OpenAI Releases 722 Math Papers Written by Unreleased Model
The manuscripts cover 372 result families, and the advisory group AGMAI said they include solutions to hundreds of open questions.
OpenAI has released a batch of 722 manuscripts containing solutions to long-standing mathematics problems, all generated by a frontier model the company has not yet released, according to The Verge.
The manuscripts cover 372 result families, which group related papers together. AGMAI, the Advisory Group on Mathematics and Artificial Intelligence, said the release includes solutions to hundreds of open questions.
The published materials include summaries of the model's reasoning, estimates of the compute used, and statistics on the number of problems attempted. OpenAI claims the average result used the equivalent of three hours of ChatGPT Pro thinking.
OpenAI published the results in a GitHub repository that also sets out protocols for paper revisions and citations. The company said it is continuing to explore other community-hosted alternatives for the release that meet the committee's guidelines.
Earlier commitments on disclosure
In September, OpenAI said its model had resolved more than 100 long-standing open problems across most areas of mathematics. AGMAI published its first recommendations in late September, urging AI labs to release mathematical results promptly through established academic channels and to disclose model names, prompts and compute costs, and asking companies not to treat such releases as marketing for their models.
Quick answers
How many manuscripts did OpenAI release?
OpenAI released 722 manuscripts covering 372 result families.
Which model produced the solutions?
An unreleased frontier model, according to OpenAI.
Where were the results published?
In a GitHub repository that includes protocols for paper revisions and citations.