Science Friday Move over, vibe-coding. Vibe-proving is here for math
23 snips
Mar 27, 2026 Daniel Litt, an associate professor studying AI’s interaction with math, and Emily Riehl, a Johns Hopkins category theory researcher, discuss AI’s leap from flubbed arithmetic to contest-level wins. They debate AI-generated proofs, the rise of ‘vibe-proving,’ risks of bogus preprints, and the role of formal proof assistants. The conversation weighs changing workflows, verification standards, and what mathematicians will need going forward.
AI Snips
Chapters
Transcript
Episode notes
AI Is Progressing From School Math To Research Problems
- AI performance in math has progressed from basic arithmetic errors to contest-level proofs and is now being tested on research-level problems.
- Emily Riehl and Daniel Litt note models recently won medals at the IMO and are starting to tackle problems of interest to professional mathematicians.
Problem Selection Drives Mathematical Progress
- Mathematicians care not just about answers but about which problems to ask and why those problems matter.
- Emily Riehl emphasizes that proposing interesting problems and conceptual framing drive the field as much as proving results.
Early Erdős Problem Wins From AI Assistance
- A few Erdős-style problems have been solved with AI involvement: about three fully autonomously and several more with human-AI collaboration.
- Daniel Litt counts roughly three autonomous solves and six or seven human-assisted solutions extracted from literature.