Independent math group forms to advise OpenAI on AI-found proofs

Nine unpaid members launched the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), hosted at the Institute for Advanced Study, after OpenAI approached some of them about an external board, and they are already advising OpenAI on how to coordinate the release of a large number of significant math results the lab says an internal model produced.

FUTURE desk — frontier labs are producing research-grade math claims faster than the community can absorb them, and an independent unpaid group is now the public adviser on how OpenAI should release a large batch of those claims.

The group says it is hosted at the Institute for Advanced Study in Princeton, and online at agmai.org. The Institute for Advanced Study, often shortened to IAS, is a research institute. It is not a company, and it is not OpenAI. The group’s purpose, in its own words, is to advise AI companies on how they deal with mathematical research and with mathematicians, including how results are presented and released. It says it wants to work for the best interest of mathematics, and to be one channel between mathematicians and the AI industry. One channel means other channels can exist. File that purpose as the group’s.

Independence is written on the same page. The group says it operates independently of any AI company, and that members do not accept payment for this work. Unpaid means no fee for this advice. It says it will publish its recommendations on its website. A recommendation is advice. It is not an order. The group says it is willing to offer that advice to any AI company whose models are likely to have a significant impact on mathematics. It also says it has no decision-making power at any AI company, and that the company remains responsible for its own decisions. File those limits as the group’s. Advice is not a veto.

The page lists nine members, with the homes it prints: François Charles (ENS-PSL), Camillo De Lellis (IAS and GSSI), Timothy Gowers (Collège de France and Cambridge), Martin Hairer (EPFL and Imperial College London), Nikhil Srivastava (Berkeley and the Simons Institute), Ulrike Tillmann (Oxford and the INI), Ravi Vakil (Stanford), Edward Witten (IAS), and Melanie Matchett Wood (Harvard). Nine is this desk’s count of that printed list. The page does not print the numeral. ENS-PSL is the École normale supérieure in Paris. GSSI is the Gran Sasso Science Institute. EPFL is the Swiss federal institute of technology in Lausanne. The INI is the Isaac Newton Institute in Cambridge. File the names and the printed homes as the group’s. This desk did not check each appointment.

How the group formed, still that page. It came together after OpenAI approached some of the members about an external advisory board. An external board, here, means advisers who are not employees. In agreement with OpenAI, those members decided to make an independent group and invite others to join. “Some” means the page does not say OpenAI asked all nine. “In agreement” means the page says OpenAI accepted that independent shape. It does not say OpenAI pays them, appoints them, or must follow them. File that origin as the group’s.

The current job is narrow, and it is the news. The page says the group is 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.” An internal model is a model the company has not handed to the public. The page does not name the model. “They report” means the results are what OpenAI says it has. The group is not saying it has checked them. “A large number” is the only size the page gives. Do not turn that into a count. The group asks mathematicians to send thoughts through a form at agmai.org/input, as soon as possible. It says those answers will inform the recommendations and will not be made public without the sender’s approval. File the task, the form, and that privacy line as the group’s. This desk did not open the form or count replies.

Do not read this as a claim that the results are verified. Do not read it as a claim that mathematics is solved. Do not invent a number of problems past “a large number.” The group does not list the results, does not name a model, and does not say a journal has accepted anything. A plan to advise on a release is not the release.

Earlier context, fenced off. On 1 Aug 2026 OpenAI posted “Ten advances in mathematics and theoretical computer science.” That page says an internal version of a model the company calls Astra produced a selection of ten results, and that people then prepared the write-ups. That is an earlier company post. The 21 Sep AGMAI text does not say the batch under advice is those ten, does not name Astra, and does not give a count past “a large number.” Do not import ten as today’s number. Do not treat the August page as a check on the batch the group is discussing now.

Plain English for the rest of the card: AGMAI = Advisory Group on Mathematics and Artificial Intelligence. IAS = the Institute for Advanced Study in Princeton, the host named on the page. external advisory board = advisers who are not employees of the company. internal model = a model the company has not released; the 21 Sep text does not name it. recommendation = advice the group says it will publish; the company still decides. unpaid = members say they take no fee for this work. a large number = the only size the group gives for the batch. This filing is the 21 Sep announcement of the group and its OpenAI task. It is not a proof check and not a product launch.

PRIMARY here: agmai.org and the 21 Sep 2026 guest post on Terence Tao’s blog — Tier A PRIMARY, the group’s own announcement, the originating record. The host at the Institute for Advanced Study, the unpaid and independent lines, the no-decision-power line, the willingness to advise other AI companies, the nine printed members and their homes, the OpenAI approach and the independent-group choice, the “large number” of results the company reports from an internal model, the form at agmai.org/input, and the promise to publish recommendations are the group’s, on both pages. Tao’s AI-conversion note is his header, about the blog text. The 1 Aug 2026 OpenAI page is prior context only, not today’s count and not today’s model name. NOT claimed: that the results are verified, that mathematics is solved, a number of problems past “a large number,” that this batch is the August ten, that the group can block a release, that this desk checked a proof or opened the form, a stock tip, or investment advice. Distinct from the already-filed un-ai-panel-agents, gates-ai-language-partnership, and microsoft-global-ai-diffusion-q2.