Eliezer Yudkowsky wrote on November 20, 2026 that "The Riemann Hypothesis has fallen," and called it the result most often named in advance as the canary for dangerously advanced AI. He was reacting to Anthropic mathematician Levent Alpöge's claim that he and Claude proved the Riemann Hypothesis. The proof checks out in Lean so far. The risk argument built on top of it is a separate matter, and it is an opinion. This post covers the whole story: what happened, what is verified, what Yudkowsky argues, how others replied, and what it changes for people who build with AI.
For the day-by-day verification log, see our claimed vs verified tracker on the Alpöge proof. This post is about the argument and the aftermath.
TL;DR: the story in one table
| Question | Short answer |
|---|---|
| What happened? | Alpöge, an Anthropic mathematician, posted a Claude-assisted Lean proof of the Riemann Hypothesis on November 20, 2026 |
| Is the proof checked? | Largely. MIT's Andrew Sutherland and others compiled it in Lean; comparator and statement checks were still finishing |
| What did Yudkowsky argue? | The methods should generalize to anything you can definitely test, including making AI better at making AI |
| Is that a finding? | No. It is his inference, stated as a warning |
| What do critics say? | Nothing harmful has happened yet, and math is not physical action |
| What should builders do? | Treat checkers as the scarce part, and put guardrails on what agents can do |
Agents that act on your behalf need rules. See how explainx helps secure your agents →
What happened on November 20?
Alpöge published an essay and Lean code on his personal website. Scientific American quoted his statement: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He also claims the generalized version, which covers a whole family of related functions.
TechCrunch's Tim Fernholz reported that Alpöge messaged Anthropic CEO Dario Amodei the previous Sunday night to ask how to announce the result. The essay opens, per TechCrunch, with the line "The precise timing is horrible." That timing matters because Anthropic had become the first publicly traded frontier AI lab the same week. Our coverage of the listing is in Anthropic IPO priced at $83.
Alpöge told TechCrunch he felt guilty about releasing a proof without an explanation people can read, and that he could not wait. He has committed to work toward a human-readable version. Commenters on Hacker News reported the full written argument runs past 250,000 pages. No second source confirms that number, so treat it as a reader's report.
Nature and Scientific American both covered the claim the same day. Science asked the question many readers have: can humans understand the answer?
Who has checked the proof?
Scientific American's Clara Moskowitz reported that Sutherland and other independent researchers "have now separately compiled Alpöge's result in Lean without issues." The article quotes Sutherland directly: "The Lean proof has been compiled and independently rechecked by the Lean kernel with only the three standard axioms," and "Honestly the chance that anything is fishy here is approximately zero."
Six proof tiles with two green checks, showing how a Riemann Hypothesis proof is checked step by step in Lean
Two checks were still open at the time of writing:
- The statement check. Kevin Buzzard of Imperial College London asked whether the theorem in the Lean code corresponds to the actual Riemann Hypothesis. A proof can compile perfectly and still prove a slightly different statement.
- The comparator run. Buzzard said on the Lean community chat that a successful comparator run was still awaited. It confirms the code does not redefine symbols or exploit a bug in the checker, and typically takes hours.
At least one specialist has also started engaging with the mathematics itself. TechCrunch reports that University of East Anglia number theorist Chris Birkbeck posted a technical comment on an automorphy-lifting step in the Lean chat, which drew about twenty reactions.
Not everyone is convinced about the framing. Science reported that number theorist Andrew Granville called the released document "garbled," and linked the release timing to Anthropic's stock. Alpöge rejected that, saying "There is an obvious cynical view, to do with the timing, but it is false."
One more fact belongs in the record. The same day, Stanford's Ravi Vakil posted a draft by postdoc Alessandro Masullo claiming a proof of the Generalized Riemann Hypothesis with, in Vakil's words, "no AI." Masullo said it took 12 years. Our write-up is here.
What exactly is the canary argument?
Yudkowsky's post, published at 11:50 PM on November 20 and viewed about 708,000 times at last count, has three moving parts.
Part one: the canary. He says RH was "the singular math result that had been named, in advance, as the obvious coal mine canary for dangerously advanced AI, over the last three decades." A canary in a coal mine dies before the miners do. The claim is that this problem was a pre-agreed warning sign.
Part two: the generalization. He writes that the methods used "should generalize to 'make AI very good at anything you can definitely test,' and unfortunately that category includes 'make AI better at making AI.'" This is the technical heart of the argument. A Lean proof is the cleanest possible example of a task with a definite test: the checker says yes or no. If training can push a model to superhuman results wherever such a test exists, then any AI research task with a clear score becomes a candidate.
Part three: the uranium analogy. He compares the result to a private company announcing "a million tons of 15% enriched uranium." It is not weapons grade, so you can say the output is not immediately deadly. The worry is that the "process line may generalize" to an extinction-level threat "within the unpredictably immediate future." He adds that forecasters still had Riemann at 2035 one week earlier. That forecast is his claim, and the post gives no source for it.
He also makes a point about selection. Earlier AI math wins could be dismissed as AI picking off famous problems that happened to be easy. RH, he says, was the problem most often named as a warning sign, which makes that dismissal harder.
How did people reply?
The replies under his post are a useful map of the disagreement. These are commenters' views, not findings.
- "How do we know it solved it?" One reader asked whether anyone can check the answer. Yudkowsky replied by asking the Grok bot on X to explain Lean. The short answer is the one above: a machine checks each step.
