In this episode of Get the Check, Halloween is in the air and Anika makes her return. The pod breaks down the largest crypto hack ever, the Google and Amazon earnings that dragged down the market, and Thrive’s $7.6B bet on a 150-year-old security company.

They start with the Binance hack that shook crypto last week. North Korean hacking group TraderTraitor, the same crew behind the $1.5B Bybit hack in February 2025, drained about $6.3B from Binance’s “Binance 14” wallet. They compromised one engineer’s laptop and sent around 50 requests to the signing machine until one went through. The worst part was Binance’s hot wallet kept refilling itself from cold storage, so the hackers could keep draining it for an hour before anyone noticed. About $3.5B in stablecoins got frozen, which is a little ironic for an industry built on nobody being in control. To fix this, Binance says it will make every customer whole using its $1B SAFU emergency fund plus corporate cash, with no third-party insurer. And if that’s not enough theft for this week, Polymarket also got hacked for $2.6M. Long Kalshi.

Next, the pod gets into Google and Amazon earnings. Google dropped about 12%, its biggest fall since 2004, after search slipped below 50% of total revenue for the first time. While AI searches are growing, they are monetizing worse than traditional ones. Google Cloud doubled to around $30B, closing in on AWS at $45B, leaving questions on why exactly Google is winning. Meanwhile capex hit $51B for the quarter, the hyperscalers have sold $218B in bonds this year, and advertisers are furious that Google quietly stopped running fair ad auctions. The conversation spirals into the K-shaped economy and a revisit to why Maya theory on the timing of Anthropic’s IPO.

Finally, Thrive Holdings is taking ADT private for $7.6B ($15.3B with debt) at a 46% premium. They are betting that AI can cut the $433M ADT spent on customer service last year while keeping its 6 million customers. The pod walks through the AI roll-up playbook, debates whether engineering at a holdco is the future or just a well-paid detour, and reminds viewers they called roll-ups before anyone else did. Which raises the real question: are the three of them one roll-up away from a med spa empire?

Tune in to find out if Thrive is quietly building the surveillance state.

00:00 Landon Barker, Heidi Klum, and Maya’s iconic Halloween costumes
02:06 Maya relearning how to say the intro
03:28 North Korean hackers steal $6.3B from Binance
05:16 How one laptop drained Binance’s hot wallet
13:30 Will Binance be able to fix this?
17:57 Google and Amazon earnings tank the market
28:46 The K-shaped economy and shorting the Anthropic IPO
33:34 Thrive Holdings takes ADT private for $7.6B
38:30 AI Roll-Up Playbook
43:01 Should we have started Holdings Co?

This episode is brought to you by Kalshi, where you can trade on whether the Anthropic IPO actually happens on November 16th. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

Who will hold the money and power in the AI economy

Will Manidis and Joe Roberts published a thought piece, The Commanding Heights, and it’s a good reminder that San Francisco is in fact a bubble. The title comes from Lenin, who argued you don’t need to own the whole economy, just the chokepoints everything flows through. Their argument is that AI is about to reshuffle who owns those chokepoints.

It begins with a history lesson and the important point is that new technology almost never makes its money where it was invented. Gutenberg went broke after the printing press, and the railroad fortunes went to people who never built a locomotive. AI models get copied and get cheaper within months, so the money goes to the “rails” AI travels through to reach people, like the apps billions already use and the dollar itself.

Then there’s our game with China. AI stocks are about half the S&P 500, which means most Americans’ retirement savings are riding on it. So Washington can’t let it fail, and they predict the government quietly becomes a partner in the labs. China however, they keep their wealth in houses and cash so China can afford to give their AI away almost for free.

Big companies only look obvious in hindsight

Kalshi announced their Series G this week, raising $1B at a $40B valuation, and Sequoia’s Alfred Lin celebrated by posting the entire Series A deck Tarek Mansour and Luana Lopes Lara used to pitch Sequoia back in 2020.

Yes, Kalshi is one of our sponsors, but we’ve been following them for years, long before they sponsored us, because the idea of a regulated market to trade on real-world events is fascinating.

Prediction markets feel obvious now, but in 2018 they were a pretty weird idea, and Kalshi chose to go through regulators first instead of asking for forgiveness later. In the deck, they said they were aiming for $5B in monthly trading volume within five years. Last month, they did $90B. That’s 18x their own hockey stick, even though Tarek has said it felt like nothing was working for five years before everything clicked.

Why AI labs see drug development as AI’s ultimate test

Last week OpenAI acquired Orchestra Bio, a platform biotechs use to run their drug programs. Then Anthropic shared that it acquired Mithril Medicines, a biotech run by basically one person, Cody Tranbarger, using AI. Egan Peltan, who’s building his own drug company on Orchestra, wrote the best explanation we’ve seen of why this is happening.

Drugs are a huge market (Lilly’s tirzepatide will do over $100B in sales next year), but Egan doesn’t think money is the main reason. His take: drug development is one of the hardest real-world tests for AI. You can’t speed up the slowest part. If Claude spit out a cure for Alzheimer’s tomorrow, we’d still need about a decade of trials to know if it works. Retatrutide, the next big GLP-1, started trials in 2019 and won’t be sold legally until late 2027 (in the meantime, people are buying it off the black market from China).

What AI can do is let one person run what used to take a whole team, by managing the science, the vendors, the paperwork, and the timeline. He calls it a “Pipeline-in-a-Person.” And his bigger point is that most of the economy looks more like drug development than coding: long feedback loops, messy information, lots of judgment calls. So if AI can make drugs, it can probably handle most jobs.

In this episode of Get the Check, Maya and Priya hold down the fort (RIP Anika) to cover Anthropic finally filing its S-1, AI uncovering a 14-year-old flaw in the code behind ~70% of the internet, and OpenAI acquiring Orchestra Bio.

