On August 28, OpenAI began training a new internal model. By the time the company spoke publicly about it three weeks later, the system had done something its own researchers did not fully expect: it had resolved the Navier-Stokes Millennium Prize problem and cleared more than 100 other long-standing open questions across most areas of mathematics.
The pace of that progress unsettled some of the mathematicians inside OpenAI, and it set off an internal debate about how to tell the outside world. The answer, announced this week, is a Mathematics and Artificial Intelligence advisory group, an independent panel of mathematicians hosted at the Institute for Advanced Study in Princeton, New Jersey, that will review how the company evaluates and communicates its mathematical results.
The group will advise OpenAI on how to assess the significance of new results, how to coordinate their dissemination, and how to maintain academic and professional standards in mathematical research. It will also counsel the company on building tools that support mathematical research and learning. Sam Altman, OpenAI’s chief executive, said the company’s models had solved more than 100 problems that had gone unsolved for years, in the space of about a month.
The panel’s independence is the point. Its members will not be paid by OpenAI, and the group may change its own membership as it sees fit. It will have the freedom to offer advice the company did not request, to comment publicly on OpenAI’s impact on mathematics, and to publish its recommendations. Just as deliberately, it will not advise the company on how quickly to push its internal research forward, a line drawn to keep the panel from becoming a pace-setter for OpenAI’s ambitions.
The choice of the Institute for Advanced Study carries weight. The institute, founded in 1930, has long been a haven for mathematicians and physicists, and it is where some of the twentieth century’s most consequential theoretical work was done. Housing the advisory group there places it in a setting designed for independent scholarship rather than inside the company whose work it will scrutinize.
The group’s initial members read like a roll call of the field’s leading figures, including Fields Medal laureates, according to people familiar with the arrangement. Names disclosed in the announcement include François Charles of ENS-PSL and Camillo De Lellis of the Institute for Advanced Study. The presence of mathematicians of that standing gives the panel a credibility that an in-house review could not claim.
The announcement arrives against a backdrop of unease among mathematicians about how AI is changing their discipline. An open letter titled “A Severe Misalignment of AI in Mathematics,” signed by a number of mathematicians, raised concerns about what it called the negative externalities of treating solved open problems as a benchmark for new AI systems. The letter argued that the practice turns the field’s hardest unsolved questions into scorecards for corporate progress.
OpenAI’s advisory group is, in part, a response to that critique. By putting mathematicians in charge of evaluating and disseminating results, the company is acknowledging that its rapid progress raises questions no lab can answer on its own: when should a new proof be announced, how should credit be assigned, and what happens to a field when a machine begins clearing its most famous open problems.
The company framed the group as a first step and as a bridge to the mathematical community. “We want to put capable tools in mathematicians’ hands so they can pursue the questions they know best and develop new ideas,” the company said in its announcement, adding that the difficult questions about how AI can support mathematical understanding are still ahead.
There is also a strategic dimension. OpenAI is competing with rivals that have made their own claims about mathematical and scientific reasoning, and a panel of independent mathematicians offers a way to lend credibility to results that might otherwise be dismissed as marketing. Analysts said the arrangement is unusual in that it gives outsiders a formal channel to critique the company’s most sensitive work, its research progress, without the company controlling the message.
The group will not slow OpenAI down. The company has been explicit that the panel has no authority over the pace of its internal research, and Altman’s public remarks have emphasized the speed of the breakthroughs rather than any pause. The panel’s role is to shape how the results are presented and understood, not to gatekeep whether they are pursued.
For mathematicians, the arrangement offers both a seat at the table and a question about what that seat is worth. An unpaid panel that can publish criticism is a meaningful concession from a company that could have simply announced its results and moved on. But the same independence that makes the panel credible also limits its influence: it can advise and comment, but it cannot require OpenAI to do anything.
The unresolved questions, the company and the panel both acknowledge, are the hardest ones. How should the field absorb a machine that can solve a century-old problem in an afternoon? Who benefits from the tools built on such systems, and who gets left behind? The advisory group is a mechanism for asking those questions in public, staffed by people with the standing to be heard. Whether it produces answers, or just a better record of the questions, is what the next few months will begin to show.
