What Today’s Best Models Still Can’t Do in Math

Video summary

Daniel Litt praises AI’s Erdős unit-distance result but warns that proof generation is not the same as mathematical understanding.

a16z’s Lisha Li interviews University of Toronto mathematician Daniel Litt about AI’s growing role in mathematics. Litt calls the AI-generated solution to the Erdős unit-distance problem his favorite autonomous result: it brought a classic technique from another field into the area and helped mathematicians find counterexamples to other conjectures. Still, he says frontier models excel mainly at computation and applying known methods, while intuition, theory-building, and judging whether an argument makes sense remain difficult. AI has helped his work with coding and examples, and it quickly proved a stronger lemma after he discovered the right formulation. Litt warns that AI-generated papers can reward paper-counting over understanding, citing three near-identical proofs and an 800-page claimed proof he considers implausible. He argues that mathematics needs diverse human curiosity and a strong education pipeline; the goal is understanding, not simply producing papers.

Chapters

1. 0:00Intro: Litt on mathematical understanding and AI’s unit distance result
2. 1:00Meet Daniel Litt: University of Toronto mathematician on evolving AI views
3. 2:24The Erdős Unit Distance Problem: AI result sparked counterexamples to other conjectures
4. 6:12Working mathematicians’ AI tools: computation is strong, intuition remains limited
5. 12:12Intuition and mathematical practice: Litt’s open-problem method and algebraic-geometry analogy
6. 16:41Intuition and AI’s role: elliptic-curve data, research questions, and parallel coding
7. 20:26Deep thinking vs. pattern matching: new theory is needed for long-standing conjectures
8. 26:17Why the Unit Distance Result Was Actually Creative: Improved Lemmas and Conceptual Insight
9. 33:33How Should the Math Community Adapt to AI? Incentives, Low-Quality Papers, and Diverse Exploration
10. 38:56How Should the Math Community Adapt to AI? Preserve Human Direction and Broad Research
11. 43:26Where AI Will Impact Applied Math First: Education, Group Theory, and Computer-Based Work
12. 46:24Taking Advantage of AI Without Losing the Craft: Deepen Understanding, Not Just Output
13. 49:21Comparing Anthropic vs OpenAI in Math: Claude Fable's Rank-30 Elliptic Curve Claim
14. 51:34Why Some Labs Have Gone More Secretive: Verification Limits and an 800-Page AI Proof
15. 54:47Why Some Labs Have Gone More Secretive: Proof Harnesses, SQLite in Rust, and Human Error-Checking
16. 59:40Raising a Mathematician: Litt's Three-Year-Old, Platonic Solids, and Learning Math in an AI Era

What the full analysis includes

16 chapters · 84 key points · 16 section summaries · mind map

The remaining 81 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →

This video is 1.1 hours. The Free plan analyzes videos up to 30 min; plans from Plus cover videos up to 2 hours.