OpenAI released 722 manuscripts on GitHub on October 6, organized into 372 distinct families of mathematical results. These findings, generated by an unreleased internal AI model, span various fields including algebra, number theory, geometry, and theoretical computer science. The company stated that the results are at different stages of verification, and for many, it provided proofs written in Lean, a programming language enabling computer-checked logical arguments. OpenAI also published 10 summaries of the model's reasoning and estimated that the average result required computing power equivalent to roughly three hours of ChatGPT Pro usage.
This release follows a previous instance last month where OpenAI claimed its AI model solved the Navier-Stokes equation, a Millennium Prize problem, which drew criticism from the mathematical community for its handling. The current deluge of findings has intensified these concerns. The Institute for Advanced Study's independent Advisory Group on Mathematics and Artificial Intelligence, which OpenAI consulted on how to share these results, has stated that it does not endorse the practice of testing advanced mathematical problems on proprietary models. The group emphasized the importance of human understanding in mathematics and urged AI labs to cease this practice.
Mathematicians have expressed a range of reactions, from awe at the scale of the discoveries to anxiety about the implications for the field. Some researchers, such as Tristan Buckmaster, a mathematics professor at New York University, reported that entire research programs have been "wiped out" by the sudden release, impacting early-career mathematicians. Bryna Kra, a mathematics professor at Northwestern University, noted that while AI could advance research, releasing such a large volume of work at once could undermine the collaborative culture of mathematics.
A central point of contention revolves around the verification of these results and the attribution of credit. While some proofs include Lean formalizations, which allow for computer verification of logical steps, this does not automatically confirm originality or importance. The Association for Human Mathematics, in a statement effectively endorsed by Fields Medalist Terence Tao, criticized OpenAI's approach, asserting that "releasing over 700 files at once is not a demonstration of scholarship, but a demonstration of power." The group urged mathematicians to view the publication model with skepticism and to prioritize human understanding in scientific endeavors.
OpenAI, in its announcement, stated its intention to "push the frontier of human knowledge and enable further progress in mathematics." The company plans to fund workshops and programs to facilitate understanding of AI-generated mathematical results. However, the lack of public access to the underlying AI model and comprehensive experimental procedures limits independent reproduction of the research process. Questions also persist regarding whether OpenAI's models may have utilized unpublished work submitted to AI tools, raising ethical concerns about data usage and academic norms.
The release included a breakthrough on a modified version of the Riemann hypothesis, a notable achievement in analytic number theory. However, it did not include solutions to any of the remaining Millennium Prize problems. The ongoing debate highlights a tension between the rapid advancements in AI capabilities and the established norms and collaborative spirit of the mathematical research community.