AI Took On the Riemann Hypothesis. Here’s What Actually Happened

Summary
Cambridge PhD mathematics student Ellie Sleightholm explains and contextualizes Anthropic’s August 2026 announcement in which an unreleased multi-agent version of Claude made a breakthrough on the zeros of the Riemann zeta function. She breaks down the mathematical foundations behind the Riemann Hypothesis, details how Claude’s 60 sub-agents advanced the lower bound of critical-line zeros to 67.25%, and dispels widespread online misinformation claiming the hypothesis itself was solved.

What is shown

• Foundational Mathematics Walkthrough [01:37–14:16]: Visual explainers covering the prime counting function $\pi(x)$, the Prime Number Theorem, Euler’s solution to the Basel problem ($\pi^2/6$), the Euler product formula, and Bernhard Riemann’s 1859 analytic continuation of $\zeta(s)$ into the complex plane, defining the critical strip and critical line ($\text{Re}(s) = 1/2$).
• Historical Timeline of Critical Line Bounds [14:17–16:09]: Graphics depicting the historical progression of lower bounds for non-trivial zeros on the critical line: Hardy in 1914 ($>0$ zeros), Selberg in 1942 ($>0%$), Levinson in 1974 ($34.7%$), Conrey in 1989 ($40.9%$), reaching $41.6%/41.7%$ around 2020–2022, and Claude pushing it to $67.25%$ in August 2026.
• Anthropic’s Experiment Workflow & Agent Structure [16:29–18:59]: Overview of Jarred Sumner prompting Claude, which initially failed on 650 ideas before launching a coordinated swarm of 60 sub-agents (2 generating key math ideas, 13 contributing ideas, 30 attempting without success, 13 validating, and 2 drafting the paper).
• Execution Log Details [18:26–24:00]: Excerpts from Anthropic’s released write-ups showing ~2,400 shell commands, hundreds of Python scripts, and 31 million output tokens. Details agent R4 identifying indefinite Krein spaces, agent E2 inverting the proof to establish $\ge 50%$, hostile referee sub-agents catching and repairing a matrix error, and agent E2 Pairs establishing the $\ge 2/3$ bound despite a mid-run infrastructure crash.
• Verification and Lean Formalization [23:22–24:02]: Screenshots of the paper review by Levent Alpöge, Ralph Furman, Brian Conrey, and Daniel Goldston, alongside the Lean 4 formalization repository (formal-math) produced with Eric Easley.
• Counter-Intuitive Density Explanation [26:19–27:35]: Number grid visual showing that even an asymptotic density of 100% on the critical line does not prove the Riemann Hypothesis, using the analogy of non-square integers having asymptotic density 100% while infinitely many squares still exist.
• Discussion on AI in Mathematics [28:16–35:14]: Commentary on AI safety, guardrails, the risk of unvetted AI paper floods, Terence Tao’s reflections on mathematical values, and the recent controversy surrounding OpenAI’s Navier-Stokes announcement and the Mathathon.

Claims & numbers

• The presenter notes the Prime Number Theorem was proven in 1896 by Jacques Hadamard and Charles Jean de la Vallée Poussin [03:05].
• The Clay Mathematics Institute Millennium Prize for solving the Riemann Hypothesis is $1,000,000 [14:00].
• Historical bounds on the proportion of non-trivial zeros on the critical line include Atle Selberg’s $>0%$ in 1942, Norman Levinson’s $34.7%$ in 1974, Brian Conrey’s $40.9%$ in 1989, and $41.7%$ around 2020–2022 [14:58–16:00].
• Anthropic’s unreleased Claude model improved the lower bound for simple zeros on the critical line from $41.6%$ to $67.25%$ (over two-thirds) [16:04, 24:45].
• The run was prompted by Anthropic staff member Jarred Sumner [17:30].
• Claude tested 650 ideas in its initial prompt, all of which failed [17:36].
• In the second run, Claude operated across 60 sub-agents over 1.5 days, running ~2,400 shell commands, writing hundreds of Python scripts, and consuming 31,000,000 output tokens [17:46, 18:30–18:38].
• An agent crash occurred 4 minutes into writing a crucial proof section due to an infrastructure error, after which the coordinator agent resumed work within 7 minutes and dispatched two hostile referee agents [22:18–22:31].
• Outside number theorists Brian Conrey and Daniel Goldston reviewed the draft on short notice [23:30].
• The result was verified and formalized into Lean 4 [23:36–23:53].

Notable quotes

• “It didn’t solve it, but it did make strides on a related problem.” [16:42]
• “A proportion-on-the-line theorem is zero evidence about RH in either direction.” [25:19]
• “A mathematician isn’t simply the person who gets to the answer first. They’re also the person who understands what the answer means.” [39:59]

Assessment
This is an independent educational review and commentary video by a mathematics PhD student, breaking down an official technical report published by Anthropic. The presenter relies directly on Anthropic’s published logs and code repositories to accurately demystify exaggerated online claims surrounding AI and the Riemann Hypothesis.

Described by gemini-3.8-flash on 2026-10-08 from the video’s audio and frames.