OpenAI claims ten math breakthroughs with next model Astra

The announcement follows OpenAI's May share of an AI-generated disproof of the Erdős unit-distance conjecture, which has already inspired further work. Now, with a selection of ten results, the company demonstrates that its models can tackle diverse, open problems across fields like sphere packing, group theory, and quantum complexity, each previously unsolved for years. The work is part of a broader initiative to provide free access to ChatGPT for academic researchers and reflects OpenAI's stated goal of empowering scientists with tools to accelerate discovery.

The company is careful to address concerns within the mathematical community, including those who signed the Leiden declaration on AI and Mathematics. By releasing Lean certificates and model narrations, OpenAI aims for transparency about how results were produced. They explicitly acknowledge that the arguments were generated by the system, while humans helped prepare manuscripts and formalize proofs.

The stakes hinge on whether the mathematical community will accept and build on these results. If they hold, this could mark a shift in how research is conducted, but the response from mathematicians will shape the trajectory. OpenAI has not claimed the results are verified beyond its own formalization, so independent scrutiny will be the test.