OpenAI releases findings on hundreds of math problems
Company releases research from unreleased model, says results are at different stages of verification
US artificial intelligence company OpenAI has published hundreds of mathematics research papers produced by an unreleased AI model, making the findings available for examination while warning that some may contain errors.
The collection, released Tuesday on the code-sharing platform GitHub, contains 722 manuscripts grouped into 372 sets of related results. These can include a main finding, supporting arguments or alternative proofs.
The model attempted approximately 4,000 research problems during testing, according to the company.
OpenAI said the published results were at different stages of verification.
For many results, the company also released proofs written in Lean, a programming language that enables computers to check whether mathematical arguments follow logically. It plans to add more as they become available.
OpenAI published 10 summaries of the model’s reasoning and estimates of the computing power used. It said the average result required computing power equivalent to roughly three hours of ChatGPT Pro “thinking.”
The company said it consulted the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study on how to present and share the research.
"This is an important event for mathematics, with consequences both for mathematics and for the mathematical community that extend far beyond the individual results," the group said, stressing its advisory role did not amount to an endorsement of the results
“Making this work public is a first step. This release is the beginning, not the completion, of the process of human understanding and the incorporation of the work into mathematical knowledge,” it added. “The future of mathematical research cannot consist only of understanding results produced by AI labs.”
