OpenAI opens hundreds of mathematics results to outside scrutiny
OpenAI has made 372 mathematical results public while keeping the system that generated them internal. Many arguments have computer-checkable versions, but mathematicians still need to assess their ideas and novelty. The company is withholding exact prompts despite recommendations from an independent advisory group, leaving reproduction of its claimed process out of reach.
Science··Morning
A public repository opens the arguments
OpenAI, the artificial intelligence company, published a collection of mathematical work on October 6 through GitHub, a service for sharing code and documents. The release gives outside researchers access to hundreds of results generated with a system that remains inside the company.[1], [2]
OpenAI says 372 results resolve or substantially advance open questions in mathematics and theoretical computer science. That assessment of their importance is the company's claim.[1]
Computer checks leave novelty to mathematicians
Many arguments have versions written in Lean, a language that lets a computer check logical steps against stated assumptions. Researchers also need to understand the ideas, examine their novelty and determine their mathematical significance. Those tasks extend beyond checking the steps of a formal argument. The released documents can be read even though the system that generated them has not been made available for public use.[1]
Exact prompts remain undisclosed
An independent mathematics and AI advisory group hosted at the Institute for Advanced Study recommended disclosure of the system, exact prompts and computing time. OpenAI is releasing average computing time and further statistics while withholding prompts. The company says almost every result came from a single prompt to a single AI agent, while acknowledging that some may have needed multiple attempts.[1]
Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, said that such single-attempt claims remain unverified until the system is released and others can reproduce the work. Daniel Litt of the University of Toronto welcomed public access to the answers. OpenAI said some newly released work remains unexplained even to its own mathematicians.[1]
