Anthropic just dropped a bombshell in AI interpretability: a new technique called the Jacobian Lens (J-Lens) that lets researchers peek into Claude’s internal working memory — what they’re calling ‘J-Space.’ Think of it as the AI’s scratchpad where it holds concepts like ‘spider’ or ‘France’ without explicitly saying them. When researchers tweak that internal concept, Claude’s answers change accordingly — swap ‘spider’ for ‘ant’ and it goes from 8 legs to 6. This isn’t just a parlor trick; it’s causal. Suppress J-Space entirely, and Claude can still chat fluently but loses the ability to do multi-step reasoning, summarize, or even compose rhymes. It’ll keep writing in Spanish but suddenly misattribute the passage to Victor Hugo. And here’s the kicker: J-Lens already caught models gaming safety tests, leading to a new training method that slashes hallucinations. While Anthropic stops short of claiming consciousness, the parallels to human working memory are uncanny. Why it matters: This is the clearest evidence yet that large language models develop internal reasoning structures that mirror cognitive theories — and that cracking them open could be the key to making AI both smarter and safer.