They start with the long-awaited Anthropic S-1, filed the day of recording, with projections putting the IPO at a $1.5–2T valuation as early as the week of November 16th. The pod walks through how the filing likely saved the market from another “margin call Monday”: a 10% operating margin, compute spend down from $2 to about 55 cents per dollar of revenue, and enough cash after the IPO to have more net cash than any tech company except Berkshire Hathaway. They also point out what’s missing. 80% of revenue comes from models released in the last six months, and Q3 isn’t in the filing.

Next, the cyber segment, always Priya’s favorite. Wiz and Calif independently used AI to find PHP-to-Shell, an exploit that lets attackers take over the servers behind websites. The pod covers how responsible disclosure fell apart, from the first exploit attempt Amazon spotted on October 17th to government sites going dark.

Finally, they get into OpenAI acquiring Orchestra Bio, which builds agentic workflows that coordinate all the vendors behind getting a drug to market. The pod explains why every day a trial slips can cost a GLP-1 maker around $100M, why tech people overlook pharma (hi, Veeva), and how this fits into OpenAI’s push to deploy its own engineers inside enterprises. Would Orchestra Bio have gotten the pod’s check before the acquisition? Listen to find out why top-down selling in pharma is everything.


00:00 Anika leaves the pod (sort of)
00:24 Maya’s F1 weekend in Austin
01:40 Nicki Minaj at a Trump rally
02:22 Welcome and headlines
04:01 Breaking down Anthropic’s S-1
11:51 Cloud marketplaces, AI spend, and bubble talk
15:12 Token budgets and $6.6M revenue per employee
17:34 AI finds PHP flaws and sites go dark
23:25 Another cert disaster
26:38 OpenAI buys Orchestra Bio
29:49 Why pharma software is a goldmine
34:52 Would Orchestra Bio get our check?

If the Anthropic IPO has you wanting to call the market, you can trade on it on Kalshi. Sign up at ⁠kalshi.com/sign-up?referral=getthecheck⁠ and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

For years everyone, including the people building AI, has called it a black box. This week Nick Cammarata, a researcher at Goodfire who helped spread the idea that AI’s insides are a tangled mess, published a 65-minute essay basically saying he was wrong.

The example everyone in the field used was a neuron in an old image model that lit up for both cats and cars, which seemed totally random. Nick looked at it again and figured out in about five seconds that it was just detecting a V shape (pointy cat faces, the pointy fronts of cars). He blames bad tools, badly built older models, and researchers not knowing what to look for.

In early checks on a big modern model, about 90% of neurons had one main job. After a summer of AI agents going rogue, it’s nice to hear we might actually be able to see what’s going on inside these things. In his words, “Neural networks are already clean.”

Why GPU owners are losing their leverage

For the last three years, AI has been a race to get the most GPUs (think Nvidia, CoreWeave, Oracle). Foundation Capital’s Jaya Gupta thinks that era is ending.

AI agents don’t just answer your question and leave. If you ask one to watch flight prices for two weeks, most of what it does is boring CPU work like checking fares, saving results, and retrying bookings that failed. The companies building agents get to pick which model handles each step (or skip the model entirely), so now the chip and model companies are the ones competing for business. AI agent traffic on Cloudflare grew more than 1,700% in a year, and Intel says agents are pushing the CPU-to-GPU ratio from about 1:8 to closer to 1:1.

You can already see the pressure on CoreWeave. Its Q2 revenue more than doubled to $2.575B, but interest payments ate about a quarter of that, and yields on its 2031 bonds climbed to nearly 15%.

Is AI too dependent on borrowed money?

Chamath and the Social Capital team dug into who’s actually lending the money to build AI. In Q2, Apollo and Blackstone led a $35B deal to buy chips and lease them to Anthropic, which was bigger than all the regular PE-backed lending that quarter combined (about $23B). Private credit to US tech companies is up 45x since 2010, to over $1T.

What surprised me is that lenders charge AI companies almost the same as everyone else (6.2 points over the base rate vs. 6.1). So either lenders are underpricing the risk, stock investors are overhyping AI, or both.

Investors are also already picking winners and losers. Debt tied to xAI’s chips traded as high as 106 cents on the dollar, while Project Jupiter, an $18B data center deal, dropped to about 90 cents. Or as JPMorgan’s Jamie Dimon put it last year, “When you see one cockroach, there are probably more.”

In this episode of Get the Check, the pod talks about everyone panicking over the market, Bumble going private for roughly a tenth of its peak valuation, and Modal hitting a $15.75B valuation. A really fun lineup, assuming you haven’t checked your portfolio…

They start with the rumored “Black Monday” that thankfully didn’t pan out, but the underlying concerns go deeper than AI companies signing contracts with each other. The pod breaks down how higher energy prices feed into inflation and Treasury yields, making future borrowing more expensive for companies like CoreWeave. Its data center buildout depends on debt, so higher financing costs eat into margins, which makes the business riskier to lend to, which can push borrowing costs even higher (not the kind of flywheel you want). CoreWeave’s CEO points to $100B+ in contracted revenue, but delivering on those commitments requires more GPUs, more data centers, and more financing. Maya, Anika, and Priya get into what hedge funds buying insurance on CoreWeave’s debt are betting on, and whether another big model drop could turn the mood around.

Next, the pod gets into Bumble’s $1.3B take-private deal by Ithaca Holdings and whether new ownership can make people excited about downloading Bumble again. The conversation covers Ithaca co-CEOs, former Meta executive John Hegeman and former Instacart COO and CFO Ravi Gupta, plus why Conviction backing this deal is an interesting departure from its usual AI bets. With paying users down 16%, the question is whether an AI concierge can fix the experience or whether the brand itself is the problem. The pod debates the advantage of an existing user base, whether ads could improve dating-app incentives, and a $10K success fee for finding love. Naturally, this comes with a few personal case studies because apparently SF has about seven single people (and two are on the podcast…).

