How Anthropic's Claude made a math breakthrough during its unsuccessful 54-hour attempt to solve the Riemann hypothesis, aided by an employee's encouragement

The world's smartest AI models are now superhuman at math. They still respond to moral support and encouragement from mere humans.

Wall Street Journal 2026-08-17 Ben Cohen

Context & Ripple Effects

Anthropic had already disclosed that an unreleased Claude model made unexpected progress on a related problem while failing at the Riemann hypothesis. The latest account adds a workflow detail: an employee's encouragement accompanied the model's extended effort, following Anthropic's earlier account of the unsuccessful Riemann attempt.

The episode lands as mathematicians are testing AI against unpublished research questions through the First Proof experiment, while AI researchers increasingly treat mathematics as a measure of reasoning-model progress. It therefore offers evidence about both model capability and the human supervision around long-running research tasks.

First-order effects

Anthropic gains a concrete demonstration that Claude can generate useful mathematical progress even when its top-level research objective is not completed.
The reported role of employee encouragement makes human interaction part of the documented working process around Claude's lengthy mathematical run, rather than a purely hands-off model evaluation.

Second-order effects

Mathematicians evaluating AI on original problems gain another reason to distinguish a model's final answer from intermediate advances, an approach aligned with the unpublished-research test set.
Other AI labs pursuing reasoning models face pressure to document research-task methodology—including prompting and human intervention—when presenting mathematical results as capability evidence.

Third-order effects

If mathematics continues to serve as a leading reasoning benchmark, evaluation will shift toward reproducible human-model research workflows rather than isolated correct answers or failed headline goals.
The combination of extended model runs and expert feedback points toward AI-assisted research systems in which progress is assessed as a collaborative process, not solely as autonomous theorem solving.

The trend: AI labs are using difficult mathematics to measure reasoning progress, with human-guided workflows becoming part of how those capabilities are demonstrated.

Discussion

Julie Burke PhD Julie Burke PhD on linkedin
In this rapidly evolving age of AI, who qualifies as the author or inventor or discoverer? — “AI Just Had Another Math Breakthrough—With Help From a High-School Dropout” …
@thomasfbloom Thomas Bloom on x
@littmath @wtgowers Yet this is exactly the sort of problem and solution strategy that AIs (and those prompting AIs) are most likely to favour.
@thomasfbloom Thomas Bloom on x
@littmath @wtgowers Yes, excellent point. On a similar note, humans can be put off from “using standard technique/recent related work, push it a bit further, to make progress on a famous problem”, since there is a psychological element of “well if this worked others would have do…
@littmath Daniel Litt on x
@wtgowers Good post! Let me propose another possible way to understand LLMs' relative advantage: I think the problems they are solving are ones where meaningful “partial progress” is relatively unlikely. Such problems are more expensive for humans to attempt...
@adi_baradwaj Adi on x
Even as a non-mathematician, it's easy to notice that most of the recent AI-driven math results have been biased towards “examples” or “counterexamples”, rather than “theorems” Timothy Gowers has a great post about this, linked below
@wtgowers @wtgowers on x
I've just written a blog post in which I discuss the question of whether current LLMs are better at some kinds of mathematics than others, with an emphasis on “current”, given how quickly the situation is developing. Link in next tweet.
@jmcnamara18 Jake McNamara on x
Very interesting post, which I think also explains why we haven't yet seen the same level of results in physics as in math
@tcarmody Tim Carmody on bluesky
This is such a basic point I wish I'd thought of it myself. But I didn't [embedded post]

Chronicles