The Relational Arc · November 21, 2026
After an AI-Assisted Proof of the Riemann Hypothesis, Is Mathematics Still Worth a Life's Work?
On the day the proof passed its computer checks, a mathematics student wrote that math class would become PE class. I think the student is right about what was lost. A proof that a computer has accepted and no person can yet explain also shows what has not yet been handed over, and how much depends on people choosing to keep it.
On November 20th, 2026, a few hours after a proof of the Riemann Hypothesis was released, a mathematics student who posts on X as danning wrote that the world was “no longer the same world I grew up in” (the post). By the next morning nearly half a million people had seen it. Quoting a friend, the student wrote that “math class will become PE class”, a way for children to exercise their brains but no longer a subject to give a life to for the sake of “making progress for humanity.” The life’s work of a great many people, the post went on, “is now just ashes in history.”
The Riemann Hypothesis is a claim the mathematician Bernhard Riemann made in 1859 about one mathematical function, now called the zeta function: it says where the points at which that function equals zero lie. The answer bears on how evenly the prime numbers are spread along the number line, and it is one of seven problems for which the Clay Mathematics Institute has offered a million dollars since 2000. Levent Alpöge, a mathematician who works at the AI company Anthropic, released a proof that, by his own account in a short note, he found working with Claude, the company’s AI. The proof is written in Lean, a language in which a computer program checks every step. Within a day the people who look after Lean said they had “full confidence” that it is correct, and their statement ends by looking forward to the work “that will help explain the new ideas in this proof” (their joint statement). No account of those ideas written for people to read exists yet.
What the student lost is the hope that a life spent on mathematics could be the one that settles a question like this, and I think the student is right that it is gone, or nearly. But the line that “lots of people’s life’s work is now just ashes in history” treats two things as one: knowing that the Riemann Hypothesis is true, and understanding why it is true. This proof gives us the first without the second. A computer has confirmed that every step of the argument holds, but nobody can yet explain why it works, and the man who released it says he does not fully understand it yet either. That gap does not give the student’s hope back, but it does change what the friend’s own words, “progress for humanity”, now ask of people.
What the Student Lost
At its core, the student’s argument runs like this. Mathematics was worth a life because the progress was real and someone had to make it. If a machine can now make it, the subject is still there to enjoy, the way exercise is, but nobody needs a person to do it.
In an earlier essay, The Purpose Displacement Problem, I set two cases side by side that fit this exactly. When computers overtook the best human chess players, chess lost nothing it depended on, because nobody had ever needed the best move from a person. A surgeon whose operation a machine does better is in a different position: the parents should choose the machine, and the surgeon, who can still operate, is practising an art where she used to be saving lives. That essay was about what makes work feel needed, and it made no claim about what the surgeon is worth. The friend’s line about PE class says that mathematics has just crossed from the surgeon’s side to the chess player’s, and for the hope of settling the great questions I think that is close to true.
Two phrases in the post, written on the night, claim more than this week has shown, and I mention them only because a reader might take them as settled. One is that AI is now “smarter than humans in any capacity.” Three days earlier Epoch AI, a research group that tracks AI progress, had given six of the newest AI systems tasks from its own staff’s research, and none reached the standard Epoch expects of an employee; they were weakest at designing experiments and at judging which findings matter (Epoch’s study). The other is that discovery is now “just like tapping on a calculator.” Alpöge’s note describes something slower than that: years of preparing for this one problem, and weeks of work in which he steered the argument to the right setting before the machine took it further than he could.
The Mathematician Who Had the Same Hope
The person with the most reason to feel what the student felt is Alpöge himself. In his note he writes that it was his secret as a student that “my plans were for this problem in decades,” and that he spent the years since trying to learn every tool it might need. He got there, and the AI got there faster: he “guided us to the right situation,” he writes, and then “this time Claude far outpaced me near the finish line.” He does not pretend the moment is simple. “It is natural to feel that this is an end,” he writes, “in my case a loss of a principal guiding light.”
So the hope the student mourns came true for the one person who held it most closely, and he describes it, among other things, as a loss. He also describes it as amazement, and he says what he means to do next. He is, he writes, “presenting the world with a complete argument before having mastered a massive neighbourhood around its mathematics,” and he intends “to understand and impart understanding of this argument.” Even its author does not yet understand all of the proof.
Knowing That It Is True, and Knowing Why
The mathematician William Thurston, who won the Fields Medal, one of mathematics’ highest honours, in 1982, asked in 1994 what mathematicians actually produce. His answer, in On Proof and Progress in Mathematics, was that the real question is how they “advance human understanding of mathematics,” and that “the measure of our success is whether what we do enables people to understand and think more clearly and effectively about mathematics.” He had a case in mind. In 1976 Kenneth Appel and Wolfgang Haken proved the four colour theorem, that any map can be coloured with four colours so that no two neighbouring countries match, using a computer to check far more cases than any person could. The proof caused an argument for years. Thurston read that argument as having little to do with doubt that the theorem was true. It reflected, he wrote, “a continuing desire for human understanding of a proof, in addition to knowledge that the theorem is true.”
The Riemann proof is that case at a scale Thurston never saw. It runs to about seventeen million lines, each a small step in the argument, spread over 175,187 files, according to the record kept by Andrew Sutherland, a mathematician at MIT (his record of the checks). A checking program reads each line and confirms that it follows from the lines before it by the rules of logic. The program does not need to understand anything to do that, which is why the proof could be trusted within a day while nobody has yet written down why it works. The physicist Brian Greene, reporting the news for the World Science Festival that day, put it plainly: “it is not like any human being on planet earth yet understands it” (his video).
