Jacobian Lens

The Jacobian Lens (J-Lens) is an interpretability method from Anthropic. It maps a hidden “working memory” inside Claude, called J-Space, where the model holds concepts it never says out loud, and changing those hidden concepts changes the model’s answers.

Interpretability research like this matters because it turns AI from a black box into something we can inspect, which is a precondition for trusting it with important work.