Anthropic Says It Found a Silent-Thinking Region Called J-space Inside Claude

On July 6, 2026, Anthropic published interpretability research claiming that a privileged set of internal representations called "J-space" emerged spontaneously during training inside its Claude large language models, functioning as a space for "silent thinking" where concepts are held and manipulated without being output. According to the official research page, the discovery relies on a new method called the "Jacobian lens" (J-lens), which identifies internal activation patterns that predict which words the model becomes more likely to say next.