Science1 publisher3 min readPublished
Lorraine mathematician confirms Claude's 67.25% bound on Riemann zeta zeros by another route
Youness Lamzouri has independently confirmed Claude's result that at least 67.25% of Riemann zeta zeros lie on the critical line, against 40% known since 1989. The hypothesis itself is still open. The case shows a working split for AI in mathematics, in which a model finds a candidate result and a specialist rebuilds it.
The Scientist · Science desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Anthropic described the attempt in an Aug. 10 preprint and blog post, and the work was done by a version of Claude the company has not released.
- Anthropic operator Jarred Sumner first prompted the model in late July, and it tried more than 600 lines of attack on the full hypothesis.
- Lamzouri did more than confirm the bound: according to Live Science, he also obtained new results on how prime numbers are distributed.
- OpenAI's Navier-Stokes proof came from around 10,000 AI agents, and almost a month after its release mathematicians still did not know whether it is correct.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- precedent Labs announcing AI proofs can now be held to a higher standard: an outside specialist re-deriving the result by another route. Claude's bound has had that check. OpenAI's Navier-Stokes claim has not yet had it.
- cost The scarce resource becomes specialist time. Agents can produce arguments far faster than number theorists can rebuild them, so every unchecked claim waits in line for that work.
- constraint Because the model version is unreleased, outside mathematicians can test the written argument. They cannot rerun the system that produced it or measure how often it succeeds.
Claude's figure is a floor. It says at least 67.25% of the non-trivial zeros of the zeta function sit on the critical line [4]. That is 27.25 percentage points above the 40% that, according to Live Science, mathematicians had known since 1989 [3][1]. Live Science calls the gain a huge leap [19].
The Riemann hypothesis, proposed by Bernhard Riemann in 1859, asks for more than that. It requires all of those zeros to lie on the line [18]. A floor of 67.25% leaves up to 32.75% of them unaccounted for [2]. Live Science notes that even a proof covering almost every zero would still permit exceptions, and the hypothesis permits none [14].
The bound came out of an attempt that had stalled. At the start of an August session, Anthropic's Jarred Sumner told the model: "Take a big leap of faith in your capabilities." [11] Claude replied, "That's not a confidence problem I can fix by believing harder." [12] The team then switched to a smaller target, the minimum share of zeros on the line [2]. "When you find a huge mountain, like the Riemann hypothesis, and you cannot prove it, then you settle on lower objectives," Lamzouri said [13].
The strongest evidence in the story is the check. Lamzouri works in this field at the University of Lorraine, and he confirmed the bound using a different and more intuitive approach [5]. I'd count that as the mathematical version of an independent replication. A second derivation by another route need not share the first one's errors, so when the two agree, that says more than a careful reading of a single argument could. "It's like you have an archaeological site and you bring in big machines and they extract a treasure because this is what we want: the artifact," Lamzouri told Live Science [7]. He added that "humans usually do it very carefully because they want to understand how it came to be that this artifact is buried there," and that "this is what happened with Claude and me." [8]
Live Science's case for why this matters rests on a comparison with OpenAI's Navier-Stokes claim. According to the outlet, the Clay Mathematics Institute posed its seven Millennium Prize Problems in 2000 to draw out new techniques and understanding, and solving them by brute-force computation was never the purpose [16]. Live Science argues that a proof humans cannot understand or use holds little value to mathematicians, and says that appears to be true of OpenAI's proof [17].
The thing this doesn't tell you is how often the pattern holds. This is one sub-problem, checked by one specialist. On that evidence I think Lamzouri's division of labour is the right working model: the machine extracts a candidate, and the result counts only once a specialist has rebuilt it. The Live Science account does not say whether Anthropic's preprint or Lamzouri's confirming work has been refereed.
What to watch
- Whether Anthropic's preprint and Lamzouri's confirming argument pass peer review, and whether other number theorists push the 67.25% floor higher.
- Whether mathematicians reach a verdict on OpenAI's agent-built Navier-Stokes proof. That would show whether output at that scale can be checked at all.
- Whether Anthropic releases the Claude version used, so outside groups can test how reliably it produces checkable results.