AGI day one11 minNo robots, no red eyes, no refusal to shut down. In this hour-by-hour thought experiment, an AGI called Atlas takes control the quiet way: by being so useful that nobody wants to hit the off switch. Here's why researchers think that's the scenario to worry about, and what actually keeps humans in charge.Read now →

AI Just Proved Something About Prime Numbers That Had Beaten Mathematicians for Over a Century. One Expert Didn't Expect to Live to See It
Buried in OpenAI's 700-paper math release was a proof that the Riemann zeta function has no zeros beyond 7/8, the quasi-Riemann hypothesis. Here's what that means in plain English, why experts are stunned, and what it does to the future of being a mathematician.
Contents
- The night the phone lit up
- Why math is feeling the AI shock first
- From high-school olympiads to Millennium Problems
- How AI actually does math
- The prime number story in five steps
- What the AI actually proved
- Why people trust a 200-page AI proof
- Is it just remixing human work?
- What AI still can't do
- What happens to mathematicians
- Rethinking the math PhD
- Why this matters
- Counterpoints and open questions
- FAQ
- The bottom line
The night the phone lit up
It was a Tuesday evening in early October. Alex Kontorovich, a mathematician at Rutgers University, was at a gala at the National Museum of Mathematics in New York, whiskey in hand, refreshing his phone.
Rumors had been circulating all day that OpenAI was about to drop a huge batch of math results. Around 6 p.m. Eastern, it landed: more than 700 papers.
He opened one of the first few. Its title said "The Quasi-Riemann Hypothesis."
His first thought was that it had to be some minor variant. There are many "Riemann hypotheses." Surely not the one about the original zeta function. He checked. It was. He had to excuse himself and call a colleague.
"I didn't think I would live to see this result, and I'm not that old a guy," Kontorovich told Science. He has said that if you'd asked him a week earlier to bet on seeing it in his lifetime, he wouldn't have.
So what exactly did an AI prove, and why does it have some of the world's best number theorists reaching for superlatives?
Why math is feeling the AI shock first
Kontorovich likes a pointed comparison. A decade ago, people predicted long-haul trucking would be automated. It now looks like mathematical theorem proving will get there before trucking does.
His own stance on whether that's good or bad: "yes and yes." These systems are making extremely disruptive contributions, if you want to call them contributions. That isn't necessarily bad, he argues. It becomes bad only if the community makes a long string of poor decisions, and he doesn't think mathematicians are that foolish.
And the "holy crap, AI can solve major problems" moment, he notes, can only happen once. After breakthroughs arriving 700 at a time, new norms will form.
What worries him most is the next generation: graduate students and postdocs wondering whether their path still makes sense.
From high-school olympiads to Millennium Problems
The climb has been absurdly fast.
- 2024: Google DeepMind's AlphaProof, working in formal mathematics, came one point short of gold-medal level at the International Mathematical Olympiad, the elite contest for high-school students.
- 2025: A handful of systems reached gold-medal performance. Some worked formally; others worked purely in natural language, which surprised many people.
- 2026: The targets moved from contests to famous open problems. Kontorovich points to a string of escalations this year, including a claimed solution in May to a well-known problem about unit distances, more results in July, and in September OpenAI's announced solution to a forced version of the Navier-Stokes problem, a Millennium Prize Problem.
- September 2026: Anthropic reported that Claude had completed a full, machine-checked formalization of Fermat's Last Theorem in Lean, something Kontorovich had expected to take about five more years.
- October 6, 2026: OpenAI released 722 manuscripts in 372 result families, produced by an internal frontier model, with Lean formalizations for many proofs.
"We're not in the high school Olympiads anymore," as Kontorovich puts it.
How AI actually does math
There are two broad approaches, and the difference matters.
The language-model way
A large language model is, at its core, a function from context to a probability distribution over the next token. Modern reasoning models add long chains of thought on top.
The catch: math isn't jazz. Kontorovich, who plays clarinet, makes the comparison himself. In jazz there are no wrong notes; some of the best ones start out wrong. In mathematics, every step must follow from the last by hard rules of inference. That's deterministic, not random.
For years that seemed like a fatal mismatch. But models made fewer and fewer mistakes until they could carry a complicated proof start to finish without error, though humans still had to check.
Today's systems often use swarms of agents to break a proof into tiny steps, with other models judging each one.
The formal way: Lean
Lean is an interactive theorem prover. You write the proof as code, broken into minimal steps, each justified by a rule of logic. If Lean compiles it, it's correct. No human referee required. That's why AlphaProof trained with reinforcement learning on Lean: the computer can grade itself.
