Sep 8, 2026 · 1h 5m
AI models push mathematical discovery beyond human persistence
OpenAI Researchers on the Future of Mathematical Reasoning
The episode examines whether reasoning models are becoming collaborators that can generate, test, and extend ideas in areas where mathematics remains poorly understood.
- 1Reasoning models can pursue difficult mathematical ideas, recover from mistakes, and explore multiple paths over long horizons.
- 2AI matched the best known asymptotic bound in a high-dimensional sphere-packing problem through a newly constructed optimal function.
- 3The researchers argue that mathematicians may increasingly focus on interpreting, organizing, and extending machine-generated results.
Don't miss
The guests describe a short AI-generated group-theory proof that identifies a concrete combinatorial obstruction earlier work could not rule out.
The brief
OpenAI mathematicians Mehtaab Sawhney and Mark Sellke describe reasoning models that pursue mathematical ideas, make mistakes, backtrack, and continue working in ways that resemble expert practice.
In sphere packing, a model analyzed a high-dimensional linear-programming method, constructed an optimal function, and matched the best known asymptotic bound.
The conversation then turns to coding theory, where interactive prompting helped push an initial improvement further, raising questions about human direction, model scaling, and experimental design.
A short AI-generated group-theory proof identified a concrete combinatorial obstruction connected to sofic groups and a stronger approximation conjecture.
The larger question is not only whether models can prove results, but whether mathematicians can understand, organize, and develop an expanding body of machine-generated mathematics.
Featuring
Mehtaab Sawhney
GPT-6 Astra
OpenAI
Paul Erdős
Andrew WilesListen to the full episode and explore every guest, topic, and moment on PodLume.