OpenAI on Tuesday released 722 math manuscripts grouped into 372 result families from an unreleased internal AI model, leaving mathematicians to judge whether it followed new release guidance
OpenAI on Tuesday released 722 math manuscripts grouped into 372 result families from an unreleased internal AI model, leaving mathematicians to judge whether it followed new release guidance.
Key Points:
- OpenAI posted 722 manuscripts in 372 families on GitHub, all produced by an internal model the public cannot use.
- An independent advisory group of mathematicians said its role is not an endorsement of the results or the process behind them.
- The company said some results that lack a formal computer check could have issues.
The company published the collection in a public GitHub repository, with rules for revising and citing each paper. The repository lists 722 manuscripts in 372 families, where a family groups a main result with companion arguments, consequences or alternative proofs. OpenAI said it posed about 4,000 problems to the model and kept only results it judged significant enough.
Each result used roughly three hours of ChatGPT Pro thinking compute on average, according to the company. Many of the proofs come with versions written in Lean, a programming language that lets a computer check a mathematical proof.
OpenAI said some results without that check could have issues.
Also Read:North Korea's Lazarus Moved Stolen $1B Through Chinese Launderers, ZachXBT Alleges
The Advisory Group on Mathematics and Artificial Intelligence, nine unpaid mathematicians hosted at the Institute for Advanced Study, said its role is neither a judgment of the results nor an endorsement of OpenAI's process. It added that only the wider mathematical community can assess how far the company followed its recommendations.
Those recommendations, issued Sept. 29, ask AI labs to cite earlier papers, write proofs in the style of a traditional paper and disclose prompts, reasoning summaries and computing costs. They also call for results to sit in scholarly repositories that no AI lab controls. OpenAI said it is still exploring community-hosted alternatives to GitHub and will fund workshops on major AI results.
Melanie Matchett Wood, a Harvard mathematics professor and group member, warned in September that AI math output "often includes plagiarism" and usually has no human willing to take responsibility for it. A company spokesperson separately acknowledged that many of the new results are not yet understood by OpenAI's own mathematicians.
The release follows a month of friction between OpenAI and mathematicians. On Sept. 8 the company announced what it called a solution to the Navier-Stokes Millennium Prize problem, which set off a credit dispute with New York University mathematician Tristan Buckmaster.
Three days later, 25 Fields Medalists signed a declaration against using famous open problems as AI benchmarks, and OpenAI said on Sept. 21 that the model had resolved more than 100 other problems.
Read Next:AMD Says AI Demand Is So High It Will ‘Substantially’ Lift Chip Supply In 2027