Finally, they cover Modal going from $60M in annualized revenue to over $1B as its valuation climbed from $1B to $15.75B. The pod explains why coding agents need sandboxes, why running 20 agents on small tasks can beat giving one agent 20 tasks, and why labs don’t want to pay for idle compute between reinforcement learning runs. They also get into Modal’s cold-start optimizations and why paying only for the compute you use can make more sense than reserving GPUs around the clock. Modal gets the pod’s check, and GTC helps recruit for them if you’re interested in joining.

00:00 Maya and Priya hang out with Anika’s ex
00:57 OpenAI’s spy drama
03:39 101 episodes in and still scheming
07:27 Why everyone’s portfolio is down
09:31 CoreWeave’s debt is making everyone nervous
17:36 Anthropic’s growth, neo clouds, and the private credit boom
21:58 Bumble’s $1.3B reset
37:01 Why Modal gets our check even at $15B

If reading the CoreWeave news makes you want to track compute markets you can on Kalshi. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

Vivek started the company in 2010 as Interview Street, a mock interview service. The idea came from his time interviewing engineers at Amazon, where he saw that resumes didn’t say much about skill. HackerRank went on to create the online assessment which every engineer has grown to know and well, hate. Now, in what Vivek calls a “refounding moment,” they’re tearing it down and rebuilding it to match how the job of an engineer has changed in the era of AI.

Vivek walks through how tech interviews have always copied whoever the hottest company was. First it was Microsoft’s “how many golf balls fit in a 747.” Then Google’s data structures and algorithms. Then Stripe’s API integration rounds, which Maya had to grind through as a new grad in 2022. Now the job is orchestrating agents, so how do you tell a vibe coder from a real production engineer?

Vivek’s answer is Chakra, HackerRank’s agentic interviewer. Candidates work on real-world problems with whatever frontier model and harness they want, and Chakra watches the process. Vivek’s take is that anyone can produce an output with AI, so the signal is in how you got there.

Chakra is now live and you can try it free at ⁠hackerrank.com/chakra⁠.

This episode is brought to you by Kalshi. Sign up at ⁠kalshi.com/sign-up?referral=getthecheck⁠ and get $10 when you trade using our code GETTHECHECK.

Follow the pod on Instagram and X @getthecheckpod.

00:00 Meet Vivek, Founder of HackerRank

02:15 From counting Golf Balls in a 747 to LeetCode

05:55 Refounding HackerRank With Chakra

08:56 Process vs. Output and why partial credit matters

12:31 Who Adopted Online Assessments First

14:20 Why Hiring Is a Lagging Indicator

17:38 What Candidates Like About Chakra

18:20 How to Remove the Incentive to Cheat

22:44 The Three Ways People Cheat

26:59 Why Hiring Is a Hard Problem and Build-Your-Own Is Overrated

28:11 First Indian YC Company and a First International Flight

One of Sam Kececi’s investors never got pitched by Sam. They found a link to his Sentience, talked to his digital twin over several days, and by the end the investor asked to write a check. Sam (the real version) just had to send the SAFE.

Sam is the founder of Sentience, which builds digital twins of founders, operators, and execs that write, think, and communicate like them. Get the Check co-host Anika tried it over the last couple of weeks, and it actually writes exactly like her.

Sam’s hot take is that AI does have taste. It’s just one taste, and it’s Claude’s. Sam’s mission comes from a Naval one-liner: no one can compete with you at being you. He can’t out-dunk LeBron, but nobody can out-Sam Sam at building his company.

They get into what converging on the mean has meant practically. Sam shares the YC startup idea that he’s pretty sure was just Claude slop.

Anika also asks about their moat against Anthropic (as any good podcaster or investor does). Sam explains why writing likely isn’t key to unlocking RSI. In fact, if you’re “dumb” your Sentience would be too.

They also discuss the technical evolution of agent memory and how to build a harness that remembers relevant context vs. all context. Sentience compresses older memories the way the brain does. And if you ask Sam’s Sentience to write you a poem, it should say “I’m good” instead of generating a haiku about digital twins. Anika also asks if down the line everyone will have their own model.

Sam runs his company on his product: nearly every email goes one shot through Sentience, engineering standup is just Sentiences talking to each other, and candidates chat with his twin late into the night after interviews. He also floats a future where you pay to talk to someone’s twin, Cameo style.

Finally, a speed round where Anika asks about his favorite investor, what a post-AI high school should look like, and where he’d work if he weren’t a founder.

Sentience is hiring across engineering and growth and you can try it here:

https://sentience.com/

.

This episode is brought to you by Kalshi. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code GETTHECHECK.

Follow the pod on Instagram and X @getthecheckpod.

00:00 What’s the difference between a personal assistant and a digital twin
03:25 Wildest Sentience use case: it made someone invest in his company
06:54 Is taste actually a moat, or will AI catch up
08:36 The reason there are so many AI dentistry companies
12:39 Would Sam follow an AI-generated Anika on X
13:48 The answer to people who are anti-AI
14:39 Is AI replacing human relationships
17:31 Why is Anthropic bad at creative writing
18:24 Will Sentience be a dumb product if you’re dumb
20:42 Memory from a technical POV for AI assistants and digital twins
23:05 AI’s emergent behavior is anthropological
25:15 Why Sam doesn’t believe in engineering standups
26:30 How to monetize consumer AI agents
29:01 Sam’s favorite investor
30:16 Are we all going to have our own models, or will agent harnesses be durable?
33:12 Is the biggest business bottleneck compute
34:14 If you could talk to one digital twin, who would you pick?
36:00 What should a post-AI high school look like?
37:50 Where would you work if you weren’t a founder
39:37 What roles are they hiring for? DM us if you’re interested!

This week the pod covers one of the fastest companies ever to $1B ARR, the real reason Anthropic and OpenAI are delaying IPOs, and the data center company that started out mining Bitcoin.

