OpenAI releases mathematical work tied to 377 unsolved problems

OpenAI says it expanded its evaluations after the model’s performance on existing mathematical evaluations saturated, and that it posed approximately 4,000 problems to the model.[2] The evaluation figure is not the same as the 377 previously unsolved problems described in Ynet’s account of the published work.[2][3] OpenAI says the average result used compute equivalent to roughly three hours of ChatGPT Pro thinking, and the repository includes 10 summaries of the model’s reasoning.[1] Those figures offer some information about the work’s production, but do not by themselves establish the novelty or correctness of each result.

Scientific American reports that many results had already been verified in Lean, which it describes as a programming language that validates proof logic.[4] The model that produced the work has not been released publicly.[4] MIT mathematician Andrew Sutherland told Scientific American that one-agent, one-shot claims should be treated as unverified until the model is released and results can be replicated.[4]

Ynet reports that some papers concern challenges linked to three of the seven Millennium Prize Problems.[3] Scientific American reports that the work includes a claimed solution to the four-dimensional Kakeya conjecture and progress toward the Riemann hypothesis.[4] It also reports that mathematicians may need months to determine whether the proofs contain novel ideas or combine existing techniques.[4] OpenAI says it has consulted an independent advisory group on best practices and will fund workshops, conferences and special programs to help people understand major AI-produced results.[1]