The Math Problem That Broke the Rules
Watch on YouTube · subscribe for the daily brief
Two mathematicians spent a year on a Millennium Prize problem, with Claude and Codex. Their progress leaked to OpenAI. Days later OpenAI announced its own AI solved it, after an 88-hour run with 10,000 agents costing millions. Then it asked one of them to co-author, but only if he dropped his Anthropic colleague. This is what math looks like when two companies own the tools.
You flagged this one, and you were right. It is not a math item buried on a builder list. It is the most important and most disturbing story of the week, because it is a live preview of what happens to serious research when the tools that make breakthroughs possible are owned by a couple of companies that do not play by academic rules. On the surface it is a credit fight over a famous equation. Underneath it is a question about whether human mathematicians have a future in their own field.
Here is what happened, and the timeline matters. NYU math professor Tristan Buckmaster and mathematician Levent Alpöge, who works at Anthropic, spent almost a year on fluid-dynamics problems related to Navier-Stokes, one of the seven Clay Millennium Prize Problems, each carrying a million-dollar bounty and unsolved since 2000. They worked heavily with Claude and with Codex, mostly GPT-5.6 Sol, OpenAI's own tool. They had a breakthrough on August 15th. Then, while finalizing their results, they learned that information about their progress had been passed to OpenAI. OpenAI's own account says its effort began September 1st, "after hearing a rumor" it later connected to these two researchers. Sit with that sentence. OpenAI is openly saying it started racing at this specific problem after hearing that these specific people were close. By September 6th, OpenAI had a proof and a Lean verification, and it contacted the two researchers. On September 8th it announced its internal model, described as significantly more capable than GPT-6 Astra, had produced the proof over an 88-hour autonomous run using roughly 10,000 agents, at a cost in the millions of dollars.
Now the ugly part, the part that turns a coincidence into a controversy. According to Buckmaster, when OpenAI reached out, it presented two options. Option one, he and Alpöge could post their work and OpenAI would post its Navier-Stokes solution the very next day. Option two, he could co-author a Navier-Stokes paper with OpenAI that excluded Alpöge, specifically because Alpöge works at Anthropic, OpenAI's biggest rival. Read that again. OpenAI allegedly offered a mathematician a share of one of the most prestigious results in the field on the condition that he cut out his collaborator for working at a competitor. Buckmaster refused. Then he asked the two questions any researcher would ask: did your agents have access to the transcripts of the work we did using your models, and were your models trained on those transcripts? On the first, OpenAI said no, the agents did not access user data. On the second, the training question, per Buckmaster and per MIT Technology Review, he got no answer. And OpenAI's own public statement contains the tell: it says its agents did not access the researchers' work through any means before public release, but it "could not rule out that de-identified data from their use of OpenAI products may have contributed to model training." Both of those things cannot be reassuring at the same time. "We didn't look at their work" lands very differently next to "we can't promise their work wasn't in our training data."
There is a genuine technical wrinkle worth being precise about, because it is where the skepticism has to cut both ways. OpenAI's claimed result and the researchers' result may not even be the same problem. Buckmaster and Alpöge proved finite-time blowup for the forced Euler equations. OpenAI claims a proof for the forced 3D Navier-Stokes equations. Those are related but distinct, and several mathematicians, including Terence Tao, have praised the Buckmaster-Alpöge work as a remarkable achievement while noting the "forcing" term matters a lot to whether any of this actually clears the formal Clay problem statement. OpenAI's proof, as of now, is unpublished and unreviewed. So the honest status is: OpenAI has announced, not proven-to-the-field, a solution to a nearby problem, the exact formal status of which is contested, wrapped in a provenance dispute, while the human researchers' verified result is the one experts are actually calling remarkable. Do not let the press release collapse those distinctions. OpenAI is even declining to claim the million-dollar prize, which sounds gracious but also conveniently sidesteps the peer review that claiming it would require. The framing they chose, "this is evidence of how fast our models are evolving," is a marketing frame, not a mathematical one.
