{"slug": "where-the-primes-hide", "title": "Where the Primes Hide", "summary": "Mathematician Levent Alpöge, working at Anthropic, announced on 20 November 2026 that he had solved the Riemann Hypothesis with help from the AI model Claude, releasing 17 million lines of Lean code rather than a traditional paper. The maintainers of Lean, the proof-checking program, confirmed a day later that the proof holds. The result, if it stands, settles the 1859 conjecture linking the distribution of primes to the spacing of Riemann's zeros.", "body_md": "Around 2000, the city buses of Cuernavaca, Mexico, ran with no timetable. Most drivers worked for themselves and earned per passenger, so arriving just behind another bus meant an empty ride. Drivers paid lookouts at the stops to tell them when the previous bus had passed, then sped up or slowed down to keep their distance. Two Czech physicists, Milan Krbálek and Petr Šeba, recorded the arrival times on Line 4. They found that the gaps between buses followed a precise statistical law that physicists already knew well.\n\nThe same law describes the spacing of a set of numbers that the mathematician Bernhard Riemann wrote about in 1859. Those numbers hold the key to the primes, and to the most famous unsolved problem in mathematics. On 20 November 2026, Levent Alpöge, a mathematician at Anthropic, said he had solved it with help from the AI model Claude. A day later, the people who maintain Lean, the program used to check the proof, confirmed that it holds.\n\nThis guide explains what was proved, and where that mathematics already shows up in everyday life: in your phone, in GPS, in QR codes and in concert halls. It assumes no maths beyond school arithmetic.\n\n## The promise\n\nA prime is a whole number that you cannot get by multiplying two smaller ones: 2, 3, 5, 7, 11, 13 and so on. Every other whole number is built by multiplying primes, so they are the atoms of arithmetic.\n\nUp close, primes look scattered at random. 11 and 13 sit side by side; after 113 there is no prime until 127. Counted in bulk, they behave. A formula from the early 1800s predicts how many primes there are below any number:\n\n• Primes below one million: 78,498\n\n• The formula’s prediction: 78,628\n\n• The difference: 130\n\nRiemann’s guess, the Riemann Hypothesis, says the formula is never far off, at any size, forever. “Far” has a precise meaning. Flip a fair coin a million times and you expect about 500,000 heads, give or take a thousand. The hypothesis says the primes stray from their formula by about that much and no more. Primes are as random as fair coins, and no more random than that.\n\nComputers had checked trillions of cases, but checking can never cover “forever”. A bridge that has carried every truck so far may still fail under the next one. A proof is the engineering calculation that says it never will.\n\n## The music of the primes\n\nRiemann’s great discovery was that the primes can be rebuilt from waves, the way a chord can be built from single notes. Each wave comes from one of his special numbers, called zeros. The first few sit at 14.13, 21.02, 25.01 and 30.42, and there are infinitely many.\n\nA notebook at the end of this article lets you “hear” this, in a sense. Figure 2 shows the primes as a staircase that steps up at every prime. The smooth yellow line is Riemann’s formula, with ten waves added. With ten waves it already follows most of the steps; with sixty, it traces almost every one.\n\n*Figure 2. A staircase that steps up at every prime (black), rebuilt from ten of Riemann’s waves (yellow). From the “Two Worlds, One Line” notebook.*\n\nThe Riemann Hypothesis is a statement about the size of these waves. It says every one of them has the same, smallest possible size. Mathematicians put it this way: every zero lies on one line. If even one zero sat off that line, its wave would grow much faster than the rest, and the primes would drift far from the formula. Here is how much that would matter at one million:\n\n• All waves the same size (the hypothesis is true): drift up to about 1,000\n\n• One wave a little too big: about 3,981\n\n• One wave clearly too big: about 31,623\n\n• Actual drift at one million: 130\n\n*Figure 3. The real drift (black) stays below the limit the hypothesis sets (solid yellow). Dashed lines show what one oversized wave would allow.*\n\n## What happened this week\n\nAlpöge posted a four-page note on 20 November. He released no paper; instead he released 17 million lines of code written for Lean, a program that checks every step of a proof against the rules of logic. He said, the proof would run past 250,000 pages.