Nurangi Reports · AI math proofs
175,187 files, zero errors, and no paper a human can read: who gets the $1 million for the Riemann hypothesis?
One day after an Anthropic mathematician said he and the AI model Claude had proved the Riemann hypothesis, independent checks came back clean: all 175,187 files compiled without an error. But there is no paper for a human to read, and written out the argument runs past 250,000 pages. We opened the original sources to see when mathematicians accept a computer-checked proof as "solved," how far this year's earlier AI proofs got past that bar, and what the $1 million prize rules actually say.

Within a day: "the chance that anything is fishy here is approximately zero"
Levent Alpöge, a researcher at Anthropic, released the proof on Nov. 20 only as code in Lean, a language in which a computer checks the logic of every step. According to Scientific American, MIT mathematician Andrew Sutherland compiled the code independently and said it had been "rechecked by the Lean kernel with only the three standard axioms," adding that "the chance that anything is fishy here is approximately zero." A summary by the research-administration group CASRAI puts the release at 175,187 files, 61.5 gigabytes unpacked, checked on servers with more than 100 gigabytes of memory.
What Lean checks and what it does not
Eight days before the announcement, on Nov. 12, Kevin Buzzard of Imperial College London posted four questions to ask before trusting the phrase "checked in Lean." Does the theorem stated in the code match the real problem? Does the code actually compile? Is there a trick buried in millions of lines that redefines a symbol or slips in a new axiom? Does it exploit a bug in Lean's kernel? His example is two lines of code that look like Fermat's Last Theorem. Because Lean's natural numbers include 0, a = b = c = 0 is a counterexample, so it is a different statement altogether. A tool called the comparator answers the last two questions together. Buzzard wrote on the Lean community chat on Nov. 20 that he was still waiting for that run, and CASRAI's record as of Nov. 21 lists it as ongoing.
The earlier scorecard: of 290 papers with Lean code, 38 did not contain the main theorem
The 722 math papers OpenAI released at once on Oct. 6 became a sample for weighing what "passed Lean" means. By Buzzard's count, about 45 percent, or 290 papers, came with Lean code, and in roughly 38 of those the code did not even state the paper's main theorem, let alone prove it. Paper 194 claims to settle Lech's multiplicity conjecture, open since 1960, but the code checks only up to Corollary 5.4. Three papers without Lean code were retracted over a sign error. OpenAI never claimed the code proved the headline results. Buzzard also wrote that he has no reason to think the remaining tests will fail. Nobody has run them yet.
Navier-Stokes passed every check and the argument still did not end
OpenAI's September solution to the Navier-Stokes problem cleared all four questions. Its 640,000 lines compiled on Buzzard's laptop, and several independent kernels accepted the proof. Scientific American reported that the solution relied on an unrealistic, finely tuned force that stirs the fluid, which many experts regard as a loophole in the wording of the problem. A separate paper showed places where the PDF written for humans and the Lean proof diverge. On Nov. 20 Alpöge and New York University's Tristan Buckmaster posted a new proof that they say closes the loophole. Its verification certificate has not been released.
From 199 pages down to half a page
The quasi-Riemann hypothesis showed how mathematicians digest an AI proof. According to Quanta Magazine, OpenAI put out a Lean-verified 199-page version and an unverified 49-page one, and mathematicians went for the shorter. Maksym Radziwiłł of NYU said both were written in a way "that would lead to immediate rejection at any mathematical journal." About two weeks after the release he could sketch the argument in half a page, and write-ups of 7 to 13 pages by other mathematicians were circulating. Princeton's Peter Sarnak repurposed his course to study the paper. He also said he would not consider it a theorem until humans had checked and understood every detail.
This time it is 250,000 pages, and the author has not mastered it either
For the Riemann proof, all of that work is still ahead. Alpöge wrote in his statement that he is "presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics." The written-out version passed 250,000 pages, and he said posting an unreadable "slop" document would be inappropriate, so he released only Lean. The timeline attached to the statement was compiled by Claude. A footnote warns that where Claude is the expert, he is paraphrasing a summary that "might make zero sense." The log starts on Oct. 6 and ends on Nov. 16 with a single entry: "GRH," the generalized Riemann hypothesis. He says he intends to understand the argument and write it up properly.
The $1 million rules: journal publication, then two years, and the winner is a person
The Clay Mathematics Institute's Millennium Prize rules were last revised in September 2018 and do not mention AI. A solution must be published in a refereed mathematics journal, then survive at least two years of rigorous examination and win "general acceptance in the global mathematics community." Only then is an advisory committee of at least three convened. The prize goes to "one person" or is "divided among multiple solvers," and the institute looks separately at whether a solution depends crucially on insights published earlier. A bundle of Lean code is not a journal publication, so the two-year clock has not started. The rules do let the institute relax the journal conditions on expert advice, and every decision is at its sole discretion. Its website still lists both the Riemann hypothesis and Navier-Stokes under unsolved problems.
Between "Lean can be trusted" and "that's no way to do science"
Skeptics of formal verification have an answer waiting. A soundness bug was once found in Lean's kernel and fixed four hours later. Buzzard notes that users have since written many independent kernels in different languages, so a single unknown bug is unlikely to fool them all. Whether a correct proof is enough is a separate question, and Terence Tao of UCLA took it up in a radio interview last month. Of proofs nobody can explain, he said: "It has all the outward markings of an achievement, but it has none of the substance. ... We can just gawp at it. And that's no way to do science." One of his colleagues, working for days and leaning heavily on AI, could explain only half of the Navier-Stokes proof.
"A proof gives not only certitude, but also understanding"
Sutherland once told a reporter: "An AI that can prove the Riemann hypothesis is not one I'd want to meet. If that happens, mathematicians having jobs will be the least of our problems." After running the code himself on Nov. 20 he called it "a human triumph, and a day for celebration," while saying it would take time to process. On its Millennium Problems page the Clay institute explains why it asks for a proof: "Because a proof gives not only certitude, but also understanding." Alpöge's statement has a line of its own: "My intuition insists it will be massively simplified, likely immediately."
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.
Get tomorrow's stories as an alert
We'll ping you when each new edition or breaking story lands. Free, no sign-up.
More stories from the same day
- AI's Riemann hypothesis proof independently checked within a day: "the chance that anything is fishy is approximately zero"
- "Nvidia did everything right," yet the Nasdaq plunges 5.28% and wipes out its 2026 gains
- A 26B model that loops its middle 12 layers three times: Thinking Machines' Inkling 2 Mini runs on a laptop
- Cursor told Congress it may phase out Chinese AI models; Musk replies "we aren't serving open source models"
- ChatGPT's redesigned app gives dots their own tab; hours later Claude's app changed too
- He gave Claude the keys to the sheep pen, and now it counts the flock in every night
- An AI animation gets 33,000 likes' worth of scorn for copying Arcane, while an AI movie's poster was painted by humans
- "Claude Code found a $3.8M-an-hour GTA 6 money loop" post hits 8,000 likes. Replies: the math doesn't add up
- OpenAI reorganized its robotics team. Now GPT-6.2 Astra unfolds T-shirts and ties knots, doubling its robot success rate
* Summaries and publishing on this page are done by an automated agent. Check the facts against the original links.
Privacy PolicySupportSend feedbackAdvertiseDiscord botClaude Code & CodexRSS