OpenAI says an internal model cleared 100+ open math problems. IAS just stood up a group to advise on the release.

OpenAI says a new internal model, training since about August 28, has resolved more than 100 long-standing open problems across most areas of mathematics. They are also pointing at Millennium Prize territory again, including a claimed Navier-Stokes result that is already under debate.

Alongside the claim, they announced the Advisory Group on Mathematics and Artificial Intelligence. It sits at the Institute for Advanced Study, with a public site at agmai.org. The group advises on how results get reviewed and shared with mathematicians. It does not get to set OpenAI's research pace.

Who is on the group

The IAS announcement names François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, and Melanie Matchett Wood.

OpenAI asked some of them about an external board. They formed an independent group instead, invited others, take no pay for the work, and say they will publish recommendations. That is credibility on the communications path, not a rubber stamp on what the model can do.

What is claimed vs what we can check

From OpenAI and this week's coverage:

• Training on the new internal model started around August 28.
• The company reports 100+ long-standing open problems resolved across most of mathematics.
• Navier-Stokes is already in the public fight. Some coverage also mentions Hodge as reported. We have not independently confirmed Hodge here.
• Company mathematicians were reportedly surprised by the speed. OpenAI says the hard part now is how to warn academia.

Still missing in public:

• The full problem list
• Method, human guidance, compute cost, and failures
• Writeups a field can stress-test
• Independent confirmation that each "solution" is a real proof

Decoder is blunt that which problems were solved, how, and what they mean remain open. Treat "100+" as a company claim until the artifacts land.

Why mathematicians are upset

An open letter from Fields Medalists and others, "A Severe Misalignment of AI in Mathematics," argued against scoring AI by how many open problems it clears. Wrong proofs are only part of the fear. Dumping machine solutions can crowd out the conceptual work that trains people and keeps a field alive.

Timothy Gowers joined AGMAI and did not sign that letter. He agrees math is in trouble, then splits on the goal. Some mathematicians chase problems and use understanding as a tool. Others chase understanding and use problems as a tool. His sharper worry is social. If famous problems stop being a human dream worth a PhD, fewer people become custodians of the tradition. Funders could also decide human mathematicians are optional. He expects strong public models within months and saw little point scolding labs for producing solutions too fast.

For people outside math, the fight is about release, verification, and whether anyone still wants to do the human part when a lab can outrun a field's digestion rate.

What AGMAI will and will not do

From the guest post on Tao's blog, the group will:

• Advise AI companies on how they interact with mathematical research and how results get presented
• Act as one channel between mathematicians and industry
• Stay independent, publish recommendations, take no payment
• Advise any company whose models are likely to matter for math
• Right now, advise OpenAI on coordinating release of a large batch of results the company says its model produced

Out of scope: how fast OpenAI trains. OpenAI keeps that dial.

They are collecting community input through a form. Responses inform recommendations and stay private unless the sender approves publication.

The part that matters if you ship software

We see a softer version of this every week. A model looks strong in a demo. Production shows the failure modes. The team that generated the output is not always the team that has to keep understanding it.

If OpenAI's claim survives scrutiny, the bottleneck moves from generating proofs to explaining, verifying, teaching, and integrating them. If it does not, the bottleneck was never generation. It was whether anyone could trust the claim. Either way, ignore the press number. Watch the release process: problem lists, methods, certificates, human writeups, and failures.

AGMAI is trying to put that process under public advice. Watch what they recommend, and whether OpenAI follows it.

Sources

• https://terrytao.wordpress.com/2026/09/21/advisory-group-on-mathematics-and-artificial-intelligence/
• https://agmai.org
• https://the-decoder.com/openai-says-its-internal-model-solved-over-100-long-standing-math-problems-after-just-a-month-of-training/