A Solo Researcher Spent 1.5 Years Trying to Prove the Riemann Hypothesis with Claude and Lean, and Failed
Independent researcher Vladimir Radkevich spent 1.5 years trying to prove the Riemann hypothesis with Claude and Lean, but a growing factor ate up every gain, much like speeding up one operation in code only to see data preparation slow down by the same amount. According to him, on November 20, 2026, Levent Alpöge proved the hypothesis, also working with Claude; Lean and Mathlib confirmed the proof. The takeaway: an AI chain needs a strict checker, or it will produce convincing calculations that lead nowhere.
AI-processed from Habr AI; edited by Hamidun News
It happens that you spend a week optimizing code, every change looks reasonable, the tests are green, and the runtime doesn't budge. Something quietly eats the entire gain. That's roughly what 1.5 years of work looked like for independent researcher Vladimir Radkevich, who, together with the Claude and Codex models and the Lean proof-checking system, tried to prove the Riemann hypothesis. He didn't succeed. But he did write a detailed breakdown on Habr of what an "almost working" idea looks like and why people and AI are so easily fooled by their own calculations.
A caveat up front: everything below rests on the author's account, which we did not independently verify. The original Habr text is the first part of a three-part series, and our material covers only that part, so what we say about the second and third installments comes only from the author's announcement.
Where the Gain Disappeared
The author describes a typical failure. To push the error to zero, he widened the computation window: the wider the window, they reasoned, the more accurate the estimate. The derivations were correct, the auxiliary inequalities were correct, and some of it was even checked on a computer. Then it turned out that as the window grew, so did a multiplier, and it consumed exactly as much as had been gained. Several pages of convincing calculations, and at the end, zero progress.
For a developer, this is a familiar picture: you speed up one operation, but preparing the data for it gets slower by the same amount. The local gain is real; the overall gain is zero. That's why the author advises asking any new idea an uncomfortable question up front: where is its cost hidden, and will that cost turn out to equal the gain?
How a Hypothesis Becomes an "Almost ML"
Problem
The Riemann hypothesis states that all nontrivial zeros of the zeta function lie on a single vertical line in the middle of the critical strip. Computers have verified trillions of zeros, and all of them sit on that line. But you can't replace the word "every" with a large number of examples: the next zero could be the first one that breaks the pattern.
Radkevich's project is called NymanBeurling, after the Nyman–Beurling criterion. It turns the question about zeros into a question about approximation. From the fractional parts of numbers, one builds a family of "sawtooth" functions on the interval from 0 to 1.
Think of them as short sound waves: you can turn up the volume and flip the sign of each one. The task is to add them together into a flat line, the constant one, such that the least-squares error tends to zero. In the Báez-Duarte strengthening, the criterion says the hypothesis is true if and only if this error goes to zero.
To the ear, this sounds almost like machine learning: tune the coefficients and watch the error fall. But why adding up waves should know anything about the zeros of the zeta function, the author explains in a separate argument. If a "wrong" zero existed, there would be a probe that cannot see any individual building block but can see the target.
Such a probe is stable, which means you could not get close to the target at all.
Simple mistakes happened too. Through an oversight, a function identically equal to zero was included in a "linearly independent basis," and a zero row appeared in the matrix.
What Lean Did, and What to Take Away
In Lean, a statement is written as a type, and a proof as a value of that type. A dedicated kernel checks that the value was built according to permitted rules. For a programmer, this is the difference between an interface and an implementation, except that the implementation must serve as evidence of truth, not as a function that returns true. If Lean accepts a proof, it holds for all objects satisfying the conditions; it cannot be broken "on the next input."
But the checking has a boundary: someone still has to understand the meaning of the definitions and conditions. The author immediately warns that this is something they "verified on their own skin."
The roles were divided as follows. Radkevich chose the questions, argued about constructions, and tried to tie the results into a whole. Claude and Codex proposed arguments, wrote code, looked for errors in each other's work, and also made mistakes. They started in August 2025, and the main chronicle falls in January–October 2026.
The ending of the story, according to the author, is this: on November 20, 2026, another person, Levent Alpöge, proved the hypothesis, also in pair with Claude. He published a Lean library of roughly 17 million lines, and on November 21 the Lean and Mathlib communities issued a joint statement confirming that the proof and its formulation had been verified. Radkevich himself did not repeat the machine check and does not claim to have understood all 17 million lines.
The practical lesson for anyone building agents and long chains of models: without an external, strict arbiter, they produce convincing pages with no result. In this story, the formal verifier worked like a rigorous test suite. It doesn't invent a solution, but it honestly shows what was actually achieved. The project's repository is currently closed; the author grants access on request.
- Nov 22, 2026A Backend Developer With No Unity Experience Built a 3D Game With an AI Agent: Subscriptions Cost $112
- Nov 22, 2026OpenAI's Chief Scientist Proposes an AI Speed Limit: GPT-4 Perplexity at 0.91, Now Below 0.8
- Nov 22, 2026DeepSeek Veteran Open-Sources Linger: A Local AI Agent That Responds Up to 5x Faster
- Nov 22, 2026Anthropic Shares Fell 8.5% in a Day, Nearly Back to the $83 IPO Price
Key Questions Answered
Where the Gain Disappeared?
The author describes a typical failure. To push the error to zero, he widened the computation window: the wider the window, they reasoned, the more accurate the estimate. The derivations were correct, the auxiliary inequalities were correct, and some of it was even checked on a computer. Then it turned out that as the window grew, so did a multiplier, and it consumed exactly as much as had been gained. Several pages of convincing calculations, and at the end, zero progress.
How a Hypothesis Becomes an "Almost ML" Problem?
The Riemann hypothesis states that all nontrivial zeros of the zeta function lie on a single vertical line in the middle of the critical strip. Computers have verified trillions of zeros, and all of them sit on that line. But you can't replace the word "every" with a large number of examples: the next zero could be the first one that breaks the pattern. Radkevich's project is called NymanBeurling, after the Nyman–Beurling criterion. It turns the question about zeros into a question about approximation.
What Lean Did, and What to Take Away?
In Lean, a statement is written as a type, and a proof as a value of that type. A dedicated kernel checks that the value was built according to permitted rules. For a programmer, this is the difference between an interface and an implementation, except that the implementation must serve as evidence of truth, not as a function that returns true.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.