OpenAI Publishes Manuscripts Showing Its Unreleased Model Solving Hard Math Problems

OpenAI has released a set of documents with 722 manuscripts covering 372 result families. They show how an unreleased frontier model solved many long-standing math problems. People had been waiting for weeks to see the details.

The release includes summaries of the model's reasoning and estimates of computing power used. OpenAI says a typical result took about as much compute as ChatGPT Pro thinking for three hours straight. The work is posted in a GitHub repository.

A new group of top mathematicians called AGMAI wants AI labs to share math results through proper academic channels. It says they should reveal the model name, prompts, and compute costs, and not use math results to promote their own models.

OpenAI and rivals like Anthropic have been producing more and more math results, including work linked to a Millennium Prize Problem. AI labs entering math has sparked debates about research ethics and how to credit human mathematicians. The full impact is still to be seen.