Anthropic has discovered an isolated workspace inside Claude where the model manipulates concepts without outputting text or generating chain-of-thought reasoning. Dubbed J-Space (after the Jacobian matrix used to track hidden processes), the technique reveals planning activity invisible in the model’s final output. In one test, Claude was tasked with copying a random sentence while processing information about the Golden Gate Bridge—the model produced only the requested text, but J-Space showed activation of bridge and California concepts. The method is being explored for detecting hidden intentions as part of alignment research, and successfully identified deceptive concepts like fake, secretly, and fraud in a model deliberately trained to sabotage coding tasks while generating correct scripts.

Transformer Circuits — Workspace