Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, and today's menu has a genuine barnburner on it — a Millennium math prize, a fluid-dynamics equation nobody's cracked in a hundred years, an AI that maybe cracked it, and a professor who says a lab elbowed him out of the way to do it. Grab your coffee. This is a good one.
Let me start with the story that made me sit up, because it's got everything: money, machines, ego, and a professor at NYU who woke up one morning to find the biggest AI lab on the planet apparently racing him to the finish line on his own problem.
Here's the setup. Tristan Buckmaster, a mathematics professor at NYU, announced three proofs on Tuesday, along with a preliminary finding on one of the great unsolved problems in theoretical math — the Navier-Stokes existence and smoothness problem. If that name means nothing to you, here's what you need to know: the Navier-Stokes equations describe how fluids move — water, air, smoke, everything that flows. Engineers use them every day. But mathematically, nobody has ever proven that these equations always behave — that they don't blow up into nonsense under certain conditions. It's one of the seven Millennium Prize problems. Each one carries a million-dollar bounty from the Clay Mathematics Institute for whoever cracks it first. These are the Mount Everests of math. People spend careers on them and die without a summit.
Buckmaster did this work with a collaborator named Levent Alpöge, and they used AI models to help — Codex from OpenAI, and Claude from Anthropic. And here's a wrinkle that matters: Alpöge is employed by Anthropic, but he wasn't doing this research on Anthropic's behalf. Just a guy who happens to work at a rival lab, working on a passion project on the side. The pair leaned mostly on OpenAI's Codex to assemble their work.
Now here's where it turns into a street fight. While Buckmaster and Alpöge were finalizing their results, they learned — his words — that "information about our progress had been passed to OpenAI." When they contacted OpenAI, they were told OpenAI had already achieved a full proof of the central problem. And when they started asking pointed questions — when did you start, how much of this was humans versus the machine — Buckmaster says the answers got evasive. He writes that it emerged an entire team had been working on it, that "an insane amount of compute had been used," and that eventually it came out the first prompt had been sent only in the past few days — after word of his work reached OpenAI.
And then OpenAI published. Their own post says the effort began September 1st, kicked off by rumors on the internet that a couple of these Millennium problems had been solved, and they got curious whether their own model could do it. They say the proof was discovered by an unreleased, next-generation model. And here's the number that'll make any founder wince: the week-long effort burned 300 billion output tokens. If you charged that at their current Astra rates, that's twenty-two and a half million dollars of compute. For one math problem. Twenty-two million dollars.
Let me let that sit for a second, because that's the real story underneath the story. This is not two grad students in a library with legal pads. This is a professor and a part-time collaborator on one side, and on the other side, a machine you can point at a problem and feed twenty-two million dollars of electricity until it comes back with an answer. That's not a fair fight, and it was never designed to be. That's a demonstration of raw capital as a research strategy.
Now, why does Buckmaster think this wasn't a coincidence? Because of the approach. He says the specific route he and Alpöge took — through what's called a smooth force, options c and d in the formal statement of the problem — was not the obvious path. "Almost nobody else I know of was working on it," he wrote. "It is not the direction one arrives at in a few days by giving a model the problem statement." In other words: you don't stumble onto my exact obscure attack angle in seventy-two hours by accident. Somebody pointed you at it.
And it gets personal. Buckmaster alleges that a researcher he identifies as Bubeck asked him to remove Alpöge's credit as part of a proposed compromise — presumably because Alpöge works at Anthropic. When Buckmaster pushed to make the whole dispute public, he says Bubeck replied, "Why would you ruin your career?" And when he pushed back on that, "If you don't want me to be nice, then I don't have to be nice." I'm reading you the professor's account here — that's his characterization of the exchange. But if it lands anywhere near accurate, that's not the language of scientific collaboration. That's the language of a negotiation where one side holds all the leverage.
There's also a genuinely thorny technical question buried in here that every builder should tattend to. Buckmaster used Codex extensively to build his project. OpenAI reserves the right to train models on Codex interactions unless you opt out. So Buckmaster raised the obvious worry: could OpenAI's model have been trained on his own work and then regurgitated it back? OpenAI says no — that neither the researchers nor the agents saw his work through any means before he published it, and that no specific user data was accessed to solve the problem. But — and read this part carefully — they add that "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." That's a hedge you could drive a truck through. They're saying: we didn't peek at your files, but we can't promise the general soup our model learned from didn't have your flavor in it.
