On August 28, OpenAI started training a new internal model. By September, it had solved the Navier-Stokes Millennium Prize problem, one of the most famous unsolved problems in mathematics, and then kept going. As OpenAI announced on September 21, the model has now resolved more than 100 long-standing open problems spanning most major areas of mathematics. The pace was fast enough to surprise the mathematicians working inside OpenAI itself. That detail is worth sitting with for a moment.
This is not a benchmark result or a carefully staged demo. OpenAI is describing a model that is producing mathematical discoveries faster than the research community can process them. For context, the Millennium Prize problems are a set of seven problems selected by the Clay Mathematics Institute in 2000, each carrying a $1 million prize. Only one had been solved before now, the Poincaré conjecture, resolved by Grigori Perelman in 2003. Navier-Stokes, which concerns the behavior of fluid flow, has been open for over 150 years.
The reaction from the mathematical community has been mixed. A recent open letter titled “A Severe Misalignment of AI in Mathematics” raised concerns about AI companies using open problem resolution as a benchmark, arguing that this creates negative externalities for working mathematicians whose careers and research agendas are directly affected. It is a legitimate concern. When an AI system solves a problem a researcher has spent a decade on, the professional consequences are real, even if the mathematical progress is genuine.
OpenAI’s response is to form an independent advisory group hosted at the Institute for Advanced Study. The group includes some of the most respected names in modern mathematics:
- 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 Simons Institute
- Ulrike Tillmann, Oxford and INI
- Ravi Vakil, Stanford
- Edward Witten, IAS
- Melanie Matchett Wood, Harvard
The group’s mandate is deliberately broad. It will advise on how results are reviewed and communicated, help assess their significance, and weigh in on academic standards. But it also has real independence. Members are not paid by OpenAI, the group can change its own membership, and it has the freedom to offer advice OpenAI has not asked for and to make that advice public. It is also explicitly not responsible for advising on how quickly OpenAI should push internal progress. That boundary matters, because it keeps the group from being captured as a rubber stamp.
The broader implication here goes well beyond mathematics. If an AI model can resolve 100 open problems in weeks, the same pattern could eventually apply to other fields where formal reasoning and large bodies of structured knowledge matter: theoretical physics, economics, parts of biology. Mathematics is in many ways the cleanest test case because proofs are verifiable. But the governance question OpenAI is now wrestling with, how to deploy capabilities that can outpace entire scientific communities, is one the industry has not solved.
Compared to how other labs have handled analogous situations, OpenAI’s approach here is more structured than most. Forming an independent group with genuine authority to dissent publicly is not just optics. Still, the harder questions are ahead. How do you distribute access to these capabilities fairly? What happens to mathematical education when proofs can be generated on demand? OpenAI says it wants mathematicians at the center of shaping those answers. The advisory group is a start, but it is a long way from a complete answer.
