Anthropic Finds a 'Workspace' Inside Claude — and Open-Sources the Tool That Found It

A 16-person Anthropic research team has identified a structure inside Claude models that functions like a cognitive workspace — a small, privileged slice of internal activity where reasoning surfaces before it becomes an answer. The paper, “Verbalizable Representations Form a Global Workspace in Language Models,” introduces the Jacobian lens (J-lens), a tool that measures how much each internal activation pattern would push the model toward outputting a given word later. Applying it revealed “J-space”: a subset holding only a few dozen active concepts at a time, accounting for roughly 6-10% of total activation variance per layer, while the much larger remainder of the model’s computation runs automatically and never surfaces to introspection. When asked what it’s thinking about, Claude reports what’s active in J-space — and the structure emerged on its own during training, not by design. Anthropic open-sourced the underlying code under Apache-2.0 and built an interactive demo with Neuronpedia applying the method to open-weight models.

The timing lines up with Anthropic’s GRAM research, published two days later, which routes dangerous knowledge into removable model compartments. Read together, the two papers point the same direction: Anthropic is investing in tools that see and control what happens inside a model, not just grade what comes out of it. J-lens locates where reasoning becomes visible; GRAM controls where specific knowledge lives well enough to delete it.

For companies evaluating AI vendors, that distinction matters more than another benchmark score. A model that scores well on a leaderboard has been graded on outputs. A lab that can point to a tool locating where its model’s reasoning actually happens is making a different, harder claim — one worth asking any vendor to back up before trusting an agent with a decision that can’t be undone by checking the output alone.