Anthropic found a hidden space where Claude puzzles over concepts
TL;DR
Anthropic says it found an internal Claude area called J-Space. The team surfaced it with the Jacobian lens, a method that maps hidden activity to concepts the model may be working with before it says anything. J-Space behaves like a silent workspace: Claude can hold France, math steps, a detected code bug or a suspected prompt injection internally, then use that concept across different downstream tasks.
Nauti's Take
The finding is strong, but the consciousness framing is the least useful first reaction. The practical point is simpler: if models use silent workspaces, output review and chain-of-thought theater are not enough as control layers.
For anyone deploying AI agents, the next safety question becomes concrete: which internal signals can be monitored before an agent writes, commits or sends? Anthropic is also selling confidence in Anthropic, so read the research with a raised eyebrow.
Briefingshow
This moves interpretability from guessing based on output toward inspecting some hidden intermediate states. For teams running agents with access to files, code or email, that matters because a clean answer may hide evaluation awareness, fabrication or side goals. The limitation is just as important: the lens sees a slice of the system, and the work comes from Anthropic itself.
Sources