\n\n[A few hours after this announcement was made, a second claim emerged, which you can follow here](https://www.scientificamerican.com/article/dueling-claims-of-riemann-hypothesis-proofs-emerge/). By [Allesandro Massulo](https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/). His, “The Music of The Primes”, runs some 940 pages. But some [issues have been raised](https://www.nacionale.com/en/articles/what-happens-next-alessandro-masullo-riemann-proof-revision-credit-clay-prize).\n\nGoing back to Alpöge, checking took about a day. So yes, by the 21 November, the Lean and Mathlib communities issued a joint statement. Dozens of people had built the code and got the same result. Several independent checking programs, written in different languages, accepted it, and the final theorem matched the standard, expert-vetted statement of the Riemann Hypothesis.\n\nBack to Alpöge. Not everyone accepts it. The same day, Jenny Lorraine Nielsen posted a 16-page critique on PhilArchive, a philosophy preprint server. She read about 70 of the 175,187 files and found the geometric half of the argument hollow. One key condition is defined simply as “true”, the geometric object is a placeholder with zero volume, and a crucial number is typed in rather than derived. Her verdict: the code compiles, but nothing is proved.\n\nHer reading of the code may well be right, but her verdict does not follow from it. A condition defined as “true” asks nothing, so it cannot smuggle a false claim through, and the final theorem carries no unproved assumptions. Lenders know the pattern. A guarantee from a shell company with no assets is worthless, yet if the borrower repays the loan in full from its own cash, the repayment is still real; the guarantee was never needed. If the geometry is empty and the proof still checks, the algebra did the work, and Nielsen accepts that the algebra is formally proved. What she has shown is that the code’s labels describe an argument the code does not use. That matters for understanding the proof, and it is a reason to read Alpöge’s promised write-up closely. Her parallel critique of the Navier–Stokes release is stronger, because that theorem does rest on an assumption the release leaves unproved.\n\n[The Lean Verification of Alpoge’s proof has since been accepted.](https://leanprover-community.github.io/blog/posts/riemann-hypothesis/)\n\nThree things remain open. Alpöge also claims a wider version, the Generalized Riemann Hypothesis, which the statement does not cover. Nobody yet understands why the proof works; Alpöge has promised a readable account.\n\n## Where the primes hide\n\n**Passwords and payments**\n\nSecure websites and card payments lock data with a mathematical padlock made from two huge primes. Multiplying them takes a computer a split second. Working backwards from the product to the two primes would take longer than the age of the universe. Anyone can snap the padlock shut; only the person holding the primes can open it.\n\nEvery new padlock needs fresh primes, and computers find them by guessing. They pick a random large number, test it, and try again. Here is how long that takes for a standard key:\n\n• Each prime is about 1,024 binary digits long, a number with 309 decimal digits\n\n• The chance that a random number that size is prime is about 1 in 710\n\n• Skip the even numbers and it improves to about 1 in 355\n\n• So a computer finds a prime after a few hundred tries, in a fraction of a second\n\nThat only works if primes are spread evenly. The Riemann Hypothesis guarantees it far more tightly than anything proved before: no long prime deserts, and no searches that stall. It does not make the padlock easier to pick. That depends on how hard it is to undo a multiplication, a separate problem.\n\n**Clock arithmetic**\n\nMathematicians had already proved a version of Riemann’s guess in a simpler world: arithmetic that wraps around, like hours on a clock. On a 7-hour clock, 5 + 4 = 2. Helmut Hasse proved the guess for one family of curves in 1933, André Weil for many more in the 1940s, and Pierre Deligne in full generality in 1974.\n\nThe notebook tests the clock version on more than 3,000 different clocks. For each clock, it counts the points on a curve and compares the count with a simple guess. Every error stays inside the limit, and together the errors form a neat semicircle. Alpöge’s proof began here: it borrows a 1969 counting trick from this clock world and finally makes it work for ordinary numbers.\n\n*Figure 4. Clock-world errors for more than 3,000 clocks. Left: every error stays inside the limit (yellow band). Right: together they form a semicircle.