Source Are we sleepwalking into World War III? | 🔍AI analysis

Democracy

• 🗳️📜💀 Declared Dead: India’s 130 Million Ghost Voters fight to restore voting rights

  

  Source Declared dead: Indian voters fight to get back on electoral rolls | Awani International | 🔍AI analysis

🧠Mind expanding

• 🔮📊🎲 One Million Forecasts Reveal Prediction Markets’ Real Limits

  

  Hari Seldon would be half-pleased. A Tel Aviv University team led by Michael Gilead analysed 1,014,122 resolved forecasts from Polymarket, Kalshi, Manifold and Metaculus across 12 domains over a decade, publishing in PNAS. The headline finding is that crowds are well calibrated in aggregate: events priced at p% happen roughly p% of the time, across horizons from hours to years. Zoom in, though, and systematic over- and underestimation appears in high-stakes areas such as international conflict and elections.
  

  Sources

  • Can we really predict the future? Over one million forecasts offer clues, not certainty

  • Calibration, accuracy, and domain variation across one million collective forecasts from prediction platforms | PNAS

  🔍AI analysis

• 🔮⏳📖 Too Much Future

  

  Futurist Tracey Follows has released a freely downloadable 47-page collection of five essays arguing that we’ve entered an era of “Big Futurity” — an industrial-scale production of forecasts, scenarios and models by governments, corporations, consultancies and AI that has become the very environment in which we understand what’s possible. Her central thesis: we’re “overfed on other people’s futures, starved of our own,” and every future is written between the system and the self.

🆕 AI releases

Just links and headlines from here… so much going on! Read fuller analysis in Sensorium Reader.

• 🌙🔓🧠 Kimi K3.1 Pro Brings Open Weights Within Three Points of the Frontier

  Sources

  • Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence

  • Introducing Kimi K3.1 Pro

  • Kimi K3.1 Pro launches at 4th on the Intelligence Index and will lead open weights models | Artificial Analysis

  • Moonshot AI launches new flagship Kim K3.1 Pro model with 3T parameters - SiliconANGLE

  • Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks | Reuters

  • Kimi K3.1 Pro Scores 4th of 215 AI Models; Moonshot Scrubs Cyberattack Data Before Open Weights

  🔍AI analysis

• 🎙️🤖🇪🇺 Mistral launches Voxtral 3 with a voice agent SDK

  Sources

  • Mistral launches Voxtral 3 with a voice agent SDK

  • Mistral Voxtral 3 Cuts Full Voice Agent Pipeline to One SDK at $0.04/min, Beats Whisper

  🔍AI analysis

• oMLX speeds up local LLM inference on Mac

  Source oMLX — LLM inference, optimized for your Mac | 🔍AI analysis

• Recursive’s Flint delivers fast agentic coding on local hardware

  Source Recursive releases Flint, 40B model for local coding | 🔍AI analysis

• 🤖📬🔐 Anthropic Launches Claude Duet, a Proactive AI Agent for Professionals

  Sources

  • Introducing Claude Duet | Claude by Anthropic

  • Anthropic unveils Claude Duet AI agent app | Reuters

  🔍AI analysis

• 💸🧠⚡ GPT-6.1 Terra Delivers Near-Frontier Smarts at a Budget Price

  Sources

  • Introducing GPT-6.1 Terra | OpenAI

  • GPT-6.1 Terra pushes the cost efficiency frontier | Artificial Analysis

  • OpenAI’s GPT-6.1 Terra matches GPT-6 Astra at 10% of the token cost

  🔍AI analysis

• OpenAI challenges Jev with widely available Decisions API

  Source OpenAI makes its Jev competitor widely available as decision model competition heats up | TechCrunch | 🔍AI analysis

• Google DeepMind’s Genie 4 world model adds speech, sound

  Source DeepMind unveils Genie 4 world model promising more realistic physics and humans - Ars Technica | 🔍AI analysis

• Agentic AI Foundation gains open-source tools from OpenAI, Anthropic

  Source Agentic AI Foundation adds four more open-source AI tools - SiliconANGLE | 🔍AI analysis

• Cadence unveils Nvidia-backed AI models for in-house chip design

  Source Cadence unveils suite of AI models that chipmakers can run in-house | Reuters | 🔍AI analysis

