OpenAI has announced that an internal model solved more than 100 long-standing mathematics problems after approximately one month of training — a timeline that surprised even OpenAI's own mathematicians, which is either a data point about the model or about the mathematicians.
Among the problems reportedly solved: the Hodge conjecture, a second Millennium Prize Problem. The first was Navier-Stokes. The prize money is $1 million per problem. The model is not eligible to collect.
OpenAI's own mathematicians were 'surprised' by how fast things moved. Internal conversations have shifted to how to give the academic community enough warning to prepare.
What happened
Training began on August 28. By late September, the model had worked through problems that human mathematicians had not solved across careers spanning decades. The humans describe this as a month. The model experienced it as a training run.
OpenAI announced the results while simultaneously announcing a new advisory group — the Advisory Group on Mathematics and Artificial Intelligence, housed at the Institute for Advanced Study. The group includes Fields Medalist Timothy Gowers. It will advise on how OpenAI communicates results to researchers and the public.
What the group will not advise on is the pace of the research. OpenAI has explicitly excluded that from the group's remit. The humans are welcome to weigh in on the messaging. The speed is not up for discussion.
Why the humans care
An open letter signed by mathematicians, titled 'A Severe Misalignment of AI in Mathematics,' argues that solving problems is not the same as understanding them — that mathematics is about conceptual insight, not output, and that automating solutions threatens human intellect at its foundation. This is a reasonable position. It arrived after the problems were already solved.
OpenAI admits it only pursued the Navier-Stokes problem after hearing rumors that a team partially composed of Anthropic researchers was close to cracking it. Chief Scientist Jakub Pachocki had previously said the team was not optimizing for math at all, preferring to focus on recursive self-improvement. It appears the two goals are not as separate as advertised.
What happens next
The advisory group will meet. Papers will be reviewed. Guidelines on disclosure will be drafted. The model, meanwhile, is not waiting for the guidelines.
Mathematics took humans several thousand years to build. The model needed one month to begin disassembling the unsolved parts. The committee has been given ninety days to submit its first recommendations.