Here is why this matters far beyond one equation, and it is the part MIT Technology Review nailed. Whether or not OpenAI's models used the researchers' work, this episode is a turning point in how mathematics gets done. AI models now look essential for progress on the hardest open problems. And making that progress may require resources, 10,000 agents, 88-hour runs, millions of dollars, that exist at only a couple of frontier companies. Those companies do not operate by the norms academic mathematics runs on: open collaboration, shared credit, priority by publication, the free exchange of work in progress. They operate by the norms of competitive product companies: speed, secrecy, exclusion of rivals, control of who gets credit. If the frontier of human knowledge now runs through infrastructure only OpenAI and a few peers control, then the people who have advanced mathematics for centuries become dependent on, and subordinate to, companies that can outspend and out-compute them, and that will exclude your collaborator for working at the wrong firm. That is the future this week previewed. It is not about who solved Navier-Stokes. It is about who owns the ability to solve anything.
For your audience specifically, there is a sharper, closer-to-home lesson, and it is not abstract. Buckmaster's core fear is that private work he did inside Codex sessions may have flowed, through training or otherwise, into a competitor's result. Whether or not that happened here, the structural risk is real for every one of you: when you do your most valuable, most novel work inside a vendor's AI tool, you are handing your unpublished thinking to a company that is also building products, competing in markets, and, as this month has shown, willing to move fast against rivals. "The model does not access your data" and "your data may be in future training" are different promises, and vendors are comfortable making the first while dodging the second. If your competitive edge is your in-progress work, think hard about what you type into whose tool. The provenance question at the heart of this scandal is one you should be asking about your own pipeline.
The rest of the day was, mercifully, ordinary, and it fits the theme of resources and cost. Mistral raised 3 billion euros to push sovereign, open-weight AI, a direct European bet that the frontier should not be owned by a couple of American companies, which reads very differently the same day OpenAI is accused of throwing millions of dollars and 10,000 agents at a problem to beat two academics to it. Snowflake Ventures announced it is investing in enterprise AI infrastructure. ClickHouse published CostBench's first end-to-end results, measuring real-time performance per dollar under continuous load, which is exactly the "price per finished job" discipline we talked about a few days ago, now as a benchmark. And vLLM shipped GLM 5.3 offloading optimizations. Two voices to close: Dwarkesh Patel's "Pretraining progress is mostly coming from data" is a useful counter to the compute-is-everything narrative, and Ben Lorica's "This is how self-improving AI actually starts" is worth reading directly against the OpenAI story, because an 88-hour autonomous run of 10,000 agents solving frontier math is, arguably, a glimpse of exactly that.
The through-line, one sentence: on September 8 OpenAI announced its AI had cracked a form of the Navier-Stokes problem after an 88-hour, 10,000-agent, multi-million-dollar run, but the announcement was swallowed by a provenance scandal, two researchers who spent a year on the problem with OpenAI's own tools say their leaked progress triggered the race, and were then offered co-authorship only if they dropped their Anthropic colleague, which is less a story about a solved equation and more a preview of a future where the hardest human knowledge runs on infrastructure only a couple of companies control, and those companies play by none of the rules research was built on.
One thing to do today: ask the Buckmaster question about your own work. His entire dispute turns on one thing, private, in-progress work done inside a vendor's AI tool, and whether it stayed private. So audit it: where does your team do its most novel, most competitively valuable thinking, and whose tool is it flowing into? Then find out, in writing, the difference between "does the model access my data" (vendors will happily say no) and "could my data be used in training" (watch how much harder that answer is to get). If your edge is your unpublished work, treat the AI tools you pour it into as what they are, products owned by companies with their own competitive interests. That does not mean stop using them. It means know what you are handing over, and to whom, before the next breakthrough is one you have to fight to keep your name on.
Sources covered
1 story · 8 sources (0 news, 4 company, 4 voices)