Tao: Mathematicians Should Prioritize 'Digesting' AI Proofs Over Speed
量子位 · wechat · 2026-08-20
Faced with the crisis of AI batch-generating mathematical proofs, Terence Tao and Fields Medalist Hong Wang point out that the mathematics community needs to shift from merely 'discovering proofs' to 'digesting proofs'.
Core Perspectives:
- The Complete Proof Process: Tao breaks down mathematical achievements into five steps: generating argument, verifying correctness, peer interpretation, publication review, and knowledge consolidation. AI currently accelerates the first two steps, while the latter three still require deep involvement from mathematicians.
- Importance of Digestion: Tao spent days 'digesting' the AI-assisted proof of the Sendov conjecture. By tracing literature and extracting core identities, he not only compressed the Lean code from 90k to 15k lines but also covered the stronger Phelps–Rodriguez conjecture, proving that the digestion process itself can broaden mathematical results.
- Call for Rule Change: Tao advocates changing the old norm of 'first to prove gets the credit' and elevating the status of 'digesting and organizing proofs'. If an author cannot explain the results at an expert level, the work should not be published even if AI verified it correctly.
Project Palomar: Tao released Palomar, a registry for Lean verification results. It serves as a relay station between verification and publication, recording four time nodes—generation, verification, interpretation, and publication—to determine attribution.
More from AGI Musings
- "An LLM doesn't understand anything" gets the perfect parody reversal: neither do you — VoidStateKate · 2026-10-16
- Nina Schick: Europe's Policy Circle Still Doesn't Believe FSD, Orbital Compute Are Real — NinaDSchick · 2026-10-16
- Meta's Superintelligence Lab Runs 200+ Direct Reports—A Lesson for Flat Multi-Agent Systems — peterjliu · 2026-10-16
- Andy Masley lays out his version of effective altruism, pluralism over utilitarian math — RobbWiller · 2026-10-16
- We're forgetting how much of a miracle GPT-3.5-level intelligence was, and it may cost us — EigenGender · 2026-10-16
- Ex-Anthropic's McDonagh Builds Compute Intelligence Agency to Study Frontier AI Costs — McDonaghMatthew · 2026-10-16