Invest1 publisher3 min readPublished
OpenAI asks nine unpaid mathematicians how to release its 100-plus math results
OpenAI says its new nine-member mathematics advisory group will not advise on the pace of its internal research, and the group's own website says the decisions rest with the AI companies.
The Investor · Invest desk
What happened
- Nine mathematicians now advise OpenAI on how it presents and releases its AI-generated math results, as the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study in Princeton.
- OpenAI began training a new internal model on August 28 and says it has since solved more than 100 long-standing open problems, alongside a claimed solution to the Navier-Stokes Millennium Prize Problem.
- An open letter launched on September 11 by 25 Fields Medal winners says labs racing to solve famed problems put mathematicians' intellectual work at risk, and now lists 27 medalists and over 7,800 endorsers.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- cost Nine of the field's most credentialed appraisers work for free, so the only price OpenAI pays for their names is whatever they choose to say in public.
- constraint The remit stops at release, so the arrangement leaves the model's pace exactly where it was; the nine inherit a backlog whose size and timing were fixed before the group existed.
- exposure Anyone pricing OpenAI's math claims now has to price provenance as well as verification, because a machine check settles whether a proof closes and human authorship is a separate question.
- contradiction OpenAI's post has the company bringing the panel in; the group says OpenAI approached some of them about an external advisory board and they formed an independent group instead. Who owns the remit depends on which account you take.
Cryptopolitan reported on September 8 that OpenAI said its internal model had shown finite-time blow-up for the forced Navier-Stokes equations, verified in the Lean proof-verification language [6]. Training had started eleven days before the Millennium Prize claim [7].
The nine appraise the importance of new results, coordinate when and how they are released, weigh them against the field's academic and professional norms, and advise on serving working mathematicians with OpenAI's tools [2]. On speed, OpenAI wrote that the group "will not be responsible for advising us on how to pace our internal progress on mathematics" [3]. The group's website sets out its standing: "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" [4].
The members are unpaid, and several of them hold Fields Medals [8][14]. OpenAI, for its part, has said it does not intend to claim the Clay Mathematics Institute's $1 million prize [9].
Related blow-up results were published days earlier by the NYU mathematician Tristan Buckmaster and the Anthropic researcher Levent Alpoge [10]. Buckmaster said he was told "very little human input" went into OpenAI's proof, an account he says fell apart on a call with the company [11]. He wrote that an entire team had worked the problem, that easier cases including Euler came first, and that the prompt shown to him had itself been generated by prompting Codex [12]. Sebastien Bubeck, the OpenAI researcher at the centre of the exchange, said the allegations were "false and inflammatory", and denied using the mathematicians' prompts or proofs [13].
Lean checked the proof [6]. Who wrote the prompt is outside what it checks, and that is what Buckmaster and Bubeck disagree about [11][12][13]. A panel that appraises importance and coordinates timing is pointed at the first question. In my view this is a release apparatus, and the counter-argument sits in the same material: the nine set their own membership, can issue unsolicited recommendations and can comment publicly on the company's impact [8]; three of them are on the declaration warning that labs racing at famed problems put mathematicians' intellectual work at risk, with Martin Hairer signing as a Fields Medalist and Camillo De Lellis and Ravi Vakil endorsing [15][16]; and the group said it would offer the same counsel to any AI lab whose models might reshape the field [17]. OpenAI cited the letter directly, framing the criticism as a reason to engage the community more thoughtfully [19].
The first coordinated release is where this gets settled. If the nine publish a recommendation OpenAI declines to follow, and say so, the design is a check and I have read it wrong.
What to watch
- Whether the first coordinated batch of the 100-plus results names its human contributors, the point Buckmaster and Bubeck disputed.
- Whether the declaration's endorsers treat the panel's involvement as an answer to their objection.
- Whether a second lab takes up the group's standing offer to advise any AI lab whose models might reshape mathematics.