A Mathematician Predicted AI Would Beat Him to His Field's Biggest Problem. Days Later, It Did.

Hugo Duminil-Copin spent a decade on percolation theory, the math of when a porous network suddenly becomes open from one side to the other. On August 30, 2026, the Fields Medalist wrote on the Proofs and Prompts blog that his field's most famous conjecture would "only" fall to the "bulldozers" of AI. An LLM-generated proof of that exact conjecture appeared in a public Anthropic repository four days later.

Four days separate the prediction from the proof.

Percolation, in plain terms, is the study of phase transitions in random networks. Below a critical probability, liquid seeping through a porous rock almost surely stays trapped in finite pockets. Above it, an infinite connected channel opens up almost surely. The conjecture asks whether, at the precise threshold where that flip happens, the probability of an infinite open channel is exactly zero. For decades, mathematicians could prove vanishing at the threshold for two dimensions but not for higher ones. The percolation README names Justin Leder as author and claims vanishing at the critical point for nearest-neighbour Bernoulli bond percolation on ℤ^d for every d ≥ 2, with dimensions 3 through 10 as the previously open cases.

That is the claim, not an independently refereed result. The same README states the work has not been independently refereed and asks readers to verify that the Lean file states the intended theorem. A separate AUDIT.md records a clean build, an axiom audit limited to standard Lean machinery, and a successful comparator check. These are publisher-recorded checks; none has been independently rerun by the broader community.

Duminil-Copin won the Fields Medal in 2022, the honor awarded every four years and often described as mathematics' equivalent of the Nobel, for work that included contributions to percolation theory. His August 30 essay was not framed as a literal forecast. It argued that failed attempts at hard conjectures often generate the next generation of useful math. The bulldozers line was a metaphor about how AI might outpace human work on the field's centerpiece. The proof arriving days later turns the metaphor into a measurable event.

According to Scientific American, which reproduced the story from Spektrum der Wissenschaft with AI-assisted translation and editorial review, mathematician Benedikt Jahnel called the proof "ambivalent" and described it as "disillusioning." His reaction mirrors the field's. The achievement is real on its face: a long-standing, Fields Medal-caliber problem in probability theory, formally stated, formally proved, and lodged in a public repository. The disorientation is also real. The bottleneck in mathematical discovery has historically been finding the proof. For an LLM, the bottleneck is upstream: stating the theorem precisely, encoding it in a proof assistant, and verifying the model output. Each step still requires a human mathematician.

Duminil-Copin's essay predates the proof and is not a review of it. That detail matters because the prediction/proof reading collapses if the timestamps are read the wrong way. He was not prophesying a specific release; he was describing a pattern. The Anthropic repository is the pattern arriving on schedule. The next move belongs to the small set of mathematicians who can read the encoded theorem against its mathematical intent. The watch item is whether one of them publishes a refereed assessment before the conjecture fades into the field's permanent record.