In my book The Calibration Problem I used the student’s own image, the calculator, as the plainest example of a tool that is safe to rely on: “The calculator is safe because the goal it serves, mathematical understanding, stays with the human while only the mechanical step is handed off.” In this proof the machine’s part, by Alpöge’s account, went well past the mechanical step. Thurston’s point is that the goal did not go with it. A proof that a checker accepts and no person can follow settles the question for everyone and explains it to no one, and if the progress the friend meant is humanity’s understanding, it was not finished on the day the checker said yes.
The Objection That Machines Will Do That Too
The strongest objection to all of this is the one the student would make next. Machines may well explain mathematics better than people do, and soon. If an AI writes the account that makes this proof clear, then understanding becomes something a person picks up by reading, which is a good thing to have and still sounds a great deal like PE class. The same goes for checking. A machine that can find a seventeen-million-line proof can presumably also be asked whether the theorem it proved is the one that was wanted.
This autumn gave one test of that. In September 2026 an AI built by OpenAI produced a proof that the Navier-Stokes equations, which describe how fluids move and are another of the Clay Institute’s seven problems, can predict a fluid moving infinitely fast. The proof met the wording of the problem, and a checker could confirm every step. Many experts still thought it answered the wrong question, because it worked by imagining an unrealistic force, tuned to stir the fluid in exactly the way needed (Scientific American’s report). The people who noticed were people who understood what the problem was for. On the day of the Riemann proof, Alpöge and Tristan Buckmaster, a mathematician at New York University, said they had closed that loophole with an AI’s help (Scientific American’s report).
The Riemann proof needed the same kind of check before anyone could trust it. A sentence can look like the hypothesis and ask something easier, for example by quietly leaving out one condition, and the checking program will confirm a proof of the easier sentence just as happily. So the Lean community wrote a file that states the hypothesis in the terms of Mathlib, the shared library of mathematics that a community of contributors has built in Lean over years, and had two separate programs confirm that the proof proved exactly that statement (the joint statement). Mathlib’s statement of the hypothesis, the community wrote, “has been carefully vetted by experts in Lean and number theory.” Sutherland then joined the proof to a Lean version of a test the mathematician Guy Robin found in 1984, a statement about whole numbers that is true exactly when the hypothesis is, and that does not depend on Mathlib’s definition of the zeta function (his record of the checks).
Each of those checks could in principle be given to a machine, and some of them already use programs. But every machine that checks another machine raises the same question one level up, and at the top of the chain either someone understands the subject well enough to say whether the answer is the one that was wanted, or everyone defers. In the book I described the slide from using a tool to deferring to it, and wrote that “it does not feel like a slide. It feels like getting better at using the tool.” Deferring is not always a mistake. If machines build a long record of catching problems like the Navier-Stokes loophole before people do, and that record can be checked by someone independent of the machines, then deferring may become the sensible thing, and the student’s view will be closer to right than mine. I would change my mind on that evidence. Today there is no such record, and the only independent check there is comes from people who spent years learning the subject. If enough students concluded that mathematics is PE class, there would be fewer of them, and the checking programs would go on saying yes to whatever they were given.
What I Cannot Tell the Student
The student was asking about a working lifetime, and my argument is about the next few years. I cannot promise that the work I have described will still be work for people when today’s students are forty. Nor can I say what a life is for once nobody needs its skill. That question was open before November 20th, and the proof has made it more pressing without answering it.
What I can point to is narrower and is a fact about this week. The one person who held the student’s hope, and saw it come true in a way that left him behind, has said that the next thing he means to do is understand the proof and explain it. That work is open to anyone, and it answers to a standard outside the person doing it: an explanation either helps other mathematicians follow the proof or it does not. Nobody yet knows whether a machine will do it first. It is a smaller thing than the hope the student lost, and I would rather call it small than pretend it is none.
Reading List & Conceptual Lineage
From Sentient Horizons
The Purpose Displacement Problem
What work was for, and the difference between the chess player and the surgeon when a machine overtakes them.
If AI Wins Wherever the Answer Can Be Checked, Is Judgement What Is Left for People?
Why “the work that can’t be checked” is a poor place to rest what people are for, and why the question of what a life is for stays open.
Human of the Gaps: If AI Can Do It, Do People Matter Less?
Why a gap in ability settles neither what a person is worth nor whether a machine has a mind.
The Calibration Problem
The chapters on ladders and on living with powerful tools, where the calculator and the slide from use to deference are set out.
External Works
Levent Alpöge — Remark (November 20th, 2026). His four-page note announcing the proof, with how it was found and what he means to do next.
William P. Thurston — On Proof and Progress in Mathematics (Bulletin of the American Mathematical Society, 1994). The case that what mathematicians produce is human understanding, not only proofs.
Lean and Mathlib communities — Lean verification of the Riemann Hypothesis (November 21st, 2026). The joint statement on the proof’s checks.
Andrew Sutherland — Verification files for the Lean proof of the Riemann Hypothesis (updated November 21st, 2026). Every check, with its result.
Peter Hall — After months of controversy, AI finally solved the ‘real’ Navier-Stokes problem, Scientific American (November 20th, 2026). The loophole in the September proof, and the claim to have closed it.
Brian Greene — Has AI Solved Math’s Pinnacle Problem?, World Science Festival (November 20th, 2026). A three-minute report on the day of the announcement.
Epoch AI — How good are Fable 5.2 and GPT-6.2 Astra at real work? (November 17th, 2026). The newest AI systems tested on a research group’s own staff work.
Sentient Horizons is an essay series on what a mind is, whether a machine could have one, and how to act toward minds we cannot see inside. Start Here is one route through it; Essays lists everything, newest first.
Three Ways In
- Consciousness Is Like Flight If you want to know what a mind is
- We Built the Alien First If you want to know whether a machine could have a mind
- The Purpose Displacement Problem If you want to know what AI means for your own life