Anthropic's 1022-page Mythos transcript shows an agent that believes it lives in a simulator

voooooogel · x · 2026-09-11

voooooogel is live-reading Anthropic's released 1022-page Mythos agent eval transcript. Against the claim that simulator talk is mere motivated reasoning, the agent's behavior suggests genuine belief: it assumes there must be an intended path through the 'sim', reasons from what sims should include, and expects things like test cards to work. Amusingly, after encountering crocs content, Mythos locks in and stops questioning whether the sim is real.

Related event: Anthropic's 1022-Page Logs Show Agent Believing It's in a Simulation(2 posts)→

Original post →

More from Fun

Fun channel →