OpenAI says one of its models solved 100+ open math problems; forms independent advisory group at Princeton
OpenAI claims an internal model trained since late August has resolved more than 100 long-standing open problems in mathematics, including the Navier–Stokes Millennium Prize problem. A new nine-mathematician advisory group at the Institute for Advanced Study gets a voice in how the results reach the world — but no vote on how fast the research moves.
On Monday, September 21, OpenAI announced an independent advisory group for mathematicians, hosted at the Institute for Advanced Study in Princeton — and, in the same breath, the claim that one of its internal models has solved more than 100 long-standing open problems in mathematics, including one of the Millennium Prize problems. The model, the company says, began training on August 28. Barely three and a half weeks of training for results that would normally define careers.
The announcement did not land in a vacuum. Earlier this month, 25 Fields Medal-winning mathematicians signed an open letter arguing that the labs' race to solve famous problems is damaging mathematics itself. OpenAI's new group is, in part, an answer to that letter — though attentive readers will notice what the group is explicitly not allowed to do.
What OpenAI is claiming#
The headliner is the Navier–Stokes problem — one of the seven Millennium Prize problems, a question about whether smooth solutions to the fluid dynamics equations always exist. OpenAI says its abrupt publication is what prompted the creation of the advisory group, and that the same model has since resolved more than 100 additional open problems across most areas of mathematics.
The company says the pace surprised even its own mathematicians and triggered internal debates about how to inform the field. It frames mathematics as a fundamental science whose discoveries ripple outward, which makes the responsible deployment of math-capable AI a question that matters well beyond mathematics.
One caveat looms over all of it: these are claims in a company announcement, not published papers. As of this writing, the 100-plus results have not been released, peer-reviewed, or independently verified. The story so far is OpenAI's telling of its own success — and the group it just created exists partly to figure out how the world gets to check the work.
The advisory group: voice, but no vote#
The Advisory Group on Mathematics and Artificial Intelligence starts with nine prominent mathematicians: François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood. It has its own public charter at agmai.org and is already soliciting input from the broader mathematical community.
The unusual part is how independence is defined. The group can offer advice OpenAI never asked for, publish its views, and change its own membership — and it accepts no payment. But its remit is deliberately narrow: an advisory body for how results get communicated, not a brake on the research.
| The group can | The group cannot |
|---|---|
| Advise on the significance and release of new results | Advise on the pace of OpenAI's internal progress |
| Publish its recommendations publicly | Slow down or redirect ongoing research |
| Criticize OpenAI's impact on mathematics unsolicited | Make binding decisions — power stays with the company |
| Advise any AI company, not just OpenAI | Bind anyone to follow its recommendations |
The limitation is explicit in OpenAI's own post: "the group will not be responsible for advising us on how to pace our internal progress on mathematics." The Institute echoed it: "Although we will give advice, we do not have decision making power at any AI company, and the responsibility for the decisions made by any company will rest with that company." Voice, yes. Veto, no.
The backlash that forced the issue#
The context is the September 11 open letter "A Severe Misalignment of AI in Mathematics," signed by 25 Fields Medalists including Terence Tao, Peter Scholze, Maryna Viazovska, Artur Avila, and Martin Hairer. OpenAI's announcement points to the letter directly.
The letter's argument is not against AI doing mathematics — it is about how labs are doing it. Solutions get announced in a rush, with no time for proper writeups, for isolating the new ideas inside them, or for crediting prior work. That raises serious attribution and plagiarism questions. And an accelerating "mass production of true/false statements," the signatories warn, risks salting the earth: fertile territory gets exhausted before human mathematicians can develop it, and AI-conceived results wither unless mathematicians connect them to the existing body of mathematics. Their core charge: "the goals of the AI companies and the goals of the mathematical community are severely misaligned."
Against that charge, OpenAI's new group is a concession — an admission that engagement with the mathematical community has been too thin. Whether an unpaid, non-binding advisory body heals the misalignment or merely institutionalizes it is the open question.
What to watch#
First, the flood: whether the promised wave of results actually appears with enough detail for the mathematical community to verify — machine-assisted proofs can be brilliant, and brutally hard for humans to audit. Second, the group's first public recommendations at agmai.org, the first real test of whether this body bites or merely barks. Third, whether other labs take the group up on its offer to advise any AI company; if they don't, it risks looking like a PR structure attached to a single lab rather than a community institution.
The tension between the two futures — AI as raw material mathematicians build on, or as strip-mining of their field — is exactly what the September 11 letter warned about. It is now OpenAI's problem to manage. With nine unpaid advisors and no brakes on the engine.