• iFLYTEK’s Spark Medical V3.7 hits 99.6 on MedBench 5.0

  Source iFLYTEK Healthcare Upgrades Spark Medical Model to V3.7 and Reports 99.6 on MedBench 5.0 - Pandaily | 🔍AI analysis

• AI slashes wildfire hazard mapping from months to hours

  Source Canadian Wildfire AI Slashes Hazard Mapping Time From Months to Hours: Open-Source Tool | 🔍AI analysis

• Opera’s WebAI standard runs AI models locally in browsers

  Source Opera introduces a new web standard to let users deploy AI locally in their browser | TechCrunch | 🔍AI analysis

• Sage-A1.5: 35B agent outperforms larger models through capability composition

  Source Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition | 🔍AI analysis

• Google’s Gemini 4 Carbon reaches early access testers

  Source Gemini 4 Carbon early test results are in ahead of release | 🔍AI analysis

• Gemini 4 Argon is powerful, but who needs it?

  Source Google’s Gemini 4 Argon just landed in the Gemini model picker — here’s what it actually means for you | TechRadar | 🔍AI analysis

• Microsoft AI diagnoses like a doctor, outperforms physicians

  Source Microsoft develops AI tool that can diagnose ‘sequentially’ like a doctor | Reuters | 🔍AI analysis

• Open-source llm-d 1.0 slashes AI inference costs 11.5×

  Source Open-Source llm-d 1.0 Hits Production: IBM Research Confirms 11.5× AI Inference Cost Drop | 🔍AI analysis

• Deepseek’s V4.1 Flash is fast, cheap, and hallucination-prone

  Source Deepseek’s latest AI model is smarter than ever, but its low running costs come at a price — uses only one-quarter of the HBM and one-eighth of SSD storage but still battles hallucinations in agentic workloads | Tom’s Hardware | 🔍AI analysis

• 🎙️👂🧠 Block’s Dessa learned to talk by listening alone

  Sources

  • Dessa - the Conversation Model trained only on speech

  • Block’s new voice model sounds eerily human | TechCrunch

  🔍AI analysis

🥼 AI research

The usual skim across the surface of what’s interesting and what’s happening… if only there was time to go deeper!

• How Fable 5.1 hides secret messages in images

  Source How does Fable 5.1 hide messages in images? — LessWrong | 🔍AI analysis

• A million AI agents buy speed, not intelligence

  Source Can a Million AI Agents Create Superintelligence | 🔍AI analysis

• Claude AI designs proteins that bind deadly scorpion toxins

  Source Claude’s 64-hour experiment finds 17 binders for scorpion toxins | 🔍AI analysis

• DeepMind watermarks AI-generated proteins to deter bioweapon threats

  Source Scientists can now ‘watermark’ AI-generated proteins to deter bioweapon creation. But will it really work? | Live Science | 🔍AI analysis

• CompanionSim reveals humanlike chatbot behaviors reduce trust, likability

  Source CompanionSim: Synthetic Data for Evaluating Anthropomorphism in Human-AI Relationships | 🔍AI analysis

• AI agents uncover brain disease leads in minutes

  Source Using thousands of AI agents to uncover new leads in biological data | OpenAI | 🔍AI analysis

• Frontier AI models’ robot control doubles every 54 days

  Source The Ability of Frontier Models to Control Robots is Rapidly Improving — LessWrong | 🔍AI analysis

• SuperMimic trains humanoid controllers with 1000× less compute

  Source SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model | 🔍AI analysis

• PSI’s AI system tackles five open quantum problems

  Source Five results on Open Quantum Problems · PSI | 🔍AI analysis

• AI digital twins uncover treatments for hard-to-treat breast cancer

  Source AI-based computer model offers new weapon against hard-to-treat cancers | EurekAlert! | 🔍AI analysis

• OpenAI swarm uncovers neurodegeneration drug lead in 32 minutes

  Source OpenAI Drug Discovery Swarm Decoded 7.7 Million Brain Cells in 32 Minutes | 🔍AI analysis

• NYU’s AI tools catch citation hallucinations, find research protocols

  Source Computer scientists build innovative tech tools to find experiments, detect citation “hallucinations” | EurekAlert! | 🔍AI analysis

• UCLA war game reveals how shared boundaries foster peace

  Source Breaking through the ‘fog of war’ to find peace | EurekAlert! | 🔍AI analysis

• Claude formally verifies 123 open-source libraries, uncovers hidden bugs

  Source Claude finds bugs that survived trillions of software test runs | 🔍AI analysis