Education
Mathematicians are amazed by ‘holy grail’ proof of Riemann hypothesis
Key Points
If the past year was the “mathpocalypse”, a claim that AI has helped produce a proof of the Riemann hypothesis is its most stunning horseman yet. Levent Alpöge at AI company Anthropic says he has used AI to solve this problem, which has sat at the heart of mathematics for nearly 200 years. This is not just the latest in the series of famed mathematics problems to fall to AI, but could be the biggest mathematical result of the century.
If the past year was the “mathpocalypse”, a claim that AI has helped produce a proof of the Riemann hypothesis is its most stunning horseman yet. Levent Alpöge at AI company Anthropic says he has used AI to solve this problem, which has sat at the heart of mathematics for nearly 200 years. This is not just the latest in the series of famed mathematics problems to fall to AI, but could be the biggest mathematical result of the century. Accordingly, mathematicians have received the news with a combination of exhilaration, bewilderment and caution.
“If true, it would, in my opinion. be the greatest achievement yet by AI, and one of the greatest proofs in the history of mathematics,” says Thomas Bloom at the University of Manchester in the UK. “The Riemann hypothesis has represented a major barrier in our current understanding of the way numbers behave – if the AI proof has correctly solved it, this would represent a huge leap in what we know.”
Rainer Dietmann at Royal Holloway University of London in the UK is similarly in awe. “If it turns out to be correct, this is not only huge, it’s the biggest result in pure mathematics ever…It’s like having found the Holy Grail,” he says.
Advertisement
Dating back to the 1800s, the Riemann hypothesis concerns the distribution of prime numbers. Because they cannot be factored into any smaller whole numbers, primes are like the atoms of numbers – they are the simplest units out of which any other number can be made. However, predicting when a prime will show up in a list of numbers is notoriously difficult. In 1859, Bernhard Reimann hypothesised a formula that would stunningly reduce that difficulty, but didn’t finish a proof before he died. Mathematicians have been trying to wrangle it ever since. After years of approximate results, including an AI-generated proof released by Open AI in October, Alpöge’s work may have finally confirmed Reimann’s centuries’ old intuition.
“It has captured the imagination, presumably because it is about something that we can all understand – the prime numbers – and because it has resisted the efforts of so many for so long. And it is important to mathematicians because it is so useful for other problems: a lot of theorems begin, “Assume the Riemann hypothesis…,” says Chris Daw at the University of Reading in the UK.
The new proof has not been formally reviewed by other mathematicians, which would be required for publication in an academic journal, and Alpöge has not released the full proof, which he described in a public statement as being over 250,000 pages long. Instead, he has released the version of the proof that has been adapted for verification in the proof-checking software Lean, which is 17 million lines long. Anthropic and Alpöge did not respond to New Scientist’s request for comment.
Michael Levitin at the University of Reading in the UK says hundreds of thousands of pages of Lean output will not be easily checked. “In proportion to the magnitude of this claim, we should be appropriately sceptical,” says Bloom. “Even after the Lean formalisation has been independently checked, it remains possible that the AI might be taking advantage of some subtle undiscovered bug in the code. I’d feel more confident once we have a combination of the Lean compiling correctly, with no bugs found, and also human experts [having] had a chance to discuss the ‘big picture’ of the proof and whether it seems plausible.”
“Mathematicians of course won’t be satisfied with 17 million lines of Lean. They’ll be satisfied when, and only when, they’re able to understand the proof themselves. And that will of course take time, since we only move at biological speeds,” says Scott Aaronson at the University of Texas, Austin. It is here that this new result, with its prospect of incredible historical significance, exemplifies the trouble mathematicians have had with AI over the past year, as it has been used to check one problem after another off their most coveted list of open questions.
Put simply, AI-assisted work benefits tremendously from prior work by human mathematicians, yet its end result is not immediately understandable to them. Coming to terms with the new proof will almost certainly take a long time, and it brings to the fore big questions about the future of mathematics. “Pretty much every mathematician will need time to process this. If AI can do this, what can it not do – in mathematics, at least?” says Daw. Many mathematicians already feel a sense of grief and anguish over how quickly AI tools have been solving problems that would have previously consumed a human mathematician’s whole career, and Dietmann says this new proof could have a similarly negative effect.
“What looked impossible one year ago has become reality. Many other problems in mathematics, computer science and related fields will fall as well, and it will become increasingly difficult for humans even to follow and understand,” he says. “While this progress is tremendous, it will massively amplify the despair by many in the mathematical community.”
Some mathematicians are more optimistic. “This is a very exciting time for pure mathematicians, and we are needed more than ever,” says Alex Bartel at Glasgow University. “AI is capable of spitting out proofs at a much faster pace than we are used to, but our task remains that of properly understanding what is going … and that work is as fulfilling as ever, but there is now a lot more of it. The announcement of a very long proof of the Riemann Hypothesis only reinforces my view.”
Mere hours after Alpöge’s announcement, president of the American Mathematical Society Ravi Vakil wrote in a blog post that mathematician Alessandro Masullo has also put forward a proof of a generalised version of the Riemann hypothesis. This proof was completed without the use of AI, and it is 940 pages long. It has also not yet been checked by other mathematicians nor confirmed using Lean.
This day will be remembered as a shocking one, and the aftershocks may continue. “I can only compare this achievement to the aliens with a vastly superior intelligence coming to the Earth and explaining to us all the sacred mysteries we ever dreamed of,” says Alexei Skorobogatov at Imperial College London. “Humanity managed to create an amazingly powerful, yet completely alien form of intelligence. The only event in history I can compare it to is the discovery of a nuclear bomb. Hopefully, AI will be put to better use!”
[Image text:] 28X10
POAND
M
M
243A0633
LA
ZNAOH
HOLL
AI (ORG)
Levent Alpöge (PERSON)
Thomas Bloom (PERSON)
the University of Manchester (ORG)
UK (LOCATION)
Rainer Dietmann (PERSON)
Royal Holloway University (ORG)
the Holy Grail (PERSON)
Advertisement Dating (PERSON)
Bernhard Reimann (PERSON)
Open AI (ORG)
Alpöge (PERSON)
Reimann (ORG)
Riemann (PERSON)
Chris Daw (PERSON)