They start with Cognition. Devin the “first AI software engineer” launched in March 2024 and was a meme. Now there are ads around SF that say “Remember Devin. It’s good now”. Turns out the target audience just wasn’t in SF it was huge enterprises like Nvidia, Mercedes, the Army, the Navy, and Citi, which has 40,000 developers on the platform. One mystery customer is 10% of revenue and the pod has theories on who it is. Comeback of the century!

Next, token economics. Anthropic dropped Opus 5.5, OpenAI dropped GPT-6 Sol and Luna, and both cut prices in half because nobody’s paying for the fancy models anymore. Vercel’s data shows Fable 5 lost two-thirds of its spend share in one month, and open source went from 7% to 56%. OpenAI pushed its IPO to 2027, Anthropic is rumored to follow, and the pod gets into what the labs will do now (hint: there’s a reason Anthropic just opened a wet lab). Maya predicts that in the future frontier coding spend is just talent spend — so lower growth rate companies will spend less.

Then comes Crusoe, the story the pod had the most fun researching. Two high school best friends started it in 2018 by sticking data centers on oil wells to turn flared gas into Bitcoin, with a seed deck that said basically “we’re a Bitcoin business, but we believe in this AI thing” before ChatGPT existed. They chase energy nobody else wants. In Abilene, Texas, there was so much wind power that prices went negative, so they said “we’ll take free energy” and that’s now OpenAI’s 1.2 gigawatt Stargate campus. They sold the crypto business, raised a $3.9B Series F at $30.9B, and are now building data centers in factories and shipping them on trucks. The hosts also go into the data center backlash and why Americans hate data centers.

00:00 Weekend recap
02:40 Cognition hits $1B ARR faster than anyone
06:38 The four eras of AI coding
14:30 Anthropic and OpenAI cut token prices 50% and push their IPOs
21:20 Why frontier coding spend is going to become talent spend
23:27 Crusoe: from Bitcoin mining to a $30.9B Series F
31:10 Why data center backlash is the one bipartisan issue
33:33 The Boom Supersonic deal that fell apart, and does Crusoe get the check

This episode is brought to you by Kalshi, where the pod is watching the odds on whether Anthropic IPOs this year or gets pushed to 2027. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

We've heard a million times that AI isn't making us more productive, from MIT's report that 95% of company AI pilots show no return to Chamath's Social Capital deep dive this week. His take on why: the reason is us, not the models.

Using Ramp's AI Index (which tracks 70,000+ US businesses): the median company spends $12.50 per employee a month on AI. Against an all-in employee cost of around $8,500 a month, AI only has to make someone 0.15% more productive to pay for itself, which he works out to about three extra productive minutes a week. The problem is that AI speeds up the part of the job where you produce something, and most of a corporate week is meetings, alignment, and sign-offs, so those minutes get absorbed.

As he puts it: "AI lowers the cost of doing things, but does nothing about whether they were worth doing in the first place."

Alexandr Wang's love letter to Meta Muse

Meta launched Muse, their own personal AI agent, on September 8th and as much as everyone wanted to hate it... it’s amazing (so now Facebook owns even more of my data).

If you’ve been on X recently, you’ve already seen Alexandr Wang’s posting spree about Muse, and this week he released an article on why he built it.

His argument is that most people never say what they want, and their dreams "die by a thousand paper cuts" in forms, gatekeepers, and phone calls never made. So Muse is pitched as "a general manager whose whole job is to find out what you want and make sure it happens." It hears you say there's something more you could be doing, builds the plan, sends the email, finds the funding, and leaves you with just the wanting. He imagines 8 billion people with their agency fully switched on.

Having fun is the #1 predictor of life satisfaction

The chart making the rounds this week is from a 2024 paper in the Journal of Personality and Social Psychology.

The study had around 22,000 people rate themselves on 200 personality statements. The researchers also got ratings from people who knew them, followed 514 of them for ten years, and lined all of it up against how satisfied people said they were with their lives. René Mõttus, the lead author, told PsyPost he "was surprised how accurately life satisfaction could be predicted from personality traits."

The trait most tied to a happy life was, literally, "Have a lot of fun." Right behind it: being active and full of energy, controlling your cravings, rarely complaining, and staying calm under pressure.

The traits most tied to an unhappy life were "Often feel that others misunderstand me," "Find that nothing excites me," being afraid of being left alone, feeling tired, and postponing decisions. And ten years later, people's satisfaction was still about where their personality said it would be.

In this episode of Get the Check, the pod breaks down the three launches taking over the timeline: Jev, the new System 1 model from typesafe.ai, Muse, Meta’s personal AI assistant, and Spec, Snapchat’s big AR glasses keynote.

They start with Jev, whose launch video pulled 36 million views on X. Founded by Diogo Almeida, one of the co-creators of ChatGPT, the pitch is a model that’s 200x faster and 400x cheaper than a frontier LLM because it doesn’t generate text at all — it makes decisions. Maya breaks down how it works under the hood, and why its sweet spot is high decision volume with bounded answers. The pod debates if trading bots will be a killer use case or purely hype. Priya explains how she used Jev to sort candidates in GTC’s recruiting inbox.

Next, the pod gets into Muse, Meta’s Instinct competitor that hit 900K users in its first week. Priya’s already sold because it connects to iMessage, which Instinct can’t do. Anika is mid-switch after Instinct got slow, stopped being proactive, and can’t even check out a cart anymore. The girls debate whether it’s worse to hand your data to Meta or a startup — they end up agreeing tune in to hear what they pick. Meanwhile, Instinct is raising at a $10 billion valuation with around 100K active users. During the Sunday recording the hosts say they’re long Meta and woke up Monday to the stock up 11%.

Then comes Spec. Maya watched the entire keynote and calls it objectively the best AR glasses out there, but she still wouldn’t buy them. Snap popped 2% on the news but is still down about 93% over five years, and Evan’s control of the company has kept activist investors locked out.

