An AI Model Just Outpaced the Mathematicians Who Built It
The mathematics community just organized its first formal response to a problem it didn't expect to face this soon: what happens when an AI model starts solving problems faster than mathematicians can process them.
OpenAI disclosed that an internal model it began training on August 28 has since resolved the Navier–Stokes Millennium Prize problem — one of just seven Clay Mathematics Institute problems carrying a $1 million prize, and one of math's most stubborn open questions in fluid dynamics — along with more than 100 other long-standing open problems spanning most areas of the field. OpenAI said the pace of progress "surprised the mathematicians within OpenAI" itself, prompting internal debate over how to inform the broader field without causing chaos.
A Response Built by the Field, Not the Company
Rather than leaving that decision entirely to OpenAI, a group of prominent mathematicians has stepped in. The newly formed Advisory Group on Mathematics and Artificial Intelligence is hosted at the Institute for Advanced Study in Princeton — the same institution where Albert Einstein once worked — and publishes independently at agmai.org.
Its membership reads like a shortlist of the field's most respected names: theoretical physicist Edward Witten, Fields Medalists Timothy Gowers and Martin Hairer, along with mathematicians from Harvard, Stanford, Berkeley, Oxford, and the École Normale Supérieure. According to the group's own announcement, it came together after OpenAI approached several of its eventual members directly about forming an external advisory board — meaning this wasn't purely researcher-initiated pushback, but something closer to a negotiated arrangement between the company and the field it's now reshaping.
What the Advisory Group Will Actually Do
The advisory group's stated mandate is threefold: help assess the significance of new AI-generated mathematical results before they're publicized, advise on how those results should be coordinated and disseminated to the field, and set expectations around academic and professional standards for AI-assisted mathematical research. It will also weigh in on how AI tools can be built to genuinely support mathematicians' own work, rather than simply outpacing it.
Crucially, the group has no authority to compel OpenAI to act on its advice. It can publish recommendations publicly and criticize the company's choices in the open, but final decisions — including what gets released, when, and how — remain entirely OpenAI's. Members are unpaid, and the group describes itself as available to advise any AI company whose models meaningfully affect mathematics, not OpenAI exclusively.
The Unease This Is Responding To
This formation doesn't exist in a vacuum. It follows renewed concern in the math community after Fields Medalist Terence Tao recently warned, via a widely shared post, that indiscriminate AI-driven problem-solving risks depleting the field's supply of good open questions faster than mathematicians can identify new ones — solving problems in the short term while eroding the ecosystem that produces the next generation of worthwhile questions to ask.
An OpenAI model quietly clearing more than 100 of those open problems in a matter of weeks is close to a real-time test of exactly the scenario Tao described.
Early Skepticism From the Sidelines
Reaction outside the advisory group itself has been more pointed. Commentary circulating on Hacker News speculates OpenAI may be sitting on other major unpublished results — the Hodge Conjecture has been floated as a rumor — and questions whether a small, volunteer advisory board can keep pace if the rate of new results keeps accelerating, or whether it risks becoming a bottleneck instead. Independent commentators have also noted that an advisory relationship works only through persuasion and reputation: a recommendation is not a veto, and readers will need to judge OpenAI's actual releases against the standards the group calls for, rather than assuming the group's existence guarantees responsible disclosure on its own.
For now, the group's website functions as little more than a public intake channel for community input — but its formation marks the first organized attempt by mathematicians themselves to shape how the field absorbs a wave of AI-generated results that, by OpenAI's own admission, arrived faster than anyone inside the company expected.
