Debate: brute-force Lean proofs beyond LLM context limits may nuke math
doomslide · x · 2026-11-22
doomslide argues we may live in a world where brute-force, sprawling blind Lean proofs — orders of magnitude beyond any LLM's context length — are enough to crack math, though raw scaling seems useless elsewhere unless it becomes a universal math oracle. He adds that GPT-2's charming incoherence still shows up today at the edges of latent space, and it's getting easier to hit with each model iteration.
More from AGI Musings
- Mathematicians weigh in on Riemann Hypothesis breakthrough: a 25,000-page 'wormhole' through existing math — Vjeux · 2026-11-22
- Math paper volume jumped 30% with AI, but creativity didn't budge, compressibility study finds — jamestagg · 2026-11-22
- AI isn't the end of art: a Star Trek thought experiment on how art evolves — dreamwieber · 2026-11-22
- 'Intelligence can turn every person into a giant, or into a number' — Anthropic IPO reflection — markjeffrey · 2026-11-22
- Gary Marcus fires back: Claude Code is actually neurosymbolic AI — GaryMarcus · 2026-11-22
- The 70-year-old questions: Tokyo's CTG anticipated today's AI debates — Merzmensch · 2026-11-22