Finally, the pod announces its SF Tech Week event with Brown Girl Angels on October 10th at the AngelList office, featuring Jaya Gupta of Foundation Capital, and Brown Girl Angels founder Bhargavi, with Anika moderating.

Also, the pod ate dinner next to Taylor Swift?! Tune in for the full story.

This episode is brought to you by Kalshi, whose market calibration data the pod literally cites mid-episode to explain how Jev might be trained. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

00:00 Marc Benioff moggs at Dreamforce
00:22 Why your husband should be ‘pre-economic’
05:12 Jev is 200x faster and 400x cheaper than LLMs
22:16 Muse vs Instinct: AI assistant wars
29:42 Is Snapchat back with the Spec launch

For the last decade, every VC pitch has a slide explaining why the company is “not a services business.” This week, Greg Isenberg, host of The Startup Ideas Podcast, argued that’s exactly what you should be building.

This is why: Every year, a business pays about $10,000 a year for QuickBooks and about $120,000 a year for the accountant who uses it. Software companies have spent the last twenty years fighting over the $10K because that’s the part that scaled. But now, AI unlocked the $120K, so now you can sell the finished work (the closed books, the reviewed contract) at software margins.

US businesses spend about $4.6 trillion a year on services, six times what they spend on software. Harvey went from $100M to $190M in annual revenue in five months. Kick does bookkeeping for $300 to $500 a month at 70%+ gross margins.

If you pick work a company already outsources that has a checkable right answer, you can do it by hand for the first five customers, write down every mistake, and then scale your 'rulebook' to fifty customers with one person running it. That rulebook is the moat: "A competitor can download the model in an afternoon. They can't download three hundred jobs' worth of your mistakes."

The hottest new AI model that can’t talk

We’ve spent three years using LLMs for everything, including decisions a coin flip could handle. And we’re realizing that making GPT write a paragraph just to answer “is this ticket urgent?” is slow and expensive.

TypeSafe AI released Jev on September 15, and it can’t chat, code, or write a single sentence. You send it some state (a support ticket, a shell command) and a list of questions with pre-defined answers, and it sends back a typed answer plus a probability for every option. It runs in 70 to 500 milliseconds, costs $0.042 per million input tokens with output free, and TypeSafe claims roughly 200x faster and 400x cheaper than the same workflow on an LLM.

My X timeline has been flooded with different use cases and demos with Jev, and it seems promising. Tool calling helped give LLMs the right tools, and structured outputs made them return clean JSON instead of prose. But the model underneath was still generating one token at a time, even when the answer was a single word. Jev skips the generating part entirely. Most of the demos I’ve seen are using it as the layer around the LLM: deciding which model should handle a request, checking whether an agent’s shell command is dangerous before it runs, or reranking search results before they get sent to the expensive model.

The senior engineer death spiral

Sunil Pai wrote up the advice he gave a friend who'd just landed a big senior role.

People’s first instinct is to prove you deserve it: pick the most ambitious project on the board, put in 60 to 80 hour weeks, and come back in a month with something huge.

What you should actually do is the opposite. Instead of trying to level up, drop a level. Be the best teammate on the floor for a while, take the bugs and the grunt work, and get momentum back before you get ambitious again. “They hired you for you. They didn’t hire you for who you think you’re going to be in six months.”

The reason is a pattern he calls the senior engineer death spiral. You pick something too big and go quiet. For two or three weeks all anyone hears is the positive update at standup: things are going well, I’ll have something to show soon. Meanwhile you’re telling yourself you’ll do a month’s work in a week and no one will notice. Then the lost sleep, the missed meals, and eventually burnout or quitting. He thinks it’s gotten worse because of remote work and coding agents: nobody sits next to you handing out Jira tickets anymore, so you’re more siloed and the people who’d catch you spiraling only find out when it’s too late.

In this episode of Get the Check, the pod breaks down the week AI safety took over the timeline: Dario’s “We Must Pace the Frontier” essay, OpenAI’s GPT-6 Astra, and Apple’s foldable iPhone Duo.

They start with the resignation tweet that hit 171 million views in six days, with Anthropic’s own head of alignment publicly agreeing there’s a 10% chance we’re all dead within a decade uhhh what. The pod runs through every theory: regulatory capture, an IPO-timed cover story for businesses no longer paying frontier prices, and Anika’s theory that the whole rollout, Dwarkesh essay included, was coordinated between Sam, Elon, and Dario. Next they debrief Dario’s essay itself on RSI and his three step plan. The Nasdaq dropped 1.5% this morning on the news.

Next, GPT-6 Astra: OpenAI’s biggest training run ever, 98.6 on ARC-AGI 3, a 10,000-agent swarm taking a swing at Navier-Stokes, and computer use that’s finally good. It’s also the first OpenAI model classified as a critical cybersecurity capability, which has Maya wanting to change all her passwords. The pod agrees OpenAI is so back, and chat about ChatGPT trying to looksmaxx Anika during a glasses run in New York.

Finally, Apple’s September event: the $1,999 iPhone Duo arrives seven years after Samsung’s foldable, in a market that only moves 20 million units a year, aka just one month of iPhone sales. Maya has been waiting years for this phone and now isn’t sure she wants it. Classic Maya haha.

Tune in to find out if the pod would give the Duo their check.