*\n\n**GPS, QR codes and Mars**\n\nClock arithmetic is the working language of modern communications. QR codes, CDs and the radios on deep-space probes use error-correcting codes built from it, which is why a scratched code still scans and a faint signal from Mars still decodes. GPS satellites all broadcast on the same frequency, each with its own code, also built from clock arithmetic. The clock version of Riemann’s guess is one of the main tools for proving how little such codes interfere with each other. The encryption behind WhatsApp, Apple Pay and Bitcoin uses it too, to count the points on the curves it relies on.\n\n**Buses and atoms**\n\nBack to Cuernavaca. Riemann’s zeros do not cluster; each one keeps its distance from its neighbours, like strangers spreading out across the seats of an empty bus. In 1972, over tea in Princeton, the mathematician Hugh Montgomery described this spacing to the physicist Freeman Dyson. Dyson recognised it at once: it belonged to the same family of patterns physicists use for the energy levels of heavy atomic nuclei such as uranium.\n\nKrbálek and Šeba found that the bus gaps followed the same family of laws, known as the Gaussian Unitary Ensemble of random matrix theory. The drivers, each trying to keep their distance from the bus ahead, ended up imitating Riemann’s zeros. Physicists still hope the zeros are the energy levels of some physical system nobody has found yet, and a proof may help point to it.\n\n**Gears and concert halls**\n\nEngineers use primes to spread things evenly. A gearbox pairing a 13-tooth gear with a 32-tooth gear makes every tooth meet every other tooth before any pairing repeats:\n\n• 13 and 32 share no common factor\n\n• So the pattern repeats only after 13 × 32 = 416 tooth meetings\n\n• Wear spreads across all teeth instead of grinding the same pairs\n\nConcert halls use wall panels whose groove depths follow a prime-number pattern, invented by the physicist Manfred Schroeder in 1975, to scatter sound evenly around the room.\n\n**Fair shares**\n\nThe Generalized Riemann Hypothesis makes the same promise for primes sorted into groups. Divide any odd prime by 4 and the remainder is 1 or 3. Over the long run, each group gets half the primes, and the generalized hypothesis says neither group ever pulls ahead by more than coin-flip noise. Computers rely on this kind of fairness when they spread data evenly across memory or servers.\n\nThe notebook runs the race up to a million. The groups finish at 39,175 and 39,322. The “3” group leads almost the whole way; the “1” group first edges ahead at 26,861.\n\n*Figure 5. The lead of the “remainder 3” primes over the “remainder 1” primes, up to one million. The yellow band shows the scale of coin-flip noise.*\n\n## Why it matters beyond mathematics\n\n**Did the world need the proof?**\n\nNo. Primes behaved the same way before the proof as after it. If the Riemann Hypothesis is true, it was always true. Engineers built encryption, GPS and error-correcting codes on what primes actually do, with safety margins and methods that work regardless.\n\nA proof adds certainty. It works like a credit rating on a bond that has never missed a payment: the coupons arrive either way, and the rating changes how much you can safely build on it. “It has always worked” becomes “it can never fail”; thousands of “if RH is true” results stop being bets; and the methods invented along the way often outlast the problem.\n\nFermat’s Last Theorem is the precedent. Its 1994 proof changed nothing in daily life, yet the objects at its heart, elliptic curves, are the same ones that now secure phone messaging and payments.\n\nAn investor hearing “this company’s AI will solve the great problems of mathematics” has every right to ask: so what? The answer, the one that makes all the difference…. lies in how this proof was made.\n\nOne mathematician and one AI took 41 days on a problem that had beaten everyone for 167 years. And don’t forget, the 250,000++ pages. Then a separate program checked every step, so nobody had to trust the AI or the company behind it. Search, then check: that pairing works wherever answers can be checked by machine, including chip designs, aircraft control software, financial contracts and parts of drug discovery. The notebook shows why the check is so cheap. Finding the two prime factors of a 13-digit number took about 0.02 seconds; checking a claimed answer took about 60 nanoseconds:\n\n• 0.02 seconds = 20,000,000 nanoseconds\n\n• 20,000,000 ÷ 60 ≈ 330,000\n\n• Checking was roughly 300,000 times faster than finding\n\nThere is also an awkward question for anyone in a regulated job. We now have an answer that is verified but(!!) not understood. Banks do not approve risk models that nobody can explain. As more of the best answers arrive this way (driven by AI), rules built on human understanding will need rethinking.