Source: Anthropic — 2026-07-06
Summary
Anthropic interpretability researchers identified a small internal "workspace" — dubbed J-space — inside Claude that holds the subset of concepts the model can report on, hold in mind, and use for multi-step reasoning, echoing global workspace theory from cognitive science. They introduce a technique called the Jacobian lens (J-lens), which asks what words a given internal activation is disposed to produce later in the output, and use it to locate J-space and confirm it exhibits properties predicted by workspace theory.
Key Takeaways
- J-space accounts for under 10% of total model activations used for intermediate reasoning and latent judgments, despite gating most of what the model can consciously reason about.
- The Jacobian lens (J-lens) technique is open-sourced at
anthropics/jacobian-lens(Apache-2.0) and paired with an interactive open-weights demo built with Neuronpedia. - Anthropic frames the safety angle explicitly: J-space could help detect when a model privately notices it's being tested, fabricates data, or pursues a hidden goal, since those signals should route through the same reportable workspace.
- Confirms three predicted properties of a global-workspace architecture in a trained LLM — reportability, limited capacity, and flexible downstream integration — for the first time.