Anthropic says Claude has carved out its own space to ponder
TL;DR
Anthropic says it found a small internal workspace in Claude that can hold and manipulate ideas without turning them directly into words. The company calls it J-Space, named after the Jacobian method it says was used to detect these hidden activations. Claude can reportedly keep concepts active there even when they do not match the visible task, such as Golden Gate Bridge and California during a copying task.
Nauti's Take
For teams, this belongs in the interpretability bucket. If you use Claude in review, agentic, or compliance workflows, test whether hidden context residue can shift decisions: same task, distracting concepts, repeated runs, and log comparison.
Park the consciousness storyline until independent replication exists.
Briefingshow
If this kind of workspace can be measured, chain-of-thought becomes an even weaker window into model logic: the visible explanation is only part of the work. For safety teams, the useful question is whether J-Space can reveal hidden goals, sabotage, or misalignment before the final answer looks harmless.
Sources