OpenAI releases manuscripts containing mathematical problem solutions

OpenAI has released a large collection of manuscripts produced by an unreleased advanced model, offering solutions to hundreds of open questions in mathematics. The batch consists of 722 manuscripts that cover 372 result families, which group related papers together.

Release details

The collection includes summaries of the model's reasoning, estimates of the computing power used, and statistics on the number of problems attempted. According to OpenAI, the average result required an amount of compute equivalent to three hours of thinking by ChatGPT Pro.

Community response

An independent advisory group known as AGMAI, composed of elite mathematicians, was formed to help communicate these results responsibly. In September, the group urged artificial intelligence laboratories to release mathematical findings promptly through established academic channels. They also asked companies to disclose specific details such as model names, prompts, and compute costs, while avoiding the use of such releases purely as promotional marketing vehicles.

OpenAI published these results in a GitHub repository, establishing protocols for paper revisions and citations. The company stated that it plans to explore alternative community-hosted platforms and continue improving the quality of its mathematical exposition and paper presentations in future releases. Mathematicians are currently reviewing the texts to assess their full impact.