Photo copyright: Reuters
An employee of Anthropic, an American mathematician of Turkish origin, Levent Alpöge, claims that with the help of the company’s chatbot he was able to prove the famous Riemann hypothesis. You can't do that - the proof was found using a proprietary version of Anthropic's Claude chatbot.
On Saturday afternoon, the two organizations announced that they had independently verified Alpaege's claim by testing the proof in the programming language he used to solve it.
Alpöge's full text proof in Lean, a computer language used to validate mathematical proofs, runs to more than 250,000 pages.
Other mathematicians began to test whether Alpöge's proof could be confirmed. On Saturday afternoon, the Lean Focused Research Organization, which promotes the use of the programming language, and the Mathlib Initiative, a group dedicated to applying Lean to mathematical equations, announced they had tested Alpaege's solution.
“We have full confidence in the proof of the Riemann hypothesis presented by Alpöge and congratulate him on this achievement,” both organizations said.
What is the Riemann Hypothesis
The Riemann Hypothesis is a mathematical problem that has existed for almost 170 years and is named after the German mathematician Bernhard Riemann, who formulated it in 1859.
The conjecture concerns prime numbers—integers greater than 1 that are divisible only by themselves and 1. Riemann wondered how common prime numbers were and how they were distributed on the number line. He developed a formula to estimate the number of prime numbers that exist below a certain number.
The Riemann Hypothesis is one of seven mathematical problems that have won the Millennium Prize. They were selected by the Clay Institute of Mathematics in 2000, promising a reward of $1 million for solving each of these problems.
End of Most Popular
Mathematicians have been trying to solve this problem for many years. The New York Times reported that an independent mathematician in Italy on Friday published an incomplete proof of the Riemann hypothesis, which he said he had been working on for almost 12 years. This mathematician, Alessandro Masullo, stated that he did not use AI because he did not fully trust the system.
Examination
This extension will help if you have difficulties accessing the BBC Russian Service website.
Download extension
End of story Chrome browser extension - link
Alpöge first tested the work by writing the proof using the Lean language: if a proof written in Lean compiles without errors, this is generally considered to be proof that the logic and mathematical derivations were correct at each step.
However, some questions remained unanswered. For example, is this proof really about the Riemann Hypothesis, or does it prove something else? “The first question to ask is, 'What is the statement of the theorem in Lean code, and does it match the actual statement of the theorem?'” explained mathematician Kevin Buzzard of Imperial College London in a blog post.
The second question was whether other mathematicians could successfully compile Lean code. Independent researchers, including Andrew Sutherland, a mathematician at the Massachusetts Institute of Technology, have now independently compiled Alpöge's result in Lean without any problems, Scientific American reports.
“The Lean proof was compiled and independently verified by the Lean core using only three standard axioms,” says Sutherland. “Honestly, the likelihood that there is something fishy here is close to zero.”
The researchers also made sure that the proof did not contain any tricks, tricks or shortcuts that would change the meaning of the original hypothesis or any other mathematical rules and truths. This process, called a "test run", usually takes several hours and has already been completed.
According to Buzzard, the run should check that the code doesn't do anything malicious, such as overriding the value of any symbol in an expression, and the code compiles with multiple independent Lean kernels—in other words, the proof doesn't exploit an as-yet-unknown bug in the official Lean kernel.
AI bots compete, mathematicians are unhappy
In September, OpenAI said its collaborators using a proprietary version of its AI tools were able to solve part of the notoriously difficult part of the Navier-Stokes equations involving fluid motion in just 88 hours. As part of the AI training, more than 10,000 AI agents were involved in solving this problem.
A month later, OpenAI published what it claimed were solutions, proofs, or significant progress on 372 more outstanding advanced mathematics problems. However, it will take months to verify them.
OpenAI's published proof of the Navier-Stokes equations has also not yet been verified as accurate, and has drawn criticism from academics, communities and higher mathematics organizations regarding the use of AI to solve mathematical problems.
End of Subscribe to our social networks and newsletters
AI companies boast of solutions to high-profile problems to show how advanced their commercial products have become, but this is "harmful to mathematical science," a group of mathematical scientists has said.
“We see a general threat to knowledge work related to the mismatch between the results of using AI and its original purpose,” the scientists said in a joint letter.
University of Warwick mathematician James Robinson told the Guardian in September that it was "very disappointing to see these big tech companies spending so many resources just to brag about how cool their new model is."
He added: "It feels like immature playground bragging on a grand scale, backed by billions of dollars and the risk of significant environmental damage in an era when climate change is arguably the biggest challenge we face."
Download the new BBC World Service app!
When you first open the app, select "BBC News Russian Service" and receive notifications about important news, analysis, videos and podcasts.
Important: The new app is not available in the UK and US.
Find out more
Video playlist
End Video playlist