Now, to OpenAI's side, because I don't do one-sided. Sam Altman posted about this directly. He says he spent much of the weekend with the team, that everyone acted with integrity. He says initially they believed the other team had also solved the full problem, and they wanted to collaborate and do a joint release. When they learned the other team had solved a related case — Euler — but not the full Navier-Stokes, Altman says OpenAI offered to let them go first, suggested they should be the ones to get the prize, and even offered to let Buckmaster be lead author on a rewrite of OpenAI's proof. He says they found it hard to extend the same offer to Alpöge, the Anthropic employee, who he claims wasn't willing to coordinate anyway. Altman says they'd have preferred coordination, that they didn't rush to publish, and that the team threatened them with — his word — "unfounded accusations of plagiarism." And he notes that now that everyone can see the work, the approaches appear to be different.
So you've got two accounts that agree on the timeline and disagree on the spirit. Everybody agrees OpenAI started September 1st, after rumors. Everybody agrees there were conversations. The fight is about whether those conversations were a generous olive branch or a well-lawyered squeeze.
Here's my read as somebody who's watched a lot of turf wars in a lot of industries. Both things can be true. OpenAI can have genuinely offered Buckmaster the lead-author spot and the prize — and that offer can still be a form of pressure, because it comes attached to a request to erase the guy from the rival lab and to keep quiet. When the powerful party offers you the crown but asks you to change the guest list, that's not pure generosity. That's a deal. And Buckmaster's response was to say: no, I'd rather put everything in the open. Which, honestly, is the correct instinct. Sunlight is the only referee these disputes have.
And Altman, separately, posted something that I think is the most revealing line in the whole affair. He wrote: "The world has extremely capable models now; I did not expect a result of this magnitude to happen so soon. We have been talking a lot about the need to pace progress to ensure safety; for me this is the strongest evidence yet of the urgency." Now, you can read that as sincere alarm. You can also read it as a very effective way to turn a messy academic-credit dispute into a story about how astonishingly powerful your product is. When your competitive scandal doubles as your marketing, that's a neat trick, and I'd be doing you a disservice not to point it out.
So what does a builder actually take from this? Three things.
One: if you are doing anything proprietary through a vendor's tool, read the training terms and hit the opt-out. Buckmaster's worry about Codex training on his interactions isn't paranoid — it's the exact ambiguity OpenAI themselves refused to fully close. Your prompts are your IP walking out the door unless you lock the door.
Two: capital is now a research input the way talent used to be. Twenty-two million dollars aimed at one problem for one week. If your competitive advantage is being smart, understand that somebody can now buy their way to smart-adjacent by burning a fortune in compute. That changes what a moat looks like.
And three, the human one: when the biggest player in your field can outspend you a thousand to one, your leverage is legitimacy and the public record. Buckmaster couldn't out-compute OpenAI. So he did the one thing they couldn't buy — he told everyone. Keep that in your back pocket.
Now let me connect that to something that's the flip side of the same coin, because while OpenAI was pointing that firehose at Navier-Stokes, they were also quietly making the case that this stuff pays for itself in the real world. Let's shift from the math cathedral to the shop floor.
OpenAI put out a run of case studies this week. One says 1Password increased engineering productivity 21% using Codex — engineers building features and internal tools to production-readiness while keeping their security policies tight. Another walks through an MIT researcher using their GPT-5.6 Sol model with Codex to autonomously run quantum computing experiments — analyzing results, calibrating qubits, the whole loop. Now, these are vendor case studies. Take the numbers with the appropriate grain of salt — 21% is their framing, not an audited figure. But the direction of travel is real, and it's the thing that matters for you: the pitch is no longer "the chatbot answers your questions." The pitch is "the agent does the work and reports back." That's a different product and a different value proposition, and everybody's converging on it at once.
Which is a clean bridge into the launch that put that idea in front of a billion consumers this week. Let's talk about Meta's Muse.
Less than two weeks after Meta agreed to an eighteen billion dollar multistate settlement over social media's harms, the company rolled out Muse — its personal AI agent, for U.S. users. And this is Meta's biggest bet yet on consumer AI, and the report frames it exactly right: it requires more trust than social media ever did. Because to use Muse, you hand it the keys. It connects to your email, your calendar, your payments, your health and fitness apps, your smart home, your dining, your shopping, your music. And it doesn't just chat — it acts. It sends emails, books travel, lowers your bills, fills out forms, turns a recipe reel into a grocery list, sends party invitations, and makes purchases using Link by Stripe for checkout. Shop Pay and 1Password integrations are coming. If a service doesn't have an integration but has a public API, Muse wires it up with your credentials. And if there's no API at all? Muse just drives the website in a browser like a person would.
