Claude Proves Percolation’s Continuous Phase Transition

In a landmark breakthrough, Anthropic’s unreleased next-gen Claude model generated a complete, Lean-verified formal proof of the long-standing continuous phase transition conjecture in percolation theory — solving a 70-year-old “holy grail” problem in probability theory.

⚡ A Silent Revolution on GitHub

On August 30, 2026, Fields Medalist Hugo Duminil-Copin published a reflective blog post:

“In our field, the most famous conjecture falling to the bulldozer (AI) is likely only a matter of time.”

Within hours, the mathematical world was stunned: a new GitHub repository appeared — authored by Anthropic engineer Justin Leder — containing fully automated, machine-generated Lean code that formally proves the 3–10 dimensional continuous phase transition conjecture in percolation theory.

No press release. No official announcement. Just clean, rigorously verified code — and the quiet end of a decades-long quest.

Duminil-Copin's reflection

🔍 What Is the “Holy Grail” of Probability Theory?

The percolation continuous phase transition conjecture, first posed in 1957 by Broadbent and Hammersley, asks: At what critical probability $p_c$ does fluid suddenly shift from localized droplets to infinite connected clusters in a d-dimensional lattice?

This mirrors real-world phenomena:
– Coffee brewing through a filter
– Oil seeping through porous rock
– Network connectivity in random graphs

Percolation visualization

While high dimensions ($d \geq 11$) were solved using mean-field approximations, and 2D benefited from conformal symmetry, the 3–10 dimensional range remained mathematically impenetrable — a “sighing wall” where geometry, topology, and stochasticity collide chaotically.

🧠 How Did AI Crack It? Not Brute Force — But Strategic Synthesis

Contrary to assumptions, this was not computational exhaustion. Instead, Claude performed cross-domain logical synthesis:

Leveraged the pivotal 2024 paper by Kozma & Nitzan (arXiv:2401.12397) as a formal scaffold
Constructed a multi-thousand-line Lean proof under strict formal verification
Combined known analytical inequalities, auxiliary function design, and topological bounding in a novel, human-unanticipated sequence

GitHub proof repo

As Penn mathematician Ahmed Bou-Rabee confirmed:

“I generalized their proof in one day — something I’d spent eight years failing to achieve. AI didn’t replace insight; it amplified it beyond human cognitive bandwidth.”

🌐 Global Reactions: Elation, Disorientation, and Reflection

Benedikt Jahnel (TU Braunschweig): “If a human proved this, they’d win the Fields Medal. Now — AI crossed the finish line.”
Gil Kalai (Hebrew University): “If verified, this is an extraordinary breakthrough.”
Gady Kozma (Weizmann Institute): “We await Anthropic’s human-readable explanation before comment.”

Mathematicians' reactions

🌅 Beyond Proof: The New Role of Mathematicians

This milestone doesn’t signal the end of mathematics — but its evolution.

Duminil-Copin’s own work exemplifies the deeper truth: the journey matters more than the destination. His failed attempts on the conjecture led directly to his Fields Medal-winning results in statistical physics.

Just as AlphaGo reshaped Go strategy, or deep learning revolutionized protein folding, AI shifts the mathematician’s role from:
– ✅ Manual derivation → 🧭 Problem framing & conceptual navigation
– ✅ Verification labor → 🤖 Orchestrating reasoning agents
– ✅ Isolated insight → 🌐 Interpreting and contextualizing machine-discovered truths

Historical context