OpenAI publishes 722 AI-generated mathematics manuscripts from an unreleased model

OpenAI has published 722 mathematical manuscripts, organised into 372 families, in a public GitHub repository under an Apache-2.0 licence, according to Unite.AI. The results were produced by an internal AI model that has not been released or named, and span number theory, combinatorics, theoretical computer science and mathematical physics. The lab posed roughly 4,000 open problems to the model, with each result taking about three hours of ChatGPT Pro-level reasoning on average.

Many manuscripts come with formalisations in Lean, but not all do. OpenAI warns that some unformalised results could contain issues, and says it will fix issues and add formalisations over time. It released abridged summaries of the model’s reasoning for 10 results.

The release follows September 29 recommendations from the Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, which asked AI labs to stop testing advanced mathematics on proprietary models and not to treat results as marketing. Scientific American reported that OpenAI shared average compute time and statistics but no prompts; an OpenAI spokesperson said the company is taking the group’s guidelines seriously but is not bound by them.

An arXiv audit reviewed 10 results announced on August 1 and found no confirmed substantive error in a principal result, though review depth varied; that audit did not cover the October manuscripts. Reception has been mixed, with scepticism about verification and academic credit. Coverage also came from RuntimeWire, AIStockWire, The News, Archynetys and Times Now.

Read next: OpenAI coverage