OpenAI Is Pissing Off a Bunch of Mathematicians—Again

At an OpenAI meeting, mathematicians responded with a mix of excitement and apprehension, recalls Northwestern University mathematician Bryna Kra. The attendees suggested OpenAI publish papers explaining the work so that mathematicians could absorb and digest the results, according to Kra. However, OpenAI has ignored this request and plans to release hundreds of referenced results on GitHub, people familiar with the plans tell WIRED.

OpenAI announced it had begun training a new internal model to resolve more than 100 long-standing open problems across most areas of mathematics. Some leading mathematicians feel the company has not learned from previous controversies and sees the field as a playground for OpenAI and its rival Anthropic to showcase their models ahead of their initial public offerings.

A notable case occurred in September when OpenAI deployed thousands of AI agents to solve a Millennium Prize problem, leading to accusations of front-running work done by a mathematician and his Anthropic collaborator. OpenAI has formed an advisory group to determine how it assesses and communicates new results, but mathematicians say this has not resulted in meaningful progress.

Many are frustrated that OpenAI and Anthropic continue to release results through blog posts rather than scientific papers, making it harder for others to verify and often excluding prior work from other mathematicians. Some have set up new tools like Hexagon and Palomar in response to the rise of machine-assisted proofs, but these efforts have not changed the company's behavior or aligned with the Leiden declaration, a call by over 4,000 mathematicians for AI companies to meet mathematicians' standards.