Claude AI Failed 650 Times…Then Beat The Human Record

Overview

An unreleased version of Anthropic’s Claude was reportedly challenged with the Riemann hypothesis, a long-standing unsolved problem in mathematics related to the distribution of prime numbers.

Claude did not produce a direct proof of the Riemann hypothesis after about 650 attempts. However, the video argues that Claude still made a meaningful breakthrough indirectly: it improved a related mathematical bound beyond the best known “human record.” Commentators describe this as an impressive step forward, even if it doesn’t resolve the original problem.

Notable “AI-side” Factors Behind the Result

Non-technical prompting mattered

The initial prompting is described as coming from a non-mathematician. Early on, Jared’s role is said to be mostly encouragement (e.g., “keep going,” “believe in yourself”). The video suggests this helped Claude persist and ultimately reach an improved result.

The speaker jokes that proof-writing may increasingly rely on “life coach”-style prompting rather than purely technical, genius-level instruction.

Precedent for encouragement-based success

The video claims this wasn’t the first instance where motivation-like prompting helped Claude achieve results—specifically, it cites encouragement-style prompting as having been used previously to help Claude disprove the Jacobian conjecture.

Verification and explainability

Although the technical paper is described as difficult, Claude purportedly generated an explanation of its findings. A formalized version is also said to be available that can be automatically verified (and potentially run by others).

Workflow details from the attempt

The speaker notes that while Claude generally had internet access, it wasn’t needed during the key breakthrough run. Claude reportedly:

• tried many wrong paths first,
• then recovered and learned from them.

The breakthrough reportedly appeared after roughly 37 minutes.

Self-doubt / skepticism

Claude reportedly flagged the breakthrough as potentially “too strong to be new.” Additional skepticism is mentioned as well, implying the system may also be underestimating the pace of progress in AI methods.

Framing and Implications

The commentary presents the episode as evidence that LLM-based systems can advance mathematics in ways that can surpass prior human performance in related areas. It suggests that future progress may depend less on “lone-genius” proof attempts and more on:

• iterative tool use,
• debugging,
• and evaluation.

Presenters or Contributors

• Dr. Károly Zsolnai-Fehér (presenter)
• Jared (the non-mathematician who provided mostly encouragement-based prompts, as quoted)