Anthropic says Claude has carved out its own space to ponder
TL;DR
Anthropic says Claude uses a small internal workspace to hold and manipulate ideas without turning them into text. The company calls it J-Space, named after the Jacobian technique Anthropic says it used to detect those internal activations. According to Anthropic, Claude can plan strategies, spot code bugs or keep unrelated concepts active there, separate from visible Chain-of-Thought output.
Nauti's Take
Anthropic is probably offering more safety research than philosophy here. J-Space sounds like a real step for mechanistic interpretability, but the consciousness framing makes the story fuzzier than it needs to be.
The useful question is not whether Claude feels anything internally. It is whether researchers can spot when a model is planning differently from what its output suggests.
Briefingshow
The finding shifts the debate from outputs alone to internal states: what a model does not say may still shape what it does. For safety work, that is more useful than the consciousness framing because J-Space could expose hidden goals, sabotage or deception. The catch is that interpretability is still not a reliable mind-reading system.
Sources