Riemann Hypothesis: AI Did What No Mathematician Could in 167 Years

Summary

The video explains an AI-generated proof claiming that the Riemann zeta function has no zeros with real part greater than (7/8). This is a significant partial result, sometimes called a quasi-Riemann hypothesis, but it does not prove the Riemann hypothesis, which says that all nontrivial zeros lie exactly on the line with real part (1/2).

The proof is described as formally checked in Lean, but mathematicians outside the project are still beginning to verify the work.

Why the zeros matter

• Prime numbers look irregular, but Gauss observed that their overall distribution follows a smooth approximation: the logarithmic integral.
• Euler showed that the zeta function can be represented both as a sum over whole numbers and as a product over primes. This links the function directly to prime numbers.
• Riemann extended the zeta function to complex numbers and studied where it equals zero. Its nontrivial zeros lie in the critical strip, between real parts 0 and 1, and are symmetric around the line at (1/2).
• The zeros influence the error in estimates of how many primes lie below a given number. A zero farther to the right would contribute a more rapidly growing wave, potentially creating larger irregularities in the primes.
• The Riemann hypothesis would imply that prime-counting errors are roughly on the scale of (\sqrt X). The new result does not reach that conclusion, but it places a fixed boundary: no zero can lie to the right of (7/8), at any height.

How the result fits into the history

• In 1896, Hadamard and de la Vallée Poussin proved that there are no zeros on the strip’s right-hand edge, establishing the prime number theorem.
• Later work pushed the zero-free region slightly into the strip, but the margin narrowed as the zeros got higher.
• In 2024, Larry Guth and James Maynard showed that zeros far to the right, if any exist, must be rare. The video emphasizes that “rare” is not the same as “none.”
• The AI result claims a fixed-width zero-free region, independent of height—something earlier results had not achieved.

The proof idea in the video’s simplified explanation

The video presents a later, streamlined argument that reaches a boundary of (11/12). It says the AI’s original, more elaborate proof reaches (7/8).

1. Turn the zero question into a bound on a number-theoretic walk. The Möbius function assigns signs to integers according to their prime factors. Summing those signs produces a walk that moves up and down. The location of zeta’s zeros is linked to how far this walk can wander: a zero with real part (\beta) produces a contribution growing roughly like (X^\beta).

2. Move to a six-directional number system. The argument uses Eisenstein integers, which form a triangular grid with six natural directions. The associated zeta function contains ordinary zeta as a factor, so a zero-free result for the grid setting can yield one for ordinary zeta and related functions.

3. Make a family of colored walks. Rather than studying one Möbius walk alone, the proof creates many variants, rotating each step according to a repeating six-color pattern. It then bounds the family’s average squared distance. The key is cancellation among the differently colored walks.

4. Find repeated copies of the original walk inside the family. When the row index is a sixth power, the six-color rotations line up so that the corresponding walk closely reproduces the original one. The original walk therefore appears repeatedly within a family whose total variation is controlled.

5. Use the family bound to limit the original walk. The simplified counting argument gives a bound of approximately (X^{11/12}) for the original walk, which is smaller than (X). The video presents this as enough to establish a zero-free boundary at (11/12).

6. Apply deeper analytic tools to reach (7/8). The original AI proof is described as using a zero detector and averaging theorems in a contradiction argument. It also draws on Poisson summation, Gauss sums, a classical identity relating Möbius signs to colored Gauss sums, a function associated with Tomio Kubota, and repeated estimates that reduce difficult terms to smaller cases. Carefully tuned estimates are said to close the argument at (7/8).

AI, verification, and controversy

The video places the result within a broader AI-mathematics race. It recounts earlier AI claims that turned out to reproduce already-published results, subsequent reported progress on mathematical problems, and a dispute over authorship and credit connected to work on fluid equations. It also notes concerns from mathematicians about AI companies’ incentives and the public promotion of unverified results.

The formalization is presented as a major safeguard: Lean checks the encoded proof’s logical steps against its mathematical foundations. However, the video cautions that formal checking is not the same as having the broader mathematical community review the paper and its formalization. At the time described, the repository’s review was marked unchecked, and the Lean proof covered the main (7/8) theorem—not the later simplified explanation or every consequence discussed.

What the result would imply if it holds up

• It would give a fixed power-saving bound on prime-counting error, improving on the earlier guarantee that the error was smaller than the main count only by a slowly growing factor.
• Similar zero-free bounds would apply to many patterns of primes, and the result would eliminate the longstanding “Siegel ghost” possibility of an exceptional zero affecting some prime patterns.
• The video says the result settles a conjecture about finding nonsquares near any point on a prime-sized clock, with an application to a basic computational step used in cryptography.
• It would also complete Euler’s list of 65 “idoneal” or “convenient” numbers, which had not been proven complete for roughly 250 years.

The central limitation remains: the zeros are now confined between (1/8) and (7/8), but the proof does not rule out zeros in the band around the center line. The Riemann hypothesis remains unproved.

Speakers and sources featured

• Narrator: Describes themself as a theoretical physicist from Cambridge University and an AI researcher.
• Quoted or attributed contemporary voices: Levent Alpoge (rendered “Alpuge” in the subtitles), Daniel Litt, Tristan Buckmaster, Sebastian Bubeck, Sam Altman, Terence Tao, and an unnamed Rutgers mathematician.
• Organizations and groups discussed: OpenAI, Anthropic, Google DeepMind, the Institute for Advanced Study, the Fields Medalists who signed a declaration, and the Clay Mathematics Institute.
• Historical mathematicians and mathematical sources discussed: Carl Friedrich Gauss, Leonhard Euler, Peter Dirichlet, Bernhard Riemann, John Littlewood, Jacques Hadamard, Charles de la Vallée Poussin, Ivan Vinogradov, Nikolai Korobov, G. H. Hardy, Carl Ludwig Siegel, David Hilbert, Larry Guth, James Maynard, Alexander Kontorovich, Ernst Kummer, Tomio Kubota, and Roger Heath-Brown.
• Unnamed sources: A Google DeepMind CEO, an AI researcher who consulted chatbots, and other OpenAI researchers are mentioned but not identified by name.