AI Copium explains that Claude did not solve the Riemann hypothesis. The reported result instead raises a lower bound for the proportion of important zeros known to lie on the critical line, from 41.6 percent to 67.2 percent, by combining ideas from prior mathematical research in a new way.
An Anthropic employee gave an unreleased research system an open-ended instruction to attempt the larger problem. After hundreds of failed ideas, roughly 60 sub-agents, thousands of shell commands and about 31 million output tokens, the system arrived at the narrower advance while receiving little human direction beyond encouragement to continue.
Claude then searched for counterexamples, reviewed dozens of papers and asked fresh agents to reproduce the reasoning. Anthropic mathematicians and external experts reportedly reviewed the work, and a formalized Lean version passed its checker. The episode treats this layered verification as important evidence while acknowledging that the result was new and still awaited broader mathematical scrutiny.
Watch the original on YouTube