- "It is not killing people." Another reply mocked the idea that the AI would already be doing harm by now.
- The sink problem. One commenter argued that every success so far is symbol manipulation, language and representation, and that AI still cannot unclog a sink. That is a claim about physical-world limits.
- The reassurance reading. Joshua Pantony wrote that you could say "we've achieved AI smart enough to crack the Riemann Hypothesis and so far that AI seems relatively safe," and that it is unclear why higher intelligence would change that.
- The bait question. Eric Moss asked, sincerely, for an explanation of the link between solving a math problem and a threat to humanity.
The strongest skeptical case in the wider debate is not in that thread. Georgetown's Cal Newport argues that recursive self-improvement rests on unproven assumptions. We tested his three objections in Cal Newport's case that superintelligence won't happen.
Does a math proof really generalize to AI research?
A crystal ball on a stand, standing for the Riemann Hypothesis forecasts that the AI canary argument says were wrong
Both sides have a real point, and the disagreement is narrower than the thread suggests.
What supports the generalization. Verifiable tasks are where reinforcement learning works best, because the reward signal is cheap and reliable. Lean gives exactly that. Writing and tuning code also has tests. So do many parts of machine learning research: a lower loss, a higher benchmark score, a faster kernel. Anthropic's own August work shows the pattern at smaller scale. In Claude's Riemann zeta lower bound, Claude raised a known proportion of zeros on the critical line to about 67 percent, using many subagents. That was partial progress. The November claim is the full problem.
What weakens it. Much of AI research is not cleanly testable. Choosing what to study, judging whether a result matters, and getting compute and data are open-ended. A proof assistant can certify that a theorem is true, but it cannot tell you which theorem is worth proving. Math also leaves out cost and time in the physical world, which is the sink argument in more formal clothes.
There is also an honest unknown: how much of the result came from the model and how much from Alpöge's own skill. He credits Claude strongly, but the split is his account, and commenters on Hacker News disagreed about it. Alpöge has a record here: he previously posted a Claude-generated result on the Jacobian conjecture, covered in our Jacobian conjecture post.
The uranium analogy is a claim about direction and speed, not about the current state. It cannot be proven or refuted by one result. What one result can do is move forecasts, and that is the real news here: people who had 2035 in mind now have to explain the miss.
Where does this sit in the 2026 math run?
The Riemann claim did not arrive alone. It follows a stream of AI math results this year, with mixed reliability:
- September: The claimed Navier-Stokes solutions, which we sorted in our Millennium Prize fact check and the story on how Lean cut formal proof costs.
- October: OpenAI's release of hundreds of manuscripts. See what to check in the 722 manuscripts.
- October 29: A Royal Society meeting, as reported on a Substack by Stefan Bauschard, with more than 300 attendees in person and about 700 registered online. He quotes Cambridge's Miranda Cheng calling it "probably the end of mathematics as we know it." We read this as the author's report.
- November: The Alpöge claim and a second proof related to Navier-Stokes, released with NYU mathematician Tristan Buckmaster. See Buckmaster and Alpöge on Navier-Stokes.
The pattern in our tracking is that results with a short, machine-checked statement hold up. Results that rest on long text and trust take longer. The Riemann claim has the stronger form of evidence, which is why even cautious mathematicians have shifted to "almost certainly true."
What does this mean for what you build?
You do not need to settle the extinction debate to use the practical lesson.
- The checker is the scarce asset. Tasks with a fast, trusted pass or fail test improve fastest. If you build agents, put as much design effort into the verifier (tests, type checks, evals, formal specs) as into the agent.
- Verify the statement, not just the proof. The first doubt in the Lean case was not whether the logic was right. It was whether the right thing was proved. The same trap exists in code: passing tests that test the wrong behavior.
- Expect faster automation of verifiable work. Code with good tests, data cleaning with schemas, and proofs are all in this class. Open-ended judgment work is less exposed.
- Keep controls on actions, not just outputs. If agents get better at anything testable, the limit on harm is what they are allowed to do. For that layer, AgentBeam is an agent security platform from the explainx.ai team that stops agents before they take dangerous actions.
For a deeper look at how to build verifier-first agent loops, our workshops cover evaluation design, and the post on AI labs holding back math solutions explains why disclosure and checkability matter.
What is still open?
- Statement and comparator checks. Until Buzzard and others finish, the result is "almost certainly true," not formally accepted. The Clay Mathematics Institute had not recognized it at publication.
- A human-readable proof. Alpöge says he intends to write one. Number theorists will judge the ideas from that, not from the Lean code.
- Anthropic's own position. No company statement appears in the coverage. The claim is from a named employee, not a company announcement.
- The risk question. Nothing in the proof tells us whether recursive self-improvement is near. That depends on evidence about AI research automation, which this result does not provide on its own. For the broader safety framing, see our look at Sam Altman's three endgames and Yudkowsky's plan and the critique of Anthropic's safety communications.
Related reading
- Alpöge's Riemann Hypothesis claim: claimed vs verified (live log)
- Cal Newport on why superintelligence won't happen
- Sam Altman's three endgames and Yudkowsky's plan
- Masullo's human-only GRH draft
- Claude Riemann zeta lower bound (August 2026)
- Millennium Prize problems and AI: fact check
- Anthropic IPO priced at $83
Primary sources: Scientific American on the verification, TechCrunch, Nature, Science, and Yudkowsky's post on X.
Details are accurate as of November 21, 2026. Verification is ongoing and facts may change.