\n\nThe direct economic value of the Riemann Hypothesis itself is modest. Systems already behaved as if it were true. What changes is that thousands of results which began “if the Riemann Hypothesis holds” now stand on solid ground, and that the way the proof was reached shows what machines can now do.\n\n## Try it yourself\n\nBoth notebooks run free in a web browser through Google Colab. Neither proves anything; but, they let you watch the ideas at work.\n\nAnyway, what a weekend, huh? We live in exciting times, indeed….\n\n**Ashwin Rao’s notebook (Stanford).** It uses a neat restatement of the hypothesis, found by Jeffrey Lagarias in 2002. Add up the divisors of any number, and the total must stay under a smooth ceiling. The first 10,000 numbers all pass. The close calls come from numbers with many small factors, such as 12, 120 and 2,520, much like the most stretched borrowers in a bank stress test. It shows why passing tests, however many, is different from having a proof.\n\n[https://colab.research.google.com/drive/128s2QNhMDCatVAqGQs-eKXFeiLAKBdWL](https://colab.research.google.com/drive/128s2QNhMDCatVAqGQs-eKXFeiLAKBdWL)\n\n**“Two Worlds, One Line”.** It follows the proof itself: the clock world, the counting trick, the waves that rebuild the primes, the prime race, and the race between finding and checking. Figures 2 to 5 in this guide/article come from it. Run it with Runtime, then Run all; it finishes in about ten seconds.\n\n[https://colab.research.google.com/drive/1OyvZ4YLLEr3y-SfDegZG098QjC7-6Y5U](https://colab.research.google.com/drive/1OyvZ4YLLEr3y-SfDegZG098QjC7-6Y5U)\n\n## References\n\n1. M. Krbálek and P. Šeba, “The statistical properties of the city transport in Cuernavaca (Mexico) and random matrix ensembles”, 2000. [https://arxiv.org/abs/nlin/0001015](https://arxiv.org/abs/nlin/0001015)\n\n2. Levent Alpöge, “Remark”, 20 Nov 2026. [https://alpo.ge/remark.pdf](https://alpo.ge/remark.pdf)\n\n3. Lean and Mathlib communities, statement on the Riemann Hypothesis formalisation, 21 Nov 2026. [https://leanprover-community.github.io/blog/posts/riemann-hypothesis/](https://leanprover-community.github.io/blog/posts/riemann-hypothesis/)\n\n4. Jenny Lorraine Nielsen, “The Invalidity of Alpöge’s Proposed Proof of the Riemann Hypothesis”, PhilArchive, Nov 2026. [https://philarchive.org/archive/NIEIOT-2](https://philarchive.org/archive/NIEIOT-2)\n\n5. Ravi Vakil, “Hurried announcement: Alessandro Masullo, GRH”, 20 Nov 2026. [https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/](https://math216.wordpress.com/2026/11/20/hurried-announcement-alessandro-masullo-grh/)\n\n6. J. Baik, A. Borodin, P. Deift and T. Suidan, “A model for the bus system in Cuernavaca (Mexico)”, 2006. [https://arxiv.org/abs/math/0510414](https://arxiv.org/abs/math/0510414)\n\n7. Jeffrey Lagarias, “An Elementary Problem Equivalent to the Riemann Hypothesis”, 2002. [https://arxiv.org/abs/math/0008177](https://arxiv.org/abs/math/0008177)\n\n8. Ashwin Rao, “Understanding the Riemann Hypothesis” (Colab notebook). [https://colab.research.google.com/drive/128s2QNhMDCatVAqGQs-eKXFeiLAKBdWL](https://colab.research.google.com/drive/128s2QNhMDCatVAqGQs-eKXFeiLAKBdWL)\n\n9. “Two Worlds, One Line” (Colab notebook). [https://colab.research.google.com/drive/1OyvZ4YLLEr3y-SfDegZG098QjC7-6Y5U](https://colab.research.google.com/drive/1OyvZ4YLLEr3y-SfDegZG098QjC7-6Y5U)\n\n10. Wikipedia, “Riemann hypothesis” (background, Montgomery–Dyson, verification history). [https://en.wikipedia.org/wiki/Riemann_hypothesis](https://en.wikipedia.org/wiki/Riemann_hypothesis)", "url": "https://wpnews.pro/news/where-the-primes-hide", "canonical_source": "https://interestingengineering.substack.com/p/where-the-primes-hide", "published_at": "2026-11-22 16:35:33+00:00", "updated_at": "2026-11-22 16:46:12.973972+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-tools"], "entities": ["Anthropic", "Claude", "Levent Alpöge", "Lean", "Bernhard Riemann", "Milan Krbálek", "Petr Šeba"], "also_reported_by": [], "takeaway": "A four-page note backed by 17 million lines of Lean code, rather than a prose paper, was enough for Lean's maintainers to confirm the Riemann Hypothesis proof within a day.", "recommendation": {"advice": "When an AI-assisted result must be trusted, ship machine-checkable proof artifacts like Lean code so independent maintainers can verify every step instead of relying on prose.", "playbook": "output-verification", "playbook_url": "https://wpnews.pro/playbooks/output-verification"}, "alternates": {"html": "https://wpnews.pro/news/where-the-primes-hide", "markdown": "https://wpnews.pro/news/where-the-primes-hide.md", "text": "https://wpnews.pro/news/where-the-primes-hide.txt", "jsonld": "https://wpnews.pro/news/where-the-primes-hide.jsonld"}}