Inside OpenAI’s Breakthroughs in Mathematical Reasoning

影片總結

OpenAI mathematicians Mehtaab Sawhney and Mark Sellke explain Astra’s sphere-packing bound and non-sofic group proof.

a16z’s Lisha Li talks with OpenAI mathematicians Mehtaab Sawhney and Mark Sellke about how GPT-5 changed their view of AI mathematics. They say models can find relevant literature, choose promising approaches, backtrack, and persist through technical details that might make a human abandon an idea. Their examples include Astra’s sphere-packing result: it determines the asymptotic behavior of a linear-programming bound, improving understanding of high-dimensional packing, where exact answers are known in only five dimensions, including 8 and 24. Astra also improved bounds for spherical and binary codes, using representation theory and symmetry. The mathematicians then describe its proof that non-sofic groups exist—a roughly 15-page group-theory argument, far shorter and less technically sprawling than a related 250-page disproof of the Aldous–Lyons conjecture. They expect AI to speed up results while making human explanation, absorption, and connecting discoveries increasingly important; even exponential progress may leave problems such as P versus NP beyond reach.

章節

1. 0:00From Practicing Mathematicians to OpenAI: Mark Sellke, Mehtaab Sawhney, and the IMO Gold Medal
2. 2:43Why GPT-5 Was the Conversion Moment: A Five-Minute Search of Erdős Problems Found a Reference
3. 4:21Beyond Search and Connections: Unit-Distance Proofs Show AI’s Persistence and Mathematical Judgment
4. 9:51Reasoning Traces: Models Reconsider Failed Paths and Reassess Their Promise
5. 11:44Why Math Papers Are a Poor Training Set for Real Mathematics: Motivation, Backtracking, and Reasoning Traces
6. 16:20The Astra 10-Problem Set: Sphere Packing, Hexagonal Lattices, and the E8 and Leech Lattices
7. 21:16The Astra 10-Problem Set: Astra Improves the Sphere-Packing LP Bound to About 2 to the Minus 0.61d
8. 26:18Astra’s Sphere-Packing Bound: An Exact LP Result, Then Binary Codes
9. 31:38Astra’s Representation-Theory Bounds for Spherical and Binary Codes
10. 36:17Astra’s Task-Oriented Prompting: Asking It to Push Code Bounds Further
11. 40:01Taste as Useful Judgment: Pruning Search and Stepping Back
12. 43:42Astra’s Non-Sofic Group Result: Groups, Symmetry, and Finite Approximation
13. 48:57Sofic Approximation: Integers Modulo n and the 15-Page Non-Sofic Proof
14. 53:35Astra’s Short Group-Theory Proof: A Combinatorial Obstruction and Follow-Up
15. 57:32How Should the Math Community Adopt AI? Models Summarize Proofs and Broaden Participation
16. 1:00:01Empirical vs Theoretical Math & the Positive Vision: Shared Understanding, P vs. NP, and Faster Applied Math