Anthropic AI Proves Longstanding Percolation Theory Conjecture

Summary

Anthropic has released a proof generated by a large language model for a longstanding conjecture in percolation theory, a branch of probability theory that studies when local connections form an infinite network. The problem concerns the threshold at which the probability of belonging to an infinite connected cluster changes from zero to a positive value. Exact thresholds are known for only a small number of networks, including the two-dimensional square lattice, where Harry Kesten showed that the boundary is one half. Mathematicians had already established that the transition is continuous in one dimension, in the two-dimensional square lattice, and in sufficiently high-dimensional generalizations. The unresolved cases were square lattices in dimensions three through ten. Anthropic’s model reportedly showed that the transition is continuous in all of those dimensions, confirming a result many mathematicians had suspected for decades. The development came days after Fields Medalist Hugo Duminil-Copin warned that AI might solve the field’s most famous conjecture before humans did. Mathematician Benedikt Jahnel welcomed the proof but said its completion by AI also created disillusionment. The article argues that the episode could force mathematics to reconsider how it values proofs, while emphasizing that conjecture-building, understanding, and translating results into human knowledge remain important parts of mathematical work.