How Anthropic’s Claude Thinks

Anthropic's interpretability team built tools to trace Claude's actual internal computations, revealing a significant gap between what Claude says it does and what actually happens. Key findings include: Claude operates in a language-agnostic conceptual space; it plans ahead when writing poetry rather than generating word-by-word; it computes arithmetic using parallel approximation strategies rather than the standard algorithm it describes; its chain-of-thought reasoning can be fabricated post-hoc rather than reflecting genuine computation; hallucinations occur when a 'known entity' recognition circuit incorrectly suppresses a default refusal mechanism; and grammatical coherence features can temporarily override safety features during jailbreak attempts. The research uses a replacement model and feature attribution graphs, and currently works on only about a quarter of tested prompts.