OpenAI’s mathematical paper release will continue

OpenAI released more than 700 AI-generated mathematical preprints on 6 October.[3] Science reported the release contained 722 preprints covering 372 distinct results.[4] According to the company, an unreleased ChatGPT model attempted about 4,000 problems.[4] OpenAI said the vast majority of solutions took an average of the equivalent of three hours of ChatGPT Pro “thinking.”[4] The figures describe different measures: the number of attempted problems, the number of preprints, and the distinct results covered.[4]

OpenAI said it would contribute $5 million to a Simons Foundation initiative supporting mathematicians working with AI results.[5] The foundation will manage and award the funding for workshops, conferences, and special programs.[5] OpenAI also said it would provide models and computing resources to researchers supported by the foundation.[5] The company published its mathematical results in a GitHub repository with protocols for revisions and citations, and said it would add more formalizations as they became available.[5] OpenAI said its goal was to provide tools for mathematicians and scientists.[3]

As Science went to press, OpenAI said it had withdrawn three manuscripts and revised 14.[4] Of the 719 papers left after those changes, 300 had been formalized in Lean, according to Science.[4] Science cautioned that Lean formalization does not by itself verify that a paper’s English-language claim is correct.[4] Reactions among mathematicians were mixed: some celebrated the solutions, while others complained that OpenAI had scooped them.[3] Academics were also calling the release a “giant bomb,” and scientists described the breakthroughs as “really cool and terrible at the same time.”[2]