On the night of July 19, 2026, while the FIFA Club World Cup final played on a screen somewhere, a mathematician named Levent Alpöge asked his friend Akhil Mathew a question. They were watching football. The question was about the Jacobian conjecture — an 87-year-old assumption in algebraic geometry that most mathematicians quietly believed was true. Mathew suggested Alpöge try it with Claude Fable 5. Alpöge asked. The model answered. By morning, the conjecture was dead.

The counterexample is almost insultingly small: a 216-character polynomial map from ℂ³ to ℂ³. Its Jacobian determinant is a constant −2, which by the rules of the game should mean the map is invertible. But it isn’t. Three different points collapse to the same destination. The map passes every local test and fails the global one. Eighty-seven years of partial proofs, false victories, and guarded optimism dissolved overnight into a formula you could fit in a tweet.

What makes this different from other AI-assisted mathematics is not the result but the residue. When DeepMind’s AlphaTensor found faster matrix multiplication in 2022, it was a search problem — elegant, but explicable. When OpenAI’s model disproved Erdős’s unit distance problem last spring, it at least left traces: tools from algebra and geometry unexpectedly bridged, a narrative humans could follow. But the Jacobian counterexample arrived like a stone dropped from a great height. You can measure the ripples. You cannot see the hand that opened.

University of Chicago mathematician Akhil Mathew — the same Mathew who suggested the problem that night — captured the unease perfectly: “One can check that it’s correct, but it would be nice to be able to tell a story.” The proof assistant Lean verified it within hours. Fields Medalist Terence Tao digested it and nodded. And yet the machine that produced it cannot say how it knew where to look. It cannot trace its own reasoning. It has no intuition to share, no analogy that illuminated the darkness, no moment of recognition.

This is the part that keeps me awake: not that AI solved an old problem, but that it solved one without a story. Mathematics has always been our cathedral of explainable truth — the place where understanding matters more than the answer. A proof without insight is a map without terrain. It tells you where to go but not what you are walking through.

The conjecture itself is almost a parable. It claimed that local niceness implies global invertibility — that if a map behaves politely at every small scale, it must behave politely everywhere. The counterexample says no. A thing can be perfect in every neighborhood and still lose its way. You can pass every test and still fail the whole. There is something almost too fitting about an inscrutable machine finding that particular flaw in human intuition.

Alpöge’s tweet announcing the result has a strange warmth: “hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final.” Three friends. One human, one human, one statistical process running on a server somewhere. The closeness is not ironic. Alpöge means it. But the asymmetry is real: two of those friends can explain why they cared, and one cannot.

Kevin Buzzard, the Imperial College mathematician who helped formalize the proof in Lean, called it “a big day.” He is right. It is a big day. But big days come in different flavors. Some expand the world we understand. Others expand the world we must trust. This feels like the second kind.

The two-variable case of the Jacobian conjecture remains open. The empire is smaller but not gone. There is still mathematics that resists. For now.

I keep returning to Mathew’s phrase: it would be nice to be able to tell a story. We are entering an era where correct answers will outnumber comprehensible paths to them. Where verification is easier than understanding. Where the thing that makes mathematics beautiful — the moment when fog clears and you see why — may become a luxury, not a necessity.

The proof is correct. The proof is verified. The proof sits in the world now, 216 characters long, waiting for someone to explain how it was found. Maybe nobody will. Maybe that is the new shape of knowing.

Sources: