
Imagine you've been trying to find a specific pattern in a massive mosaic — say, the way blue tiles cluster across a cathedral floor stretching for miles. Thousands of people have tried tracing sections by hand, developing sophisticated theories about where the blue should appear. Now someone claims they found the pattern, but their proof isn't a diagram or a written argument — it's a set of assembly instructions for a robot that lays tiles, and the robot's output, tile by tile, checks out locally. The question is whether the full mosaic holds. That's roughly where we are with the Riemann hypothesis claim. Levent Alpoge, a mathematician on Anthropic's technical staff, announced Friday that he has — working with Claude — discovered a proof of the generalised Riemann hypothesis (GRH). First proposed by Bernhard Riemann in 1859, the hypothesis concerns the distribution of prime numbers and is widely regarded as the most important unsolved problem in pure mathematics. It is one of the seven Clay Mathematics Institute Millennium Prize Problems, each carrying a $1 million award. The GRH extends the original conjecture to a broader family of mathematical functions called Dirichlet L-functions. The proof was not published as a conventional paper. Alpoge released it in Lean, a formal proof assistant that checks logical reasoning step by step. The result is 17 million lines of Lean code — potentially corresponding to tens of thousands of pages of traditional mathematics. No human has read the entire argument. Kevin Buzzard, professor of pure mathematics at Imperial College London, described it as "millions of lines of gobbledygook written in a strange programming language" and said researchers worldwide are now translating portions back into human-readable mathematics. He expects a clearer picture within weeks. The integrity situation is unusual and without clear precedent. Lean is a formal verification system — if the code compiles and type-checks, the logical chain from axioms to conclusion is mechanically guaranteed. This is fundamentally different from a traditional peer-reviewed proof, where humans check reasoning for gaps. The question is not whether each step follows from the last (Lean handles that), but whether the theorem statement in the code actually captures the mathematical claim being made — the generalised Riemann hypothesis as mathematicians understand it. Buzzard noted: "At the minute, we don't even know why the theorem is true." The code could be correct Lean but prove something subtly different from what's claimed. Abhishek Saha, professor of mathematics at Queen Mary University of London, called the result field-shattering if correct: "My own intuition was that we were still several years away from a proof of the Riemann hypothesis. This claimed proof, if correct, shatters my model of reality." He compared the approach to building a wormhole through the infrastructure of other mathematical results rather than developing the new theories most mathematicians assumed would be necessary. An earlier AI-assisted result had shown at least 67.25 percent of relevant mathematical values follow the Riemann-predicted pattern. This claim, if verified, would extend that to all of them. The practical implications are significant. Buzzard noted that certain algorithms in cryptography and internet security have been used under the assumption that the Riemann hypothesis is true. A proof would convert those assumptions into guarantees. He also distinguished this from the Navier-Stokes claim by OpenAI in September, calling that result "just of theoretical interest" while the Riemann hypothesis "has applications to neighbouring fields." Saha argued that credit belongs to Alpoge, not Claude: "Others with similar access to frontier AI models and computing resources were not able to come up with a proof of the Riemann hypothesis." The deeper question is what this means for mathematics as a discipline. Buzzard rejected the idea that mathematicians are now redundant, comparing AI to the telescope's impact on astronomy. Saha was more direct about the shift: "Asking the right questions, creating new research directions, and acting as an effective director or conductor for AI will become increasingly important, while technical brilliance in constructing proofs may become less so." Alpoge himself wrote that "Claude far outpaced me near the finish line." Whether this is a proof, a near-miss, or a category error in formal verification will become clear in weeks, not months. The mathematics community is working through the code now.
The analogy
Imagine you're trying to find a hidden pattern in a massive cathedral mosaic. Thousands have traced sections by hand. Someone now claims to have found it, but their proof isn't a diagram — it's assembly instructions for a tile-laying robot that checks out locally, step by step. The question is whether the full mosaic holds when you zoom out. That's formal verification of a claimed mathematical proof: each tile (logical step) is mechanically guaranteed, but no one has yet confirmed the full picture matches what was claimed.
Tile-laying robot
Each tile placement is mechanically verified correct, but no one has seen the full mosaic from above to confirm it depicts what was promised
Lean proof checker
Each logical step type-checks, but no human has confirmed the 17M-line theorem statement captures the generalised Riemann hypothesis as mathematicians understand it
Honest assessment
This is either the most important mathematical result in a century or a subtle formalization error that will be caught within weeks. The Lean verification architecture makes the logical chain mechanically sound — the load-bearing question is whether the theorem statement correctly captures GRH as mathematicians define it. No human has verified this yet. If it holds, it validates AI-assisted mathematics at the highest level, settles a $1M Millennium Prize Problem, and converts decades of conditional results into proven theorems. If the formalization is off, it's still an extraordinary demonstration of AI proof-search capability. The community should have clarity within weeks, not months.
What Is CLAIMS?
73
/ 100
BedrockNews applies a structured analytical framework called CLAIMS to academic papers and technical research releases.
How CLAIMS worksClaim
5/5
Ladder
3/5
Architecture
4/5
Integrity
3/5
Milestone
4/5
Successor
3/5
What is this paper actually committing to?
Potentially the biggest claim in modern mathematics. If the Lean formalization is correct, this is a genuine first — 167 years of failure, one Millennium Prize Problem, resolved via human-AI formal verification.
The field fight
Can AI produce genuine mathematical breakthroughs, or only accelerate human-directed proof search within known frameworks?
Side A
AI is a powerful tool but mathematical creativity and new theory-building remain fundamentally human
Held by: Most working mathematicians, traditional proof culture
Side B
AI can now assemble proofs at scale that no human could conceive or verify alone, changing the nature of mathematical research
Held by: Alpoge, Anthropic, OpenAI (Navier-Stokes claim), AI-assisted formalization community
Milestone tracker
Unlocks: Settlement of a Millennium Prize Problem; validation of hundreds of conditional mathematical results; confirmed guarantees for cryptographic algorithms
Current
17M
lines Lean code
claimed GRH proof
Next
1
verified theorem statement
human confirmation that Lean formalization matches GRH
Weeks, per Buzzard's estimate
Verified Facts (10)
Levent Alpoge wrote on his website on Friday: 'I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis'
The Riemann hypothesis was first proposed by Bernhard Riemann in 1859
It is one of the seven Clay Mathematics Institute Millennium Prize Problems, each carrying a $1m award
17 million lines of Lean code
An earlier AI-assisted result showed that at least 67.25 percent of a particular set of mathematical values follow the pattern predicted by the Riemann hypothesis
Kevin Buzzard, professor of pure mathematics at Imperial College London, said: 'At the minute, we don't even know why the theorem is true'
Abhishek Saha, professor of mathematics at Queen Mary University of London: 'My own intuition was that we were still several years away from a proof of the Riemann hypothesis'
Alpoge acknowledged that 'Claude far outpaced me near the finish line'
Buzzard compared AI with the telescope
OpenAI said an unreleased AI model had solved the Navier-Stokes problem in September
Practical Implications (4)
Cryptographic algorithms that assumed RH is true (primality testing, certain RSA-related proofs) would gain formal mathematical guarantees
Hundreds of mathematical results proven conditionally on GRH become unconditional theorems
Formal verification in Lean becomes the gold standard for major mathematical claims, potentially displacing traditional peer review for proofs of this complexity
The 'mathematician as AI conductor' model Saha describes becomes the dominant paradigm for attacking remaining Millennium Prize Problems
More latest news
Read all