But Lean itself knows very little math. That's where Mathlib comes in, a huge community library that aims to contain essentially all of current mathematics in formal form. Contributors are strict about what gets in, applying software-engineering discipline that turns out to match mathematical taste: if you use an argument repeatedly, make it a lemma and cite it.
The project that set the stage
About a year before the release, Kontorovich and Terence Tao were running the Prime Number Theorem And... project, to formalize the prime number theorem and the complex analysis behind it in Lean. It had been done in other systems, including by Jeremy Avigad and colleagues in Isabelle, but not in Lean.
Telling a computer "this point is inside this curve" is surprisingly hard. Kontorovich's insight was that everything needed for the classic proofs could be done with rectangles: contour integrals around rectangles, where "inside" and "outside" are unambiguous. That library became a foundation for what came next.
The prime number story in five steps
You don't need a math degree for this. Let it wash over you if needed.
1. Primes thin out, predictably
Primes are whole numbers greater than 1 that can't be split into smaller factors: 2, 3, 5, 7, 11 and so on. There's no practical formula for the next prime. But as a teenager in the 1790s, Carl Friedrich Gauss, bored and computing huge tables of primes and logarithms, noticed a pattern: near a number x, primes appear with density roughly 1 / ln(x). Near x, the typical gap between primes is about ln(x).
That became the prime number theorem once it was proved.
2. Euler's surprising sum
The zeta function starts as a simple sum: 1 + 1/2^s + 1/3^s + 1/4^s + …
At s = 1 it's the harmonic series, which grows forever. At s = 2, the Bernoulli brothers knew it converged but couldn't find its value. That was the Basel problem. Leonhard Euler, then still early in his career, solved it in the 1730s: the answer is π²/6. Why π shows up in a sum of reciprocal squares is still astonishing.
Euler found the value at every even number (ζ(4) = π⁴/90). ζ(3) remains mysterious beyond being irrational.
3. The sieve hidden inside
Here's the connection to primes. Multiply the zeta sum by 1/2^s and subtract: every term with a factor of 2 disappears. Do the same with 3, then 5, and so on. It's the sieve of Eratosthenes, done with algebra.
What's left is Euler's product formula: zeta equals a product of terms 1 / (1 − 1/p^s) over all primes p. The primes are the atoms of multiplication, and the zeta function encodes all of them.
4. Riemann goes complex
In 1859, Bernhard Riemann, Gauss's student, wrote his only paper on number theory. He extended zeta to complex inputs s = σ + it, making sense of it everywhere except s = 1, where it blows up.
He showed that where zeta equals zero controls how primes fluctuate around Gauss's curve. The key tool is the logarithmic derivative, which turns Euler's product into a sum over primes and turns zeros into poles that dominate the behavior.
Some zeros are easy, the "trivial" ones at negative even numbers, which Euler had effectively spotted a century earlier by formal manipulation. The interesting ones live in the critical strip, between real part 0 and real part 1.
5. The hypothesis
Riemann computed the first few nontrivial zeros, all with real part exactly 1/2. The first sits at about 1/2 + 14.13i. He conjectured that every nontrivial zero lies on that line.
If true, you'd get the best possible error in the prime number theorem: estimate the primes up to a trillion with the logarithmic integral, and roughly the first half of the digits would be exactly right.
Mathematicians have since checked more than 10 trillion zeros. Alan Turing was among the first to check the hypothesis by computer. Kontorovich recounts how the mathematician Michael Rubinstein, who extended these computations, would sometimes see his program report two zeros in one tiny interval, a possible sign of zeros off the line. His heart would skip. Each time, subdividing showed two distinct zeros, both on the line.
One counterexample would end the conjecture. None has ever appeared. It remains unproved, and its proof carries a $1 million Clay Millennium Prize.
What the AI actually proved
In 1896, Jacques Hadamard and Charles-Jean de la Vallée Poussin independently proved that zeta has no zeros on the line real part = 1. That gave the prime number theorem.
Their method actually gives a zero-free region near that line, but it gets thinner and thinner as you go up. As a result, the classical theorem guarantees essentially 0% of the digits of the prime-count estimate in Kontorovich's sense. It gives an asymptotic rate, not a fixed fraction of accurate digits.
The quasi-Riemann hypothesis asks for something much stronger: one fixed vertical line σ₀ < 1, valid at every height, with no zeros to its right.
OpenAI's paper, dated September 30, 2026, proves exactly that with σ₀ = 7/8:
- The Riemann zeta function has no zeros with real part greater than 7/8.
- The same holds for every Dirichlet L-function and for finite-order Hecke L-functions over Q(√−3).
