Alpöge Claims a Generalized Riemann Proof with Claude

Levent Alpöge, a mathematician at Anthropic, announced on Friday a proof of the generalized Riemann hypothesis obtained with Claude; just a few hours later, Stanford released an independent draft by Alessandro Masullo. Difficulties have been reported regarding the human work, while Alpöge claims that his proposal benefits from comprehensive computer verification.
Hundreds of Results Await Confirmation
Hundreds of mathematical papers assume the validity of the Riemann conjecture, indicates Alex Kontorovich from Rutgers University. If proven, their conditional results would become established. A confirmation would provide specialists with an exact control of the distribution of prime numbers on the number line. The two proposals concerning the generalized version would extend from simple equations to high-dimensional geometry.
Prime numbers are also fundamental to modern digital encryption. In 1972, physicist Freeman Dyson discovered, during a tea with a mathematician, a correspondence between the spacing of Riemann's zeros and the energy levels of atomic nuclei. The relevant formula also appears in random particle movements, chaos theory, and calculations related to black holes.
Before these announcements, few high-level mathematicians were directly engaged with the problem due to a lack of a starting point. The discreet research conducted by Masullo and Brian Conrad was among the exceptions.
Problems Reported in Masullo's Draft
The initial project planned for Alessandro Masullo a year-long stay at Stanford, dedicated to improving his proof before making it public. Anthropic's announcement hastened the release of the available draft. Ravi Vakil, a mathematics professor at Stanford and president of the American Mathematical Society, shared it on his blog just before a board meeting. He presents this research as independent and conducted without the use of AI.
According to Vakil, Brian Conrad, a number theorist at Stanford, had been examining Masullo's reasoning for months. His detailed remarks had already been incorporated into the circulated version. However, about eight hours after his first message, Vakil reported several alerts regarding difficulties in the proof. Various AI models had helped identify them, with several mathematicians among the first to signal the issues.
Vakil does not take a position on the correctness of the chosen strategy. He specifies that acceptance requires resolving all problems to the satisfaction of the mathematical community. He has requested a rapid review of the draft, its translation into verification tools such as Lean, and the filling of gaps.
Lean Checks Deductions, Not Their Starting Point
On his website, Alpöge has posted a presentation of his results along with the corresponding code. He claims that his proof benefits from comprehensive computer verification. The reasoning itself is only available in Lean, spread across 14 data packages. According to his estimate, a printout would exceed 250,000 pages.
This length is partly due to the explicitness of elementary rules: a number different from a half must, for example, be less than or greater than it. Lean checks the validity of successive deductions according to mathematical rules, without admitting any unproven assumptions. However, a caveat remains upstream: the original statement may have been poorly translated into code. This initial check could be quick for experts, but its execution remains necessary.
Alpöge also acknowledges that he does not fully master the mathematics surrounding certain results on which Claude has outpaced him. He believes that the proof can be significantly shortened in the near future. A version intended for human reading is planned for the coming months.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Alpöge Attributes a Key Idea to Claude
Alpöge says he has maintained the ambition to solve this problem since his studies, which he expected to dedicate decades to. His work with the Anthropic model took place over the last few months. According to his account, he set the research direction and combined two approaches in a simplified case; once the complete problem was addressed, Claude had significantly outpaced him in the final part of the reasoning.
The claimed method extends an idea from mathematician Stepanov, previously applied to a related question in another system. It employs a process analogous to differentiation, while isolated integer values are normally not treated as a continuous curve. Alpöge and Claude say they have overcome this difficulty through a composition of transformations reproducing the desired effect. For a sequence of length X, the idea attributed to the model consists of composing exactly a number of transformations equal to the square root of X.
Alpöge considers this finding likely to surprise mathematicians. He also praised the choice of problems studied at OpenAI, claiming to have drawn essential ideas for his own proof from them. His claim goes beyond the initial hypothesis: it encompasses its generalized form, which concerns a family of related equations. He hopes to also address, with more time, the so-called extended version.
Riemann Sought to Clarify the Distribution of Primes
Formulated by Bernhard Riemann in 1859, the hypothesis is included in the selection of major problems established by David Hilbert in 1900. Since 2000, it has been accompanied by a reward of 1 million dollars from the Clay Mathematics Institute, as part of the seven Millennium Prize Problems. It concerns prime numbers, such as 2, 3, 5, 7, or 17, which are divisible only by one and themselves. The factorization into prime factors is unique, as illustrated by 15, the product of 3 and 5; Conrad compares their importance in number theory to that of forces in physics.
Euclid demonstrated around 300 BC that prime numbers are infinite in quantity. Their placement, however, appears random, and they become scarcer: there are 25 between 0 and 100, compared to 16 between 1,000 and 1,100. By the end of the 18th century, Carl Friedrich Gauss, still a teenager, used long tables to estimate how many are found in an interval. The actual count, however, deviates from this estimate in unpredictable ways.
A student of Gauss in Göttingen, Riemann wanted to calculate this deviation precisely and locate each prime number. His zeta function associates a value with a point identified on a plane by its horizontal and vertical coordinates. The points where this value equals zero are called zeros. They provide the corrections to be applied to Gauss's estimate.
Unable to examine an infinite number of points individually, Riemann conjectured that all so-called non-trivial zeros have a horizontal coordinate of one-half. A musical analogy associates the vertical coordinate with the pitch of a sound and the horizontal with its volume. In this representation, Kontorovich compares the hypothesis to instruments all playing at the same intensity.
The Announcement Follows Anthropic's IPO
By 2026, several open mathematical problems have already been solved by Claude. A solution to the Jacobian conjecture developed by the model was made public by Alpöge in July. In September, a dispute arose over the Navier-Stokes equations, which describe the flow of fluids and gases. Tristan Buckmaster, a mathematician at NYU, then suggested that OpenAI had used previous work done with Alpöge; in October, OpenAI published significant results concerning prime numbers.
The announcement regarding Riemann came the day after Anthropic's IPO, which took place on Thursday. Alpöge places his discovery late Sunday evening and says he subsequently sought the approval of Dario Amodei, the company's CEO. He regrets the proximity to the stock operation, denies any promotional intent, and asserts that further waiting before publication was not feasible.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.