Lean and AI proofs: how machines check a math claim

10 steps · about 8 h 45 min · Intermediate

  1. On the dangers of stochastic parrots: Can language models be too big? (doc, Emily M. Bender, University of Washington, 15 min)

    Start with the original 'stochastic parrot' argument so you know what the skeptics claim language models do.

  2. Visualizing the Riemann zeta function and analytic continuation (video, 3Blue1Brown, 25 min)

    Gives you a visual grasp of the zeta function and what a Riemann claim is about, before you look at how claims get checked.

  3. Learn Lean (includes the Natural Number Game) (interactive, Lean FRO, 2 h)

    Lean's own page recommends the browser-based Natural Number Game for beginners. It lets you write proofs that a machine checks, with nothing to install.

  4. Learning Lean 4 (guide, Lean community, 20 min)

    This map of games, textbooks and courses helps you choose what to study after the first game.

  5. Mathematics in Lean (repo, Lean community (GitHub), 3 h)

    Exercises that move from toy proofs to real undergraduate mathematics, using Mathlib, the library of already-formalized math.

  6. Machine-Assisted Proof (article, Terence Tao, Notices of the AMS, 40 min)

    A leading mathematician explains how proof assistants, SAT solvers and LLMs fit together in research.

  7. AlphaProof and the IMO (blog, Greg Burnham, Lemmata, 20 min)

    Shows concretely how a model proposes proof steps in Lean and a verifier decides whether they count.

  8. How Terry Tao Became an Evangelist for AI in Math (article, Quanta Magazine, 20 min)

    A narrative case study of one mathematician formalizing real proofs in Lean with AI help, including what turned out easy and hard.

  9. Olympiad-level formal mathematical reasoning with reinforcement learning (doc, Nature (Google DeepMind), 1 h)

    The primary source on training an AI against a Lean checker, for readers who want the technical details.

  10. Terence Tao on AI in mathematics (and beyond) (article, Terence Tao (web page), 25 min)

    End with a well-informed view on how to judge AI math claims, including when to ask for a Lean verification.