Daily · Article 02 · 4 min read
- Anthropic
Claude and the Riemann hypothesis: What have outside researchers checked?
November 20 author announcement and November 21 external checks
Summary
On November 20, Levent Alpöge released a proposed proof of the Riemann hypothesis developed with Claude. A Waterloo mathematician reported successful machine checks. Clay still lists the problem as unsolved.
Key facts
In a public note released on November 20, Anthropic mathematician Levent Alpöge said he and the AI tool Claude had found a proof of the Riemann hypothesis. He released proof files in Lean, a format in which a computer can check the logical steps.
On November 21, University of Waterloo mathematician Michael Rubinstein published results from checking the release with two verification engines. The Clay Mathematics Institute’s official problem page still read “Unsolved” when checked on November 22 Japan time. Author’s note and proof files · Waterloo verification records · Clay’s problem page

1. Counting primes and locating points on a line
Prime numbers, such as 2, 3 and 5, can be divided evenly only by 1 and themselves. There are four up to 10 and eight up to 20. For very large numbers, mathematicians want an estimate of how many primes there are and a way to bound how far the true count strays from it. In 1859, Riemann approached that question through a calculation rule called the zeta function. An input at which the function gives zero is called a zero. The locations of certain zeros are tied to the error in estimates of prime counts.
A complex number is a point with a horizontal and a vertical coordinate. The horizontal one is its real part. The hypothesis says that all the zeta function’s nontrivial zeros, the class at issue after setting aside a known set of zeros, have horizontal coordinate ½. These invented points show how to read the coordinates.
| Illustrative point | Horizontal coordinate | Vertical coordinate | Where it sits relative to the line at ½ |
|---|---|---|---|
| A | ½ | 2 | On the line |
| B | ¾ | 2 | To the right of the line |
- Illustrative point
A
- Horizontal coordinate
- ½
- Vertical coordinate
- 2
- Where it sits relative to the line at ½
- On the line
- Illustrative point
B
- Horizontal coordinate
- ¾
- Vertical coordinate
- 2
- Where it sits relative to the line at ½
- To the right of the line
The cat asks Q&A
- The researcher
If we find lots of points on the line, have we proved it?
Pocket NotebookThe hypothesis is about every nontrivial zero. Checking a large range leaves open what happens beyond it, so you need an argument that covers the rest. One actual zero of that kind off the line would disprove it.
2. From the mathematician’s idea to a checkable argument
Alpöge writes that he combined ideas in a smaller mathematical setting, then guided the work to a similar setting for the target problem. He credits Claude with proposing a new ingredient in the argument. The public note tells readers where ideas came from. The Lean files let a verifier check how the final theorem follows from definitions and earlier results.
Lean is a language and a tool for checking precisely stated proofs. For a practice example, take the statement “adding 2 to an even number gives an even number.” Write the starting number as twice an integer and show that adding 2 still gives twice an integer. Testing 2 and 4 would give examples, while the argument covers every even number. The diagram shows the roles in checking a published argument.
%% Checking a published argument · editorial diagram
%%{init: {"htmlLabels": false}}%%
flowchart TD
A["Author: releases proof"] --> B["Checker: verifies steps"]
B --> C["Reviewer: checks claim and assumptions"]The cat asks Q&A
- The researcher
If the checker passes it, can I tell whether it answers the original question?
Pocket NotebookThe checker tests the proof against the statement it was given. A reviewer compares that statement with the original problem to see whether, for instance, “all zeros” has become “some zeros.”
3. What the November 21 checks establish
Starting from the author’s release, Rubinstein reports that the theorem under test was the Riemann hypothesis as defined in Lean’s Mathlib library. Lean’s verification engine rechecked all 175,187 modules, and nanoda, an independently implemented engine, checked 1,484,729 declarations on which the main theorems depend without errors. Modules count units of the source material, while declarations count items such as definitions and theorems. Adding those counts would give no meaningful total. He reports that the five main theorems depend only on Lean’s three standard axioms, the starting assumptions the system accepts. Verification logs and statement comparison
| Stage | Recorded status | Evidence |
|---|---|---|
| Author’s release | November 20 publication of the proof claim and Lean files | Alpöge’s note |
| External machine checks | Waterloo verification record reports all four checks complete by November 21 at 19:10 JST | Rubinstein’s records |
| Clay’s status | Listed as “Unsolved” when checked November 22 JST | Official problem page |
- Stage
Author’s release
- Recorded status
- November 20 publication of the proof claim and Lean files
- Evidence
- Alpöge’s note
- Stage
External machine checks
- Recorded status
- Waterloo verification record reports all four checks complete by November 21 at 19:10 JST
- Evidence
- Rubinstein’s records
- Stage
Clay’s status
- Recorded status
- Listed as “Unsolved” when checked November 22 JST
- Evidence
- Official problem page
The cat asks Q&A
- The researcher
Will Clay make a decision as soon as it sees the verification record?
Pocket NotebookThe prize has its own criteria. Clay can consider a solution after publication in a qualifying outlet, at least two years from that publication, and broad acceptance by the international mathematics community. The two-year period is tied to qualifying publication. Millennium Prize rules
To close
What to look for in the next report
When another check comes out, look for the theorem it tested and the assumptions it allowed. A mathematical write-up helps researchers read and discuss why the argument works. This release matters because external researchers have a checkable record of an argument that AI helped develop.
Sources
- Alpöge’s notealpo.ge
- Rubinstein’s recordsmrubinst.github.io
- Official problem pageclaymath.org
- Millennium Prize rulesclaymath.org