Dietterich: Lean proof checker feeds LLMs new info, so proofs aren't interpolation
tdietterich · x · 2026-10-11
In the ongoing debate over whether LLMs can produce genuinely new mathematics, ML pioneer Thomas Dietterich counters the classic information-theory argument that computation cannot create more information than it receives: the Lean proof checker sits outside the LLM and provides new information by indicating which proposed proofs are correct. Even if the checker were treated as internal, he argues, its implicit knowledge is merely made explicit through computation — still not interpolation.
Related event: Dietterich: Lean-checked LLM proofs are not mere interpolation(2 posts)→
More from AGI Musings
- Remembering David McAllester, who predicted AI in math back in 2012-14 — _onionesque · 2026-10-11
- A product owner says one night with Claude Max matched a person-year of work by a 10-person team — Exact_Knowledge5979 · 2026-10-11
- AI now tops 10% of world goods trade, and curbing it would be disastrous — pmddomingos · 2026-10-11
- 'Alien math': AI proofs are correct but think in ways no human would — imjustnewatai · 2026-10-11
- Spencer Greenberg floats a contrarian hypothesis: can AI slop save the internet? — charles_irl · 2026-10-11
- "You're either in the last generation to die of old age or the first that lives semi-forever" — rand_longevity · 2026-10-11