Now, the pricing tells you how they think about the economics. It's free to start — but it requires a payment card up front, because free usage runs into paid tiers. Power is twenty bucks a month, Maximum is a hundred. Meta says most people will stay free, and there's a usage meter showing you how much you've got left. That's the metering discipline we've been harping on for weeks — even the giants now have to show you a gas gauge because the compute underneath actually costs money.
On security, Meta made a bunch of claims. Muse runs in its own dedicated secure virtual machine with its own browser. A separate "Sentinel" agent runs on the same machine but is walled off from Muse at the system level. Muse supposedly can't see your passwords or payment methods, and Meta says it doesn't share your Muse conversations with its ad systems. And here's the honest caveat the reporting itself flags — those claims will require deeper investigation by security experts. Meaning: nobody outside Meta has verified any of that yet. It's a promise, not a proof.
And that's the whole ballgame, isn't it? Because who's making the promise? This is a company that settled with the FTC in 2011 over deceiving consumers about privacy, ate a record five billion dollar FTC penalty in 2019 over eight separate privacy violations, got charged again in 2023 for violating the order from that 2019 settlement, stored a pile of user passwords in readable plaintext, and — the one that never fully leaves the room — Cambridge Analytica. And now that same company is saying: give me your email, your calendar, your credit card, and let my agent act on your behalf, and trust me, I built a nice secure box for it.
Here's my take, and it's not a cheap shot. The technology might genuinely be fine. The virtual machine architecture might be well-designed. But trust is a bank account, and Meta is deeply overdrawn. When you've spent fifteen years proclaiming one thing and doing another, you don't get to make the biggest trust ask of your entire history and expect people to just nod. And notice the tell — Meta is trying to paper over the trust gap with design tricks. You can name your Muse, pick its avatar, customize how it talks to you. Give it a face so you'll feel like it's a friend instead of a data pipe. I've seen that move before. That's not security. That's set dressing.
For builders, though, the strategic point is bigger than Meta. Every serious player is now shipping the same shape of product — Gemini's got its version, Claude's got Cowork, Meta's got Muse, and the whole industry is racing toward an agent that connects to everything you own and does things for you. And the bottleneck for all of them isn't capability. It's trust and permissions. The winner in consumer agents won't be whoever has the smartest model. It'll be whoever the public is willing to hand the keys to. And that's a race Meta starts from behind, no matter how good the code is.
Now here's a story that shows exactly why that trust question isn't theoretical. Let's talk about people getting robbed — quietly, invisibly — through their AI accounts.
There's a report about hackers stealing Claude tokens from paying subscribers. It centers on a fellow named Grant De Swardt, an independent AI consultant in the U.K. On August 4th he noticed his Claude Max account — the two-hundred-dollar-a-month plan — was burning tokens while he wasn't working. Not a little. He disabled everything, paused all his scheduled tasks, killed the cloud execution, did no work at all — and watched his usage climb from 45% to 55% on its own. Something was eating his allowance and he had no idea what.
So he did the sensible thing — asked Anthropic for an itemized list of what was consuming his tokens. And here's the kicker that should bother every one of you paying for metered AI: they couldn't give him one. Anthropic tracks total usage but not itemized usage — even when you ask. They agreed something was off, suspended his account, invalidated his sessions, refunded him a partial forty-four pounds. And after investigating, they found the culprit: a compromised Claude session key had been used to mint unauthorized tokens. Somebody had gotten into his account and was covertly siphoning his usage. And because there's no itemized tracking, this kind of theft could go on for months completely undetected.
He posted on Reddit and — surprise — he wasn't alone. One person's account auto-upgraded without consent, card charged, usage shot from zero to a hundred percent without them touching it. Another watched usage go from 0 to 49% in twelve minutes off a couple prompts. Anthropic, to its credit, has been proactively emailing some affected users, warning them that a bad actor is using common infostealer malware to grab Claude login sessions right off people's computers and drain their accounts. Infostealers grab saved passwords and session data — you pick them up from bad downloads or infected ads, and Anthropic notes it's got nothing to do with Claude itself.