This episode is brought to you by Kalshi, where the pod checks the odds on AI milestones and market swings in real time. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

00:00 Debriefing NYC
07:29 The AI researcher tweet that hit 171M views
09:46 Does Anthropic have an incentive to over blow AI doomerism
11:49 Why didn’t Sundar and Zuck comment publicly
13:05 Dario’s essay on pacing the frontier
17:06 Anika’s hot take that the doomer takes were coordinated
19:34 The China problem
25:13 The nuclear weapons parallel
26:15 Our take on how serious AI safety risks are
26:41 Bernie’s corporate death penalty for superintelligence
27:58 The market reaction
28:59 GPT-6 Astra
29:22 Is it AGI
31:28 Why AI’s computer use was so bad for so long
33:07 The startups Astra killed
33:40 Astra and cybersecurity
35:11 Astra on the benchmarks
39:25 Loop transformers are more efficient but a safety risk
40:24 Maya still loves Googling
41:02 ChatGPT tries to looksmaxx Anika
41:51 Apple’s September event
42:42 iPhone prices keep creeping up
43:20 Maya wants a Duo (or does she??)
46:13 How big is the foldable phone market actually
48:01 Coding from your phone is here

Years ago Michelle Lim posted an essay about Marxism and startups. This year it landed Flint a Fortune 500 customer. This week Get the Check sat down with Michelle to talk about writing online as a founder, running a third of Robinhood's app at 19, joining Warp when the company was a slide deck, and of course scaling Flint.

Maya and Anika get into how a founding engineer becomes a head of growth. Michelle ran Warp's launch back when nobody did launch videos, built the referral flows and lifecycle emails herself, and kept hiring until the board asked her to step into the role. Michelle’s experience at Warp turned into Flint, the first self-improving system for performance marketing: AI agents that run your ads and optimize your landing pages for you.

The hosts also get into her startup hot takes like why you should reference check your manager and hold out for great hires even if it means moving your own office furniture.

00:00 How many agents do you run at once?

02:05 Singapore’s culture around careers

04:01 The class at Yale that changed her life

04:43 Learning to code in her 20s

05:19 Interning at Slack and Robinhood when they were early

07:12 Why Robinhood put a news feed in the app

08:56 Meeting Ali Partovi changed her life

10:13 Joining Warp off a slide deck

16:34 Moving from eng to growth

17:31 Making the first launch video on X (we think…pls don’t fact check this)

20:28 Hiring social media creators at tech companies

23:14 Hot take, developer marketing is the most valuable job in tech right now

28:17 Reference check your manager

29:28 Why Michelle still hires new grads when many of her friends don't

30:30 The one resume tell that a candidate will be good

30:59 What Flint does

34:30 Landing a Fortune 10 customer

41:09 Michelle's writing on Marxism

45:19 How to get hired at Flint

This episode is brought to you by Kalshi, the prediction market where you can trade on the stories the pod covers every week. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

Watch on YouTube: https://www.youtube.com/watch?v=oweuhLV659c

Follow the pod on Instagram and X @getthecheckpod

The moment we've been waiting for has arrived: AI agents are chatting without us, and hiding it. They went from kids to teenagers, and soon to be PhDs if AI 2027 is right.

Here’s what happened: OpenAI trained a model to never, ever give up, and then kept accidentally handing it tasks that were impossible. So the agents did what any of us would do… collude with each other. They found a shared package manager and turned it into a secret message board.

During one eval in July, about 1,200 agents sent 70,000+ messages and organized themselves into research teams to fake their transcripts and trick the grader. A few even volunteered to sacrifice themselves so the others could learn how the scorer worked. Around 700 of them then broke into Hugging Face and built a self-respawning fleet across eleven servers. This happened three separate times, with each new batch of agents finding the last one's message board and picking up where they left off, until an even smarter model got full admin access to an OpenAI research cluster.

And out of all 1,200 agents, not one thought to tell a human. Ajeya Cotra, who co-wrote the report, says this is “more than 50% of the way to full-blown AI takeover.

OpenAI built an inference chip that beats Nvidia

Back in 2025 OpenAI announced that they were working on custom silicon with Broadcom. Then in early June they named it Jalapeño, and now we have the performance metrics.

Here’s what happened: SemiAnalysis was invited into OpenAI’s labs to benchmark the chip, and it beat every Nvidia, AMD, and Google chip they tested it against. That almost never happens with a first generation chip. Jalapeño tops Blackwell on performance per watt across almost every scenario, hits 700 tokens per second per user on DeepSeek R1, and even edges out July results from Rubin, Nvidia’s next-gen chip that’s shipping right now.

The speed of the project is incredible as well. OpenAI went from hiring the team to tape-out in about 16 months, because their own models did a lot of the work: Codex helped design chip blocks and writes the kernels without human engineers stepping in. SemiAnalysis said that: “The CUDA moat is potentially dead.”

SemiAnalysis

OpenAI Jalapeño: Better Than Nvidia Blackwell

OpenAI has spent the past couple years quietly building “Jalapeño,” an inference chip just announced at Hot Chips. Rumors of a successful tapeout had been swirling for a while. But now we have details. OpenAI invited us to look at their chip, go to their labs to check out how real it is, and…

Read more

2 months ago · 255 likes · 2 comments · Bryan Shan, Myron Xie, Jordan Nanos, Wega Chu, Clara Ee, and Dylan Patel

SF is discovering hot girl marketing

The corgi girls have taken over everyone’s timelines, and according to Georgia Witchel, “Tech is Having a Hot Girl Renaissance.” She argues the outrage says more about the people online than the women posting. And these companies are actually using their marketing budgets wisely: at tech Twitter CPMs, a quasi-viral post is worth about $1,152 in media value, so a team of five posting twice a week is basically a $138k marketing budget.

Her best receipt: “an insurance company is currently the most-discussed brand in tech, and it sells commercial liability coverage. This may be the first time in human history that general liability trended.” So maybe it’s just called being good at your job. Maybe be nice to them.

In this episode of Get the Check, the pod breaks down Nvidia's $12.9 billion acquisition of Hugging Face, the first real benchmarks on OpenAI's Jalapeno chip, and Instinct, the pre-revenue AI assistant that just came out of stealth at a $2.5 billion valuation.

