OpenAI Podcast Episode 17 - What happens now that AI is good at math?
Apr 28, 2026
Ernest Ryu, applied mathematician who used ChatGPT to help solve a 42-year-old problem, and Sébastien Bubeck, OpenAI researcher bridging math and ML. They discuss how AI leapt from basic arithmetic to research-level math. They talk about using models for long workflows, automating research, literature search, verification, and the risks of shallow understanding.
AI Snips
Chapters
Transcript
Episode notes
How ChatGPT Helped Solve A 42-Year Problem
- Ernest Ryu solved a 42-year open problem by iterating with ChatGPT over three days while acting as a verifier.
- He corrected mistakes, steered approaches, and validated the final proof, turning ~12 hours of focused chat work into a publishable solution.
Why Math Is A Perfect AI Benchmark
- Mathematics is an ideal benchmark because questions are unambiguous and answers are verifiable.
- That verification discipline forces models to think long and consistently, correcting single-point failures that would otherwise invalidate whole arguments.
From Literature Search To New Mathematical Discoveries
- Large models first excelled at deep literature search, finding existing but hidden solutions across fields.
- That capability led from rediscovering Erdos problem answers to generating more than ten novel, publishable combinatorics results months later.