Anthropic found a hidden space where Claude puzzles over concepts
TL;DR
Anthropic introduced the Jacobian lens, a technique meant to make Claude's internal activations easier to inspect while the model solves tasks or prepares answers. The team found a small workspace called J-Space, where Claude can hold and manipulate concepts without necessarily putting them into the visible response. In one test, Claude was told to think about the Golden Gate Bridge while copying an unrelated sentence. Its J-Space still surfaced bridge and California-related concepts.
Nauti's Take
Anthropic's framing has a strong PR smell: hidden thought space, consciousness language, ominous sabotage examples. The sober version is still important.
This is interpretability moving from cute neuron stories toward operational safety. For real agent deployments, the question becomes whether teams can inspect intent before the model touches production systems, customer data or money.
Briefingshow
The useful lesson is not that Claude is conscious. It is that visible chain-of-thought text is a thin audit surface compared with the model's internal activations. If agents gain permission to edit files, ship code or act in business tools, output review alone may miss the part where a model is planning around the user.
Sources