Summary
Hosted by a16z, this interview brings together Mark Selki and Matab Swani — two practicing mathematicians who joined OpenAI after witnessing GPT-5’s capabilities firsthand — to discuss how AI is transforming mathematical research. Their entry points: GPT-5’s IMO gold medal performance last summer and its ability to instantly locate references to open problems that human researchers had spent hours failing to resolve, including open entries in the Paul Erdős problem collection.
The mathematicians walk through specific recent results, including OpenAI models making progress on high-dimensional sphere-packing — a problem connected to the linear programming bound and Maryna Viazovska’s 2022 Fields Medal. They explain how models can now both construct a bounding function and prove that no better function exists, moving beyond numerical conjecture to genuine proof. The conversation traces how GPT-5 was given a high-dimensional linear program and, without further guidance, produced a result that resolves a longstanding conjecture.
The discussion also addresses limits: problems like P versus NP likely remain permanently beyond reach, and the community is still calibrating how to integrate AI as a collaborative proof assistant rather than a black-box oracle. For anyone tracking the frontier of AI reasoning, this interview offers rare firsthand perspective from researchers operating at the intersection of professional mathematics and frontier model development.
📺 Source: a16z · Published September 08, 2026
🏷️ Format: Interview
