Fermat's Last Theorem now has a Lean proof you can compile
Published on HivePostify by @vixmemon · Sat Sep 05 2026
Andrew Wiles needed 129 pages and a whole year of panic after a reviewer found a hole.
The computer-checked version that dropped on September 4 is 13 million lines of Lean. I do not love that ratio, but I get why it exists.
Anthropic says Claude spent 11 days, mostly on its own, formalizing Fermat's Last Theorem. Not a new proof of the statement. A machine-checkable rewrite of the known path, the Darmon-Diamond-Taylor writeup of Wiles, sitting on Lean's three standard axioms. Kevin Buzzard, who started a human Lean project for this in 2024, called the artefact robust enough to build on. That is the part I trust more than any press headline.
The history still matters. Fermat scribbled the claim around 1637 and said the margin was too small. In 1908 someone put up a prize and got 621 wrong proofs in a year. Wiles lectured in 1993, broke the proof, then fixed it with Richard Taylor in 1995. Jan Bergstra suggested formalizing it a decade later. Humans have been chipping at that for years. The AI run did not invent that stack. It finished the bookkeeping.
Early agent swarms lost the plot. The thing that worked was Prove2Me, a Columbia-built board of theorem statements, so dozens of agents could pick the next lemma without overwriting each other. Failed attempts still left about 7 percent of the non-boilerplate lines. The finished file is more than five times the size of Mathlib, which tells you it is probably ugly. Ugly and checkable still beats "we are 99 percent sure," which is how Thomas Hales's Kepler proof sat for years.
I do not want papers to stop being written for people. I do want referees to get a compile button when a proof is this long. That is the useful part, not the victory lap.