OpenAI publishes math results from unreleased model

OpenAI has published what it calls "major results" in mathematics that came from an internal frontier model it has not yet released. The papers are in a GitHub repository that includes protocols for paper revisions and citations.

The excerpt available does not say what the results are, which problems they address, or how mathematicians have judged them.

The repository includes formal versions of many of the proofs in Lean, a programming language that lets a computer check mathematical proofs. OpenAI says it will add more formalizations as it gets them. The company is also publishing details on how the results were produced:

• 10 summaries of the model's reasoning
• compute estimates, expressed as equivalent ChatGPT Pro usage
• statistics on how many problems the model attempted

OpenAI says the average result took about as much compute as three hours of ChatGPT Pro thinking.

The company says it is still looking at other community-hosted places to publish this release that would meet "the committee's guidelines." The excerpt does not name the committee. For future releases, OpenAI says it will improve citations, mathematical exposition and presentation.

OpenAI also plans to fund workshops, conferences and special programs on understanding major results produced by AI. It says details will come "in the near future."

The company says it is "working to responsibly release the model that produced these results" but gave no timeline or name. It says it will keep testing its internal frontier models on mathematics and other sciences and will update its standards for disclosing major scientific advances based on community feedback.