Science

Levent Alpöge, the mathematician whose Riemann proof checks out before anyone can read it

Penelope H. Fritz
Add us on Google
Levent Alpöge
BornApril 1, 1992
Long Island
OccupationMathematician
AwardsMorgan Prize

Levent Alpöge kept a private plan as a student: spend decades on the Riemann hypothesis, learning every tool he might one day recognise inside it. The plan ended early. The problem gave way after weeks of all-nighters with Claude, the AI model built by Anthropic, the company where he now works, and what he released was a library of Lean code with no paper at all, because a written version would run past a quarter of a million pages. The people who maintain the shared foundations of formal mathematics now say the statement and the proof are both correct. Alpöge says he has not yet mastered the mathematics his own proof stands on.

Until this year his name meant something narrower. Among number theorists it stood for careful work on rational and integral points on curves, in the arithmetic statistics school of his doctoral adviser Manjul Bhargava. Over the past half-year it has come to mean a run of announcements, often a few lowercase lines on X, in which old problems fell with a machine’s help and everyone else did the checking in public.

Alpöge was born on April 1, 1992, and grew up on Long Island, New York, where he went to Half Hollow Hills High School West in Dix Hills. He holds American and Turkish citizenship. As a teenager he reached the finals of the 2010 Intel Science Talent Search with a computer-science project that analysed MRI scans of blood vessels, and the main-belt asteroid 25898 Alpöge carries his name. At Harvard he took a mathematics degree and a master’s in physics in 2014, summa cum laude, with an undergraduate thesis, supervised by Jacob Tsimerman, showing that the average elliptic curve has few integral points. That thesis won him the 2015 Morgan Prize, the award American mathematical societies give for undergraduate research.

A Churchill Scholarship took him to Cambridge for Part III of the Mathematical Tripos, and Princeton followed, where he finished a doctorate under Bhargava in 2020 with a dissertation on points on curves. After an NSF postdoctoral fellowship tied to Columbia and Harvard, he spent 2021 to 2025 as a junior fellow in the Harvard Society of Fellows. With Bhargava and Ari Shnidman he proved that at most about five-sixths of whole numbers can be written as a sum of two rational cubes, the first bound of its kind. With the same pair and Wei Ho he gave a second proof, published in Inventiones Mathematicae, that Hilbert’s tenth problem has a negative answer over the ring of integers of every number field: no algorithm can decide whether a polynomial equation has solutions there. Peter Koymans and Carlo Pagano had reached that conclusion by another route shortly before.

Then he joined Anthropic as a member of technical staff, and the tempo changed. On July 19, 2026, he posted “hello there the jacobian conjecture is false” above a polynomial map in three complex variables whose Jacobian determinant is the constant minus two but which still sends different points to the same place. Claude had found it. Ott-Heinrich Keller had posed the question in 1939; other mathematicians confirmed the arithmetic within days, and the two-variable case remains open. In August, with Philippe Voinov and Saul Reynolds-Haertle, he reported Hadamard matrices for all twelve orders below 2,000 where none had been known, including 668, the smallest open case since 2005. A counterexample to the Carathéodory conjecture followed, then a hundred-page document claiming a complex structure on the six-dimensional sphere, an answer to Hopf’s problem that others then set about formalising.

With the NYU mathematician Tristan Buckmaster he went after Navier-Stokes, one of the Clay Mathematics Institute’s Millennium Prize problems. In August the pair showed, with a Lean check, that the Euler equations, the frictionless cousin of Navier-Stokes, can blow up in finite time under smooth forcing. On September 8 OpenAI announced its own Navier-Stokes solution along closely related lines, citing their work, and the moment soured into a priority dispute when Buckmaster suggested OpenAI researchers might have seen private notes he had kept in a Codex session. On November 20 the two released a Lean-verified paper claiming blow-up with no outside force at all.

The manner of all this is what unsettles colleagues. The results arrive as posts and website notes, outside peer review, from an employee of the company whose model shares the credit, so each theorem also works as a product demonstration. The Riemann note went up the day after Anthropic’s shares began trading on Nasdaq in the largest stock-market debut on record. Alpöge raised the point himself. He called the timing horrible, said he had messaged chief executive Dario Amodei late on a Sunday night to ask whether he could announce, and wrote that the obvious cynical reading is false. The harder objection is one he also makes himself: he is presenting a complete argument, in his words, before having mastered a massive neighbourhood around its mathematics.

The Riemann hypothesis concerns where the zeros of the zeta function lie, and through them how evenly the prime numbers are spread. Alpöge says his method also settles the Generalized Riemann Hypothesis. He chose the direction and the analogy to a toy problem, by his account, and near the finish line Claude far outpaced him. He posted only Lean, a language in which a computer checks every logical step, judging that a quarter-million pages of generated prose would have been an unreadable travesty. MIT’s Andrew Sutherland confirmed that the code compiles, and Terence Tao described it as an enormous Lean-only proof. On November 21 the Lean and Mathlib communities, backed by the Lean FRO and the Mathlib Initiative, reported that dozens of people had rebuilt the 17-million-line library, that two independent tools matched its theorem to Mathlib’s statement of the hypothesis and that several separate proof checkers had accepted it. “We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct,” they wrote.

A certificate cannot supply the reason the hypothesis is true. Alpöge admits to little mastery of automorphic forms on higher-rank groups, the territory the argument crosses, and expects much shorter proofs to follow. The Clay Institute’s $1 million prize also asks for publication in a qualifying outlet and two years of scrutiny, and a Lean archive on a personal website meets neither condition yet.

His stated next step is the one a mathematician would have taken before any of this: understand the argument, simplify it and write it properly. How much of that paper reads as his, and how much as Claude’s, is the question his next year of work will answer.

Tags: , , , Levent Alpöge, ,

Add us on Google

Discussion

There are 0 comments.