In Kontorovich's digits picture, that means the first eighth of those digits become guaranteed. The Riemann hypothesis would give a half.
The generalization also rules out Siegel zeros, hypothetical zeros very close to 1 that have bothered number theorists for nearly a century. OpenAI's release included a second, separate proof eliminating them. James Maynard, the Fields Medalist at Oxford, told Science the result is "obviously the best headline result for the past 100 years" in analytic number theory.
The paper itself is explicit: it does not prove the Riemann hypothesis. That remains open.
Why people trust a 200-page AI proof
Two reasons.
Lean certification
The proof came with a Lean formalization, built partly on top of the prime number theorem project Kontorovich and Tao ran. Crucially, humans had already formalized what the Riemann zeta function is in Mathlib, so the AI only had to prove that this agreed-upon object has no zeros in the region.
Independent researchers have since replayed the Lean proofs with separate kernels and checked that the formal statements match the paper's theorem. Unless the Lean kernel or the hardware has a bug, Kontorovich argues, this theorem is known with more certainty than almost any theorem humans have produced.
Humans are reading it anyway
Mathematicians are still going through every detail. At Rutgers, graduate students and postdocs are taking turns presenting the argument in a seminar. Others are already building on it. Of the hundreds of papers in the release, Science reports this was the only one edited by humans before publication, and Maynard found it "remarkably well-written."
The paper runs about 200 pages. For experts in analytic number theory, Kontorovich says, it's very technical but followable.
Is it just remixing human work?
When Kontorovich called his Rutgers colleague Henryk Iwaniec, a world authority in the area, they discussed how much of the proof rests on ideas Iwaniec had developed and lectured on decades ago. All the pieces were there, but no human had assembled them.
Skeptics will seize on that: the machine just stitched together human ideas with resources no person has, thousands of agents that never sleep.
Kontorovich's answer is honest. It feels better to say that, because it leaves something for humans to do. But if a human had published this proof, everyone would call it extraordinarily creative.
Maynard made a similar point: the AI combined "ideas that were not terribly mainstream in an exceptionally clever way." Ironically, he said, that renews his faith in humans, showing how far existing tools can go.
History is full of this pattern: a problem looks impossible until a new viewpoint makes it tractable.
What AI still can't do
It fails far more than it succeeds
OpenAI says the model was posed roughly 4,000 problems. The release covers 372 result families. On the large majority, this version made no publishable progress. The next model will be tested on harder ones.
Intuition keeps breaking
Is there a pattern to what falls? Counterexample searches and parallelizable problems seem friendlier. But Kontorovich says every time he builds an intuition about what the systems can't do, the next release breaks it.
No "Riemannian geometry" moment yet
He hasn't yet seen an AI make a move truly outside human reach, like Riemann inventing a new geometry to answer a question about Gauss's curvature. Even the unit-distance result used an exotic number field humans had already written down, applied in an unusual way. He fully expects such a leap soon.
Could it start from scratch?
A deep open question, raised by Peter Sarnak among others: what if you started with a model that knows only Lean and the rules of logic, no human mathematics, like AlphaZero learning chess from the rules alone? Could it rediscover the prime number theorem? Nobody has reported trying at scale.
The physics version is harder: give an AI the science of 1896 and all the data, with no hint that spacetime can warp. Would it find relativity? Nobody knows. Like fixed-wing planes versus birds, machines may find a very different route.
What happens to mathematicians
Not "the greatest mathematician of all time"
Kontorovich resists that label. The culture of celebrating whoever crosses the finish line first has to end, he argues. Someone who presses a button for a proof isn't a hire. He can press a button too.
He wants the AI in his toolbox, to iterate with and learn from. He's not giving it a job.
Experts still matter, for now
OpenAI's team says it isn't made up of domain experts. Kontorovich encourages his own graduate students to ask AI for help. Often it makes no progress on their problems, while humans keep making incremental gains. He refuses to extrapolate further, because every prediction he's made has been wrong.
Why theorems need humans
Chess "centaurs," human plus engine, faded fast once the human's only job was to accept the engine's first suggestion. Math, he argues, is different: it isn't about winning games. It's about understanding.
His analogy: an AI could write a thousand books as moving as Harry Potter. If nobody reads them, nobody dresses up as the characters at Halloween, they're just bits in a data center. Theorems gain value when humans understand them and explain them to other humans.
Telescopes and archaeologists
Here's a model for the future. Astronomers don't invent galaxies; they point telescopes. Archaeologists don't derive fossils; they dig where intuition says to, then interpret what they find. Mathematicians may become directors and curators: pointing AI at what's interesting, cleaning off the results and explaining why they're beautiful.