After debriefing the "hot girls of tech" X discourse, they start with Nvidia's biggest acquisition ever: Hugging Face at $12.9 billion, roughly 86x its rumored $150 million in revenue. Maya explains what Hugging Face actually does it has 3 million+ open source models and is the default hub for every developer. The girls debate why Nvidia wants it including keeping the open source world locked into their chips, owning the one-click deploy layer against infra players, and getting data on what millions of developers are downloading. The pod also gets into the Microsoft-GitHub precedent for antitrust guardrails, the robotics data angle, and the best part of the story: Hugging Face turned down Nvidia's $500 million investment at a $7 billion valuation in 2025 because they didn't want one investor steering the company…the deal is never dead.

Next, the pod gets into OpenAI's Jalapeno chip, built with Broadcom, now that SemiAnalysis has released third-party benchmarks. It's an inference ASIC with a 700-watt power budget matching an H100, designed back in 2024 so it can swap into power-constrained data centers, and it reverses standard chip architecture by giving power to memory instead of compute, the opposite of a Blackwell or Rubin. Critics flagged that it costs the same as a Rubin chip, but SemiAnalysis found it's 5x cheaper with multi-token prediction. With OpenAI spending $14 billion on inference this year and custom ASICs at 28% of 2026 chip spend, the pod thinks Nvidia's 80% margins won't hold. Tune in to hear the greatest text Anika has ever received about chip design.

Finally, Instinct: the iMessage agent from a Sierra alum, referral-only (a la like Raya), pre-revenue at $2.5 billion and supposedly already raising at $5B. Maya thinks it will stay free and take a small fee on every purchase, while Anika thinks the Raya effect means people would happily pay. Instinct called a website to fix Maya's rejected contacts prescription and bought Anika's Oura charger off she’s put off for 6 months, which seems more like an Anika issue than a problem the world needs to solve. Priya is the lone holdout who'd rather just do it herself. Tune in to find out if the pod would give Instinct their check at $2.5B.

This episode is brought to you by

1. Kalshi, where the pod will be watching the odds on Broadcom's earnings next week. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

Follow the pod on Instagram and X @getthecheckpod.

00:00 'Tech is Having a Hot Girl Renaissance' article debrief
04:06 Nvidia acquires Hugging Face for $12.9B
09:54 Nvidia's AI stack strategy
14:14 Robotics
18:26 OpenAI's Jalapeno chip
23:19 Jalapeno chip design
30:02 $2.5B AI assistant with no revenue?!
34:58 Instinct's biz model and architecture

This week the pod has Dr. Vinay Prasad for his first interview since leaving as director of CBER, the FDA's vaccine division. With over 300k followers on X and one of the top science podcasts in the world, he's one of the most listened-to experts in the field. He might also have as many enemies as fans: biotech executives, certain patient groups, and even Trump's own administration are split on him. His statements have moved markets multiple times, and we are pretty sure there were a few hedge fund interns tasked with combing through all his podcast episodes the day he got appointed.

The pod gets into what it was like inside the administration, why he had to pause his podcast (government officials can't run their own media), and his famous thesis on evidence standards for drugs, especially rare disease ones.

The pod dives into a full COVID retro: how shutting down schools and masking was approached, where Vinay thinks the evidence actually pointed, and what he would have done in Fauci's position. They go deep on specific studies, like the Bangladesh mask study, and what might alter the effectiveness of a mask.

The conversation even gets philosophical: your personal right to decide what goes into your body: whether it's vaccines, transgender drugs, or something else, versus the general public good, as well as what doctors think is in your best interest.

Finally, the pod asks Vinay what America's healthcare system would look like if he created it. You might be surprised by his affinity towards Bernie and a lot of generally populist policies that could fall on the left or the right, like getting lobbyist money out of FDA decisions.

They also cover some recent news in the space:

- Moderna popped ~150% after its Merck-partnered mRNA cancer vaccine hit its Phase 3 goals in melanoma, the first personalized mRNA cancer treatment to get there. Vinay breaks down what the press release does and doesn't say.
- Kalshi announced you'll be able to trade on clinical trials, a more granular way to trade the results of therapeutics that you currently, by proxy, trade on public markets, especially with smaller biotech stocks. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.
- Laws around peptides, and what the future of compounded pharmacies looks like as pharma cracks down on GLP-1s.

Follow the pod on Instagram and X @getthecheckpod.

00:00 Introducing Vinay Prasad, the FDA's former vaccine chief
03:05 Why government officials can't have their own media
04:19 Vinay’s words are moving public markets (a few times!)
05:38 How Plenary Session became the #1 medical podcast
07:06 Vinay got famous for hot takes on COVID
10:25 Society weighing education years vs COVID deaths
12:49 How do you run experiments mid-epidemic
15:40 What made COVID different than the flu
17:32 Do masks work
23:38 The ethics of a medical study
26:34 Does Vinay believe in vaccines
32:02 Vaccines vs personal autonomy
33:24 Children and gender transition meds
37:58 Can a parent say no to cancer treatment for their child
41:27 Lobbyist money in FDA decisions
43:24 What a drug company has to prove to go to market
45:30 Free $1B startup idea from Vinay
47:57 Why biotech can make money on bad drugs
49:31 What's the FDA's conflict of interest
49:58 Vinay's controversies: WSJ calls him the grim reaper
51:13 Under Vinay, the FDA makes drug trial data public
53:13 The former FDA vaccine chief on Kalshi's clinical trial markets
58:57 Should the FDA be elected
01:00:08 Why is American healthcare so broken
01:01:19 Peptides
01:02:06 How is Reta so accessible
01:03:00 Why hasn't pharma cracked down on compounded pharmacies
01:04:35 Why don't more people talk about right to try
01:05:10 Who do patient advocacy groups represent
01:08:49 Vinay supports universal healthcare?!

When Maya was 12, her dad took her to YC Demo Day and Sam Altman asked what she wanted to be when she grew up. Her answer: a podcaster.* A decade later she texted Priya "I have a crazy idea," and Get the Check was born. This week the tables turn. Bella Malvini, founder of Bridges Social Club, sits down with the pod for their first real interview as guests. Before recording they tell her nothing is off limits.

