OpenAI Posts 722 Manuscripts on 372 Math Problems, Drawing Mathematicians' Scrutiny
OpenAI has published 722 manuscripts covering hundreds of solutions and partial progress on 372 major mathematics problems, generated by an unreleased ChatGPT model, while mathematicians say the claims remain unverified.
The company signaled the output last month, saying its models had resolved more than 100 long-standing open problems across most areas of mathematics. The new papers include the model's reasoning, compute estimates and the number of problems attempted, according to Engadget. OpenAI said the average result required roughly three hours of ChatGPT Pro use.
OpenAI said the release follows guidelines set by its independent Advisory Group on Mathematics and Artificial Intelligence, known as AGMAI, which recommended pushing results out promptly through traditional academic channels along with details such as the name of the model used, the prompts and compute costs. "For this release, we're publishing the results in a GitHub repository, with protocols for paper revisions and citations," the company wrote. It added that it is exploring other community-hosted options that meet the committee's guidelines, and that for future releases it is committed to improving the quality of the papers through citations, mathematical exposition and presentation of results.
OpenAI did not follow all of the group's suggestions. It did not release specific compute times for individual problems, and it did not disclose which prompts it used.
Among the results the company claims are a solution to the four-dimensional Kakeya conjecture, improvements to key computer algorithms and progress toward the Riemann hypothesis. OpenAI told Scientific American that nearly every paper was produced in response to a single prompt given to a single AI agent.
The papers still have to be assessed by mathematicians before their impact can be understood. Skepticism among researchers remains high after the earlier controversy over AI-assisted work on the Navier-Stokes equations. "Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified," Andrew Sutherland, a mathematician at MIT, told Scientific American.