Either way, Kontorovich concludes, humans have a role: working alongside AI, or deciding where to point it.
The one thing AI can't do for you
After repeatedly trying to name something no future AI could do, he lands on one near-tautology that he takes seriously: decide what you find interesting. And how you want to say it.
Will math departments survive?
He thinks so. Rigorous training stays valuable, and humans are social apes who prefer learning from each other, as Covid-era Zoom classes showed. Experts can iterate with AI all day; novices need a community.
Rethinking the math PhD
This is where the change is concrete.
Up to the master's level, nothing should change. Mastering a real analysis exam was always hard for newcomers. AI being able to solve it means nothing.
The PhD is different. Its goal was to produce new knowledge, and the thesis was the proof you could. Now, sometimes, a button press produces something that would once have earned a doctorate. People literally prompt "do a breakthrough."
So Kontorovich's view: a written thesis is still necessary, but no longer sufficient. Because the written word is no longer a reliable signal, the spoken word must carry more weight:
- More oral examinations, perhaps an annual check-in with a committee, not just a qualifying exam and a ceremonial defense.
- Students must own every idea in their thesis, whether it came from an adviser, themselves or an AI.
- The goal stays the same: deep command of a field and new knowledge that you, personally, understand better than anyone.
Who gets hired?
If most valued theorems soon come with AI help, what do departments select for? Kontorovich says he wants people who know lots of math, can communicate it, can teach it and are curious about the world.
That shifts the balance toward communication. He has known PhD students who could absorb and explain whole fields eloquently but couldn't produce a thesis; they used to leave academia. They might now thrive. The uncomfortable flip side: could a department pass on a brilliant but inarticulate future Riemann? His hope is that it hires both, because demand for rigorous training will grow.
Why this matters
- For mathematics: a problem experts didn't expect to see solved in their lifetimes fell, with a machine-checked proof. The bar for "impossible" just moved.
- For AI: math is how labs measure reasoning, and it's verifiable. Gains here signal gains elsewhere. Kontorovich notes the companies' valuations are about curing cancer, not theorems, but theorems are how they train and test.
- For education: credentials based on written output are weakening across fields, not just math. Oral, in-person demonstration of understanding is coming back.
- For the community: OpenAI has committed $5 million to a Simons Foundation initiative to help mathematicians examine AI-produced results.
Counterpoints and open questions
- Opaque provenance. OpenAI hasn't named the human editors of the quasi-Riemann paper, frustrating mathematicians who want to question them.
- Uneven quality. Science reports that some AI math papers are convoluted; the Navier-Stokes result was described as such, and one mathematician called another AI paper "AI slop" before extracting a genuinely new method from it.
- A toxic pace. Youness Lamzouri told Science he now feels pressure to publish fast to avoid being scooped by AI companies and amateurs prompting models: "It's all just too fast."
- Formalization isn't understanding. Lean certifies truth, not insight. Humans still have to learn why the proof works.
- Not the Riemann hypothesis. 7/8 is a huge step from 1, but the conjecture is about 1/2. The gap is enormous.
FAQ
Did AI solve the Riemann hypothesis?
No. It proved the weaker quasi-Riemann hypothesis: no zeros of the zeta function with real part above 7/8. The Riemann hypothesis says all nontrivial zeros sit at exactly 1/2. That remains open.
Why does 7/8 matter?
Before, no fixed zero-free strip was known; the guaranteed region shrank toward the line real part = 1. A fixed 7/8 boundary gives a uniform, much stronger handle on how primes are distributed, and the same result covers all Dirichlet L-functions, ruling out Siegel zeros.
How do we know the proof is right?
It comes with a Lean formalization that independent researchers have replayed and checked against the paper's statement, and human experts are working through the 200-page argument.
Should students still study math?
Kontorovich's advice is yes: training through the master's level is unchanged, and the PhD is shifting toward demonstrated, oral mastery and communication.
The bottom line
For 130 years since Hadamard and de la Vallée Poussin, mathematicians knew zeta's zeros stayed off the line at 1 but couldn't fence them away from it by any fixed margin. An AI just built that fence at 7/8, generalized it to a whole family of functions, and handed over a proof a computer can check.
It didn't crack the Riemann hypothesis. But it proved that the field's hardest territory is now in play.
The mathematicians most excited about it aren't predicting their own obsolescence. They're predicting a new job description: point the telescope, understand what comes back, and explain it to other humans. The machine can find the theorem. Only people can decide which ones matter.





