What was announced
OpenAI announced on 21 September that it has formed an Advisory Group on Mathematics and Artificial Intelligence, hosted at Princeton’s Institute for Advanced Study. It has nine founding members: François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten and Melanie Matchett Wood.
The group exists because of a claim OpenAI has made and not yet substantiated. The company says an internal model, which began training on 28 August, has resolved more than a hundred long-standing open problems across most areas of mathematics. That is in addition to the Navier–Stokes result the company announced earlier. The announcement does not list the problems.
That is the group’s first job. In its own launch statement it said it is “currently facing the very specific challenge of advising OpenAI on how to coordinate the release of a large number of significant results in mathematics that they report have been produced by their internal model.”
What the group can and cannot do
The limits are stated up front, which is unusual and worth crediting. Members are unpaid. They have no authority over the pace of OpenAI’s research, and OpenAI will not consult them on it. The group’s own wording is that responsibility for decisions rests with the company.
What they can do is assess how significant the results are, and coordinate their release so that the mathematical community can check them in the way it checks anything else. That is not a small task at this volume. A hundred results arriving at once, from a source that cannot be interrogated about its reasoning, is a refereeing problem before it is anything else.
The letter behind it
The group did not appear in a vacuum. Twenty-five Fields medallists signed an open letter titled “A Severe Misalignment of AI in Mathematics”, objecting to the practice of treating open problems as an AI benchmark. Of the nine members of the new group, only De Lellis signed that letter.
That composition tells you something about what the group is for. It is not a panel of the most sceptical voices in the discipline, and its membership is not broad: nine researchers cannot represent a field. Some mathematicians have read it as a constructive first step; others have asked what influence a body with no authority over research pace and no say over publication can actually exercise.
The thing to watch is not the group’s statements but its output. If the first tranche of results is released in a form that referees can work with — full proofs, the problems named, the model’s role in each one specified — then the arrangement has done something. If the results arrive as an announcement with a number attached, the group will have lent its names to a press release. The mathematicians on it are in a position to insist on the first, and the letter suggests the discipline will notice which one it gets.
