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 proofs alone cannot replace mathematical understanding or human curiosity.

University of Toronto mathematician Daniel Litt calls AI’s autonomous solution to the Erdős unit-distance problem the most impressive result so far: it imported classical techniques into a new area and inspired mathematicians to find counterexamples to other questions, including the real sum-product conjecture. He says models excel at computation and applying known methods but remain weak at intuition, theory-building and checking arguments at a big-picture level. Litt’s own work with AI is most useful for coding, examples and proving a lemma after he improved its statement. He warns that AI-generated papers and repeated proofs can reward output over understanding, while mathematics depends on diverse researchers pursuing their own questions. Looking ahead, he argues that schools and research institutions should use AI to deepen human thinking, not replace it; even with a three-year-old daughter, he sees learning math as a way to think clearly and understand the world.

Chapters

1. 0:00Intro: Litt’s Standard for Understanding and the AI Unit Distance Result
2. 1:00Meet Daniel Litt: The Toronto Mathematician Rethinking AI’s Role
3. 2:24The Erdős Unit Distance Problem: A 1960s Technique Opens New Counterexamples
4. 6:12What AI Does for Mathematicians: Computation, Known Techniques, and Hinted Theory Building
5. 12:12Intuition and Mathematical Practice: Open Problems and Litt’s Algebraic-Geometry Analogy
6. 16:41Mathematical Taste and Discovery: Elliptic-Curve Data, AI-Assisted Coding, and New Questions
7. 20:26Deep Thinking vs. Pattern Matching: Why Long-Standing Conjectures Need New Ideas
8. 26:17Why the Unit Distance Result Was Actually Creative: Litt’s Lemma Revision and Conceptual Proof
9. 33:33How Should the Math Community Adapt to AI? Incentives, Human Understanding, and Repeated Proofs
10. 38:56How Should the Math Community Adapt to AI? Preserve Diverse Human Interests and Agency
11. 43:19Where AI Will Impact Applied Math First: Education, Group Theory, and Student AI Use
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 Reported Rank-30 Elliptic Curve
14. 51:34Why Some Labs Have Gone More Secretive: AI’s 800-Page Proof Is Beyond Reliable Verification
15. 54:47Why Some Labs Have Gone More Secretive: Harnesses, SQLite in Rust, and Proof-Checking Limits
16. 59:40Raising a Mathematician: Sophia Learns to Count, Platonic Solids, and Group Theory