Anthropic Finds Hidden Thought Space 'J-space' Inside Claude

On July 6, 2026, Anthropic published interpretability research reporting the discovery of a functional workspace inside its Claude language model, dubbed "J-space," that holds and manipulates thoughts which never appear in the model's output. Using a new technique called the Jacobian lens (J-lens), the company mapped internal activations to "the words most likely to be output next," showing that the resulting set of vectors forms a single shared workspace within the model. The structure resembles the global workspace theory in neuroscience — the hypothesis of a shared workspace in the brain responsible for conscious access.