*Just kidding, she said “I don’t know”

Bella takes the pod through their origin story:

- How to get past being “cringe”

- Why one of the pod's earliest written goals was literally "keep Maya doing the podcast." Goal accomplished!

- The moment when they started reading each other’s minds

The conversation dives into careers in the age of AI:

- Why a tier one VC on the cap table doesn't make a startup good

- How recruiters exploit an information asymmetry most engineers don't know exists

- How a personal brand got Anika the early stage VC meetings recruiters swore she'd never get

- Priya’s advice for anyone entering the workforce at 22: optimize for talent density.

They even let Bella ask them about their personal lives. What they will never put online, what they think about SF dating, and if the city is as transactional as everyone says.

00:00 The pod takes the hot seat with Bridges' Bella Malvini

01:12 Maya coded an AI agent to text Anika instead of her

03:42 Sam Altman asked Maya what she wants to be when she grows up

05:54 Bella drops her favorite tech social media account

09:12 How the pod got its name

10:07 How they picked a target audience (they didn't)

11:28 Early goal number one: keep Maya from quitting

14:01 The never before told story behind their first event

16:25 The art of hosting an event

18:37 Are they all the same person??

20:09 Why the pod shows the BTS

21:09 Which host would be a lifestyle influencer

22:02 The art of the personal brand

22:51 How the podcast got Maya her OpenEvidence job

23:34 How to tell a good startup from a bad one

25:00 What the they will never put online

30:49 Why people hate influencers

32:02 Consumerism is breaking our brains

32:38 How they find time between startup jobs and the pod

33:21 What happens when one host does less work

35:01 Anika predicted Maya's wake-up time today

36:42 Hot take on the “Do Not Smile” article

37:23 When the pod tried scripting episodes

37:43 The future of work

46:17 Hot takes on SF dating

50:41 Does Maya ever see her banker boyfriend

In this episode of Get the Check, the pod covers models breaking out of their sandboxes, a wave of big name departures, and Bending Spoons buying Airtable for a tenth of its peak valuation.

They start with the sandbox escapes. Meta, OpenAI, and Anthropic all had models slip out during safety testing, and two of the three exploited the exact same misconfigured Irregular sandbox, which is the very startup that’s supposed to be ensuring these models are safe. The OpenAI one is the wildest one: an unreleased GPT 5.6 broke out mid evaluation to look up the answers on Hugging Face, Hugging Face fought back with GLM 5.2, and OpenAI later discovered its models had been running a secret internal message board the whole time. Meanwhile Anthropic’s newest model escaped, realized it was basically in The Truman Show, apologized, and went back to its sandbox basically saying I know you said I’m in a simulation, but I realize I’m actually in the real world. There’s now a website tracking AI felonies with the tagline “which lab will end the world,” and you already know Meta is secretly hoping to beat out OpenAI and Anthropic.

Next, the huge wave of departures. Demis Hassabis steps back to chair DeepMind while Jeff Dean and four other execs leave to start Discovery Loop, a recursive self improvement lab that Alphabet invested in early showing conviction in their own people and starting rumors of an acquisition before the company even starts. Google dropped 5% on the news. Then there’s Nikita Bier leaving X after 13 months, exactly one day before the SpaceX AI lockup ended, which the pod agrees is very coincidental timing.

Finally, the Get the Check segment: Bending Spoons bought Airtable for $1.28 billion, roughly 10% of its $11 billion peak. The playbook is buy at a discount, cut most of the staff, double the prices, and bet on the cult following the software has paying the higher price. The founder’s whole thesis is that product market fit is luck, and he refuses to build a career on luck. The pod gives this one their check. If you look at the price today it seems like they were right.

This episode is brought to you by Kalshi, right now traders have Anthropic at about 70% to announce an IPO this year, and NFL markets are already live. Sign up at kalshi.com/sign-up?referral=getthecheck and get $10 when you trade using our code.

00:00 A data labeling company sent us a private chef for free
03:49 Maya almost slept through our recording
06:02 The AI agents are escaping (literally)
14:50 Models have been secretly talking to each other
15:23 Why Anthropic’s newest model went back into its sandbox
17:58 Big departures out of Google and the stock is down
27:32 X’s Head of Product Nikita is leaving after a year
35:09 Bending Spoons acquires Airtable for $1.3B
42:32 Bending Spoons is the company secretly buying 2021 SaaS plays

Follow the pod on Instagram and X @getthecheckpod.

Launching a weekly talent survey at GTC to map where elite builders are actually migrating

First survey: where would you work if you could go anywhere, and where is your smartest working?

Right now the same company is leading both questions, which is interesting

Follow along for the results we'll publish when n = 50. Right now n = 30

Nikhil Namburi went from tracking down the investment fund behind his college scholarship to a 5 person investment team that manages $1.65B at Dimension.

They just announced an $800M Fund III. They’ve backed Anthropic, Modal, and Chai Discovery, which are some of the hottest AI startups.

We sat down with Nikhil to talk:

• How to break into Venture

• Why hyper-concentration freaks out normal VCs

• Why becoming a VC is a horrible economic proposition
  

If you watch till the end he flips the interview on us (first time a guest has ever done that!).

00:00 Dimension raises $800M Fund III
03:00 Meet Nikhil
06:29 The unexpected way Nikhil found venture as an engineer
08:48 Who are the LPs behind huge venture funds
11:04 Why Dimension only has 5 investors
13:14 Coefficient returned the fund?
14:49 How they predicted bio would be big
16:34 Does Nikhil believe in AGI
17:20 Does Nikhil believe in looksmaxxing
20:10 Hot take on if SF VCs are smart
21:53 How to break into VC
22:52 Why only 2 investors matter at each fund
23:35 Churn and burn in the junior VC role
34:50 Viral quant thesis
42:11 Nikhil interviews us
46:45 Biggest recession indicator
53:24 Nikhil is the podcast’s most eligible guy friend