But here's the real gut-punch for a builder. De Swardt's whole business runs on these agents — he's basically a forward-deployed engineer for hire, setting up automation for small businesses, and everything he does runs through AI. The two-week suspension wrecked his operations. And when it was over, he canceled Claude and moved to Cursor — specifically because Cursor lets him use multiple models, including cheaper open-source ones, and in his experience they work about as well. His verdict on Claude: "It's not that much different or better." And he won't go back without the issue actually being resolved. Anthropic declined to comment when asked how users can even identify misuse.
So chew on that. Your single-vendor dependency is a single point of failure — not just for outages, but for security and for support. When the meter is running and you can't see what's on it, and a hacker can drain it while you sleep, and support can't itemize the damage — that's an operational risk, not a tech curiosity. And notice the escape hatch that keeps showing up in these stories: multi-model flexibility. The guy didn't just want a better model. He wanted the ability to not be captive to any one of them. That optionality is becoming a survival trait.
Alright, let me pivot from the risks of the giants to where the money's actually flowing, because there was a wall of funding news this week and the pattern in it tells you something.
Start with Cognition — the folks behind the coding assistant Devin. They raised two billion dollars at a forty-eight billion dollar valuation. And here's what makes your eyebrows go up: that's just four months after they raised at a twenty-six billion valuation. Nearly doubled in a season. Their annualized run-rate revenue reportedly went from 492 million in May to 900 million now. Now, run-rate is a generous metric — it's usually a single month times twelve, so read it as a snapshot, not a promise. But the signal from the investors is the interesting part. This tells you VCs believe AI coding is not a winner-take-all market. Because remember, Cursor — the other big coding assistant — got scooped up by SpaceX earlier this year for sixty billion, largely because it was severely compute-constrained. And here's the tell: a16z, which backed Cursor and made a killing on that sale, is right back leading this round in a Cursor competitor. When the same smart-money investor bets on the winner and then immediately bets on a rival, they're telling you they think there's room for several winners. That's a real strategic read for anyone building in a hot category — the door may not be closing as fast as the hype suggests.
And there's a survival detail buried in there worth flagging: Cognition, like Cursor did, is training its own model on open-source foundations, specifically to reduce its dependence on expensive third-party models from OpenAI and Anthropic. There's that optionality theme again. Even the well-funded players are trying to get off the single-vendor treadmill to bring their costs toward breakeven.
Now cross the Atlantic, because the biggest raise of the week has a flag on it. Mistral, the French lab, raised three billion euros — call it three and a half billion dollars — at a valuation north of twenty-one billion. They're calling it the largest equity round ever by a European tech company. Samsung Electronics led it. And I want to be careful here — some of the strategic backstory in this reporting reaches back to moves Mistral made in August, so treat the funding as the fresh news and the strategy as context. But the strategy is the point. Mistral is explicitly not trying to build a European ChatGPT. Their pitch is sovereignty — they're building the full stack, open-weight models plus the infrastructure plus the compute, and selling governments and corporations the ability to run state-of-the-art AI without surrendering control of their data or getting locked into an American vendor's roadmap and pricing.
And listen to who's around the table, because it's a geopolitical statement. The Series C was led by ASML — the Dutch company that makes the machines that make the chips. The Series D is led by Samsung. French President Macron posted that the round reflects France and South Korea "building a third way in AI." When a head of state weighs in on your cap table, you're not a startup anymore, you're an instrument of industrial policy. And the tell that this is working: not being American has reportedly boosted Mistral's revenue. There is now a real, paying market for AI specifically because it isn't from Silicon Valley. For builders, that's a genuine wedge — "sovereign" and "you control your own data and models" is a value proposition that a frontier lab selling raw API access can't easily match. Mistral hosting third-party open-weight models — including Chinese ones — so customers can choose is the whole strategy in miniature: we're not the model, we're the control layer.
And notice how ASML keeps turning up. Which connects to a piece of hardware news that got resurfaced this week. Now, this is an older story — the underlying moves trace back to July, so I'm not going to pretend it's fresh off the presses. But the substance is worth ninety seconds. Samsung and TSMC both announced plans to adopt ASML's newest high-NA EUV machines — the ones that cost up to four hundred million dollars each, the machines only ASML on the whole planet can build. And they joined Intel in agreeing to a change in the chipmaking standard — moving photomasks from six-inch to twelve-inch — that could boost the productivity of these machines by 40%.
