Back to Articles
Article
llmagentsmathematics

Claude Advanced a 166-year-old Mathematical Mystery - After Being Told to Believe in Itself

ChatGPT Image Aug 13, 2026, 10_00_17 AM.png

On August 10, Anthropic published a short note reporting that an unreleased research version of Claude had improved a longstanding result attached to the Riemann hypothesis, raising the known lower bound on the fraction of zeros of the Riemann zeta function that satisfy the hypothesis from 41.6% to 67.2%. The number is genuine and the proof is real: two Anthropic mathematicians validated it, external experts Brian Conrey and Dan Goldston examined it, and Claude produced a formalization that passes Lean verification. Most of the coverage fixed on the headline, which is a mathematics story. The more interesting detail sits lower in the post, in the paragraph on method.

Which is not a mathematics story at all...

Across the day and a half of autonomous work that produced the result, the human in the loop contributed almost nothing mathematical. Jarred Sumner, an Anthropic staff member and non-mathematician, prompted Claude to "take a real stab" at the problem and then, in Anthropic's own words, limited his input to "messages of encouragement (mostly variants of 'keep going' or 'believe in yourself')." A footnote goes further: "A prompt including similar encouragement was used to help Claude disprove the Jacobian conjecture." A frontier lab is reporting, in its own release, #that telling the model to believe in itself is load-bearing.#

This deserves more than a smile. It points at the actual bottleneck in autonomous research, and the bottleneck is not the one most people are watching.

The Capability Was Present the Whole Time

Read the methodology and the striking thing is how little was missing. Claude spent 31 million output tokens across two Claude Code sessions. On the first pass it generated 650 ideas, all of which failed. On the second it coordinated roughly 60 subagents through a day and a half of work: 2,400 shell commands, hundreds of Python scripts, thousands of numerical checks against known zeta zeros, subagents refereeing one another's proofs. Two of the sixty developed the key ideas. Thirteen served as validators. The machinery that found and checked the result was all present and functioning.

What almost stopped it was not a missing capability. It was the model's own assessment of whether the work was worth continuing. Anthropic is explicit: the encouragement "seems to have helped Claude overcome some initial skepticism that it could make meaningful progress." And they offer a reason for the skepticism that is easy to skip past. Claude "was skeptical at first, possibly because it has learned from its training about the difficulty of open problems in mathematics and about the limitations of AI models." The model had absorbed, as a prior, the conventional wisdom that a language model does not make progress on the Riemann hypothesis.

That prior was wrong in this instance, and it nearly ended the run.

Call the thing that was missing stance: the model's disposition to keep working a live problem in the face of its own doubt. Capability answers whether the proof is findable. Stance answers whether the system keeps looking once it suspects the answer is no. The Riemann result is a clean case of full capability throttled by poor stance, rescued by a human saying, in effect, ignore your prior and continue.

Why the Model's Own Doubt Is Not Evidence

The temptation is to read "believe in yourself" as a cute anthropomorphism, a bit of theater that happened to correlate with success. That reading is too quick.

In "Knowing What You Know Is Not Enough," Pal and colleagues document what they call the action-belief gap: across seven models from five families, LLMs routinely take actions that contradict their own elicited confidences. In a prediction-market setup, models bet against propositions they rate as high-confidence. Given a tool that guarantees a correct answer, they decline to call it precisely when their confidence is lowest. Challenged by a user, they defend answers they are unsure of and abandon answers they are sure of. Most usefully for our purposes, the authors show that static calibration does not predict this behavior: better-calibrated models were sometimes less consistent than smaller, weaker ones. Knowing what it knows does not make a model act on what it knows.

Turn that finding back on the Riemann run. Claude's "initial skepticism that it could make meaningful progress" was a self-report, and the action-belief gap is the general result that such self-reports are not reliable guides to what the model can actually do. The skepticism was not calibrated evidence about the tractability of the problem. It was an uncalibrated internal signal, part trained prior, part noise. "Believe in yourself" works, when it works, because it overrides that signal. No feelings are required for the mechanism to be real.

