On Friday, November 20, Levent Alpöge, a number theorist who works at Anthropic, posted a four-page note on his website. One sentence matters: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis” (the note, PDF).
The Riemann hypothesis is the most famous open problem in mathematics. It is about how prime numbers are distributed, it has been waiting for a proof since 1859, and the Clay Institute offers a million dollars for it. According to the timeline attached to the note, the work with Claude began on October 6 and ended on a Sunday night in mid-November. About forty days.
A few hours later Salvatore Sanfilippo, the creator of Redis, published a video in Italian titled “Anche il software è ancorato al passato”: software is anchored to the past too. He talks about mathematics for sixteen minutes, but the target is us programmers. I listened to it twice and checked what he says. Below is what holds up, what needs caution, and what I think.
The Riemann hypothesis in two minutes
Prime numbers (2, 3, 5, 7, 11…) look scattered at random, but on average they follow a rule: the higher you go, the rarer they get, at a rate we know how to compute. In 1859 Bernhard Riemann tied that rule to a function, the zeta function. It equals zero at infinitely many points, and the position of those zeros says how far the real primes stray from the average rule.
The hypothesis is that all the “interesting” zeros lie on a single line, the one with real part 1/2. If that’s true, primes are distributed as regularly as they possibly can be. Computers have checked billions of zeros and they are all there, but checking many is not proving all. That’s why so many theorems begin with “if the Riemann hypothesis is true…”.
What was announced, and what wasn’t
It’s worth starting from the facts, because “AI solved Riemann” is already going around.
Alpöge did not publish a readable proof. He published the note and a fourteen-part archive with the Lean formalization: the proof written in a language a computer can check line by line. Lean is a “proof assistant”: a programming language born at Microsoft Research where every logical step has to be spelled out, and a program verifies that each one follows from the previous ones. If the file compiles, the proof is formally valid. What still has to be checked, by hand, is that the statement written in Lean really is the Riemann hypothesis and not something that looks like it. A text “for humans” exists, Claude generated it, and it runs past 250,000 pages. He calls it “slop” himself and chose not to release it.
On credit he is very honest. He writes that he set the direction, but that near the finish line Claude “far outpaced me”. And he admits he is presenting the world with a complete argument “before having mastered” the mathematics around it. The method, he says, extends to the Generalized Riemann Hypothesis as well.
Which model he used has not been disclosed: the note only ever says “Claude”. In the video Sanfilippo speaks of Fable, Anthropic’s latest model, and it probably is that one. But that’s a deduction, not a fact.
As of today, this is where things stand:
- No independent verification. Nobody outside Anthropic has yet checked that the Lean files really prove the right statement (AFP on Phys.org, DataCamp).
- Experts are cautious, not skeptical. James Maynard, a Fields medalist, suspects the proof is correct, but says checking it will take time and understanding why it works will take much longer. Youness Lamzouri talks about months or years (IBM).
- The timing is unfortunate. The announcement came the day after Anthropic’s stock market debut. Alpöge knows it; he writes that “the precise timing is horrible” and that the cynical reading is false. The coincidence is still there.
Then there is a second twist. On the same day Ravi Vakil, president of the American Mathematical Society, hurriedly posted the draft of another proof, by Alessandro Masullo: twelve years of work, “all human, with no AI” (DataCamp, Vakil’s post). Eight hours later Vakil updated the post: several mathematicians had reported problems in the draft. Found with the help of AI models.
That story alone says a lot about where we are.
Tao and 2050
Sanfilippo opens the video with a jab at Terence Tao, probably the best-known living mathematician. About a month ago, in an interview, Tao had said that for the Riemann hypothesis “something beyond current AI is needed” and that maybe “some combination of humans and AIs in 2050 or something” would make further progress. The quote is in a Hacker News thread with a link to the original video.
For Sanfilippo that sentence shows Tao has not grasped the scale of what is happening: he extrapolated AI progress badly and underestimated today’s models.
I would slow down here. Tao’s forecast looks like it aged terribly, but only if the proof holds. Today we don’t know. And a signal had already arrived in October: OpenAI announced a proof of the so-called “quasi-Riemann hypothesis”, a weaker result that experts consider largely correct (Quanta). The person who was off by twenty-five years a month ago was no amateur: he was one of the best in the world, with all the information available. More than Tao’s fault, to me that measures how hard forecasting is right now.
The job is choosing the questions
The best part of the video is a different one. Sanfilippo says Riemann’s huge contribution is not the missing proof but having stated the hypothesis: seeing a link between the zeros of a function in the complex plane and the prime numbers, which live somewhere else entirely.
He quotes Henri Poincaré, and the quote is right. In Science and Method (1908) Poincaré writes that to invent is to discern, to choose. True statements are infinite. The mathematician’s work is knowing which ones are worth looking at.
If proving becomes fast, value moves even further toward the questions. And Alpöge’s note, read carefully, confirms exactly that: he brought the starting idea and the right analogy, and Claude covered the rest of the road at a speed he couldn’t follow.
And we programmers?
Here Sanfilippo makes the jump: are we sure software has made this transition? His answer is no, with three arguments.
Companies. If AI writes code so well, why are we surrounded by terrible software? Because in big companies the person who decides a feature, the one who approves it and the one who implements it sit in three different offices. Good software is born when whoever decides and whoever builds are the same person or a close-knit group. AI writing code changes nothing if it crashes into that bureaucracy.
Individuals. Even people working alone, he says, haven’t made the leap. He draws the parallel with the arrival of compilers: assembly programmers looked at C with suspicion, then high-level languages won because they let you think about the result instead of the effort. With one difference that in his view favors AI: Python made you pay for convenience in speed, while today’s models write fast code and optimize past the point where a programmer would have stopped for lack of time.
The examples. He says a firmware update bricked his Polar heart rate monitor, and that Claude unbricked it by analyzing the app. And he mentions Stefan from CNC Kitchen, who reportedly had Claude rewrite the software of a Canon scanner. I found no written record of these two episodes: I report them as he tells them. What is documented for CNC Kitchen is ScanRuler, a tool for measuring 3D scans built with Claude’s help, so the character fits.
His conclusion: a single person, in an afternoon, can write better software than what a company ships with its hardware. And you can write “disposable” software that made no sense before: a tool only you need, for three months, built in three days.
What I think
I agree with almost all of the diagnosis. I have some reservations about the recipe.
- Code is no longer the bottleneck. I see it every day: the slow part of a project isn’t writing, it’s deciding what to write. Using AI to build the same management app as before in half the time is like using a car to travel at a horse’s pace.
- Disposable software is the real novelty. The tool made for a single problem, thrown away when you’re done, is the most concrete change for me. It used to cost too much. Now it doesn’t.
- Mathematics has Lean, we don’t. This is the difference the video skips. Alpöge’s proof, unreadable as it is, has a checker that says yes or no. Software doesn’t. Tests check what you thought of checking. I wrote about it in the post on vibe coding: when nobody understands the code, sooner or later someone pays.
- 250,000 pages of “slop” are a warning. A mathematician admitting he hasn’t mastered his own proof is the noble version of a problem we know well: code that works and nobody can explain. Tao said it in September: AI can produce solutions faster than mathematicians can understand them (IBM).
- “Better than Canon in an afternoon” is true for you. A company has to make that software work on millions of computers, translate it, maintain it for ten years and answer when it breaks. That’s no excuse for bad software, but it explains part of the gap.
Poincaré’s question remains, and it’s the right one for us too. Now that writing is cheap, the job is choosing what is worth building. And understanding it well enough to defend it when someone asks why it works.