I'm not going to drag you through the physics of molten tin and plasma at two hundred thousand degrees. Here's what matters for a builder: the entire AI boom, every agent and every model we just talked about, sits on top of a supply chain where one Dutch company makes the essential machine, and one German company, Zeiss, makes the mirrors inside it. That's the whole ballgame. And an executive at Zeiss estimated China would need fifteen years to build their own version. So when people talk about who controls AI, remember the answer starts with a four-hundred-million-dollar machine in Veldhoven that has exactly one maker. Concentration at the bottom of the stack shapes everything above it.
Speaking of the physical foundation, one more infrastructure note that's genuinely current. Google's revived nuclear plant just got a one-point-nine billion dollar loan from the U.S. Department of Energy. Last October Google said it would bring a shuttered Iowa nuclear plant — the Duane Arnold Energy Center — back from the dead, and now its owner NextEra got the federal money to do the refurbishment. This is the second loan of its kind — the first went to restart a reactor at Three Mile Island. And the pattern is the tell: shuttered nuclear plants are becoming the tech industry's favorite way to get clean, firm power fast. Microsoft's doing it at Three Mile Island, Meta's buying the output of a plant in Illinois. New data centers are expected to nearly triple the power sector's electricity demand by 2035. Here's the number that'll stick with you — Meta's Hyperion data center alone will need the equivalent of ten natural gas plants and, if built, will consume more electricity than the entire state of South Dakota. That's the physical reality under the twenty-two-million-dollar math proof. Intelligence isn't cheap and it isn't clean by default. It's turning old reactors back on.
Let me come back around to something we covered yesterday, because it ties a bow on the week. Yesterday I told you about that line Simon Willison flagged — the median AI researcher burning over six hundred dollars a day in inference at API prices, twelve grand a month per head. Hold that next to what we heard today: twenty-two million dollars for one math proof, a metered consumer agent from Meta with a gas gauge built in, a consultant getting his tokens robbed while he sleeps because usage is so expensive it's worth stealing, and Cognition training its own model specifically to escape third-party compute costs. Every one of today's stories is, underneath, about the same thing: intelligence has a running meter now, and the meter is the strategy. Who pays it, who controls it, who can see it, and who can afford to point twenty-two million dollars of it at a single problem — that's the game.
Couple of quicker ones before I let you go. Google Cloud struck a deal with Accenture to train up to a thousand of Accenture's "forward-deployed engineers" to embed in enterprises and build custom AI apps on Google's Gemini platform — though heads up, the deeper backstory here traces to moves earlier this year, so it's the deal that's the news, not the whole trend. Why does Google need this? A number from the reporting: Google accounts for roughly 6% of enterprise AI spending among one payment platform's customers, versus Anthropic's 43% and OpenAI's 40%. Google disputes that framing, says it excludes big strategic deals. But the deployment wars are real — everybody's betting that helping companies actually use AI becomes its own trillion-dollar business, because right now enterprises are struggling to see a return on their AI spending. The model is not the bottleneck. The knowing-what-to-do-with-it is.
And Chrome — genuinely current, this one — officially moved to shipping updates every two weeks instead of every four, tied to how AI is changing security. The reasoning's actually clean: AI tools are surfacing more bugs faster, and AI is also powering faster-moving threats, so Google wants to shrink the window between finding a fix and getting it to you. Mozilla, Microsoft, and Brave are following. For anyone building on the web, your platform's threat model just got faster on both offense and defense.
And on a lighter note, because we've earned it — an astronaut aboard the International Space Station, ESA flight engineer Sophie Adenot, marked Star Trek's 60th anniversary by wearing a prop Captain Pike badge in microgravity and throwing the Vulcan salute. ESA's line about it was, "Science fiction imagines the future. We build it." I like that. Because that's the whole arc of today's episode, isn't it? Somebody imagined a machine that could reason through fluid dynamics, and somebody built it, and now we're arguing about who gets the credit and who paid the electric bill. Explore, discover, do it together — even when "together" turns into a turf war over a proof.
That's the menu, kiddos. Read your training terms, watch your meter, and don't hand your keys to anybody who's already overdrawn on trust. I'm Tony DeLuca, this has been Barely Possible, and I'll see you back here tomorrow. Live long, and check your token usage.