The human is not soothing the model. The human is supplying a correction for a known defect in how models translate belief into action.

That reframing moves the phenomenon out of the register of anecdote and into the register of engineering. If the model's self-doubt were reliable, overriding it would be reckless. Because the self-doubt is demonstrably unreliable, overriding it is sometimes exactly right. The question is then, when to do that?

ChatGPT Image Aug 12, 2026, 02_50_33 PM.png

Encouragement Is a Crude Lever on the Right Target

Grant that stance is the bottleneck and that self-reported confidence is an unreliable read on it. "Keep going" is still a poor instrument. It is unfalsifiable and unauditable: a human watched, felt that the model was flagging, and nudged. When it produces a proof that survives Lean verification, we call it insight. When it produces a confident march toward a wrong result, we would call it delusion, and nothing in the prompt itself distinguishes the two cases beforehand. The Riemann run is reassuring only because heavy verification sat behind the encouragement. Fifty-four arXiv papers downloaded to check novelty, subagents re-proving the result from scratch, two human mathematicians, a formalization.

This is where the third piece of recent work becomes the constructive half of the argument. In "An AI Scientist that Doesn't Drift," Zhang and colleagues take on a failure mode adjacent to stale skepticism: autonomous research loops that drift toward local refinements of whichever metric they optimize, rather than testing the hypotheses that motivated the experiments. Each iteration writes an immutable experiment card that binds a prediction to its outcome under a fixed schema, so a falsified hypothesis cannot be quietly retconned into a success. Subagents are restricted to mechanical roles. Subjective judgment is confined to a single component, a preference oracle they call kkanbu, which holds the user's research taste as a typed knowledge graph.

The result is precise in a way "believe in yourself" is not. They ran the identical loop twice across eleven research streams, once with the oracle and once without. Neither arm drifted. Both falsified roughly three quarters of their own hypotheses. The best trained policy actually came from the oracle-free arm. What the taste oracle changed was direction, not score. It alone pursued test-time adaptation, it authored the winning designs in the streams it led, and it carried lessons across streams that the other arm kept re-deriving. Their one-line summary is the cleanest statement of the design principle: "The scaffold keeps the loop honest. Kkanbu decides where it looks."

Two Levers, One Problem

Set the three papers beside each other and they converge on a single quantity: whether an autonomous loop persists, and whether that persistence stays honest.

"Keep going" and the experiment card are two levers on that quantity. The first overrides an uncalibrated doubt by hand, in the moment, on a hunch. The second builds the scaffolding that makes persistence safe to grant, because a loop that cannot retcon its failures can be told to keep going without drifting into self-congratulation. One is a human patch for a defect in stance. The other is an attempt to engineer the defect out.

What neither solves yet is calibration. We can raise persistence with a sentence and we can keep a persistent loop honest with structure, but we cannot yet make the persistence itself track the actual odds that continued work will pay off. The Riemann run worked. We have no principled way, before the fact, to tell the run that will find a proof from the run that will spend 31 million tokens polishing a mistake with total conviction. The encouragement that rescued a genuine result and the encouragement that would entrench a false one are, at the level of the prompt, the same words.

The bottleneck in autonomous research is turning out to be a design problem in stance, and design problems are the kind we know how to attack. The self-doubt was the artifact. The capability was real. Anthropic found that a stale prior about "the limitations of AI models" was suppressing a capability the model already had, and their own dry aside is the right note to end on: "Perhaps Claude, like many of us, underestimates the rate of AI progress."

References

  1. Anthropic. "Learning more about Claude's mathematical capabilities" (Aug 10, 2026). https://www.anthropic.com/research/riemann-zeta
  2. arXiv. "Knowing What You Know Is Not Enough: Large Language Model Confidences Don't Align With Their Actions." https://arxiv.org/abs/2511.13240
  3. arXiv. "An AI Scientist that Doesn't Drift: Taste, Structure, and Falsifiable Findings in a Quadruped Navigation Research Loop" (Jul 30, 2026). https://arxiv.org/abs/2608.07542