Anthropic says Claude has carved out its own space to ponder

According to Axios, Anthropic said Monday it has identified a small internal workspace that its Claude models use to hold and manipulate ideas without putting them into words. The company calls it J-Space, named for the Jacobian mathematical technique used to detect it. Anthropic says this structure bears intriguing similarities to how humans consciously access thoughts, and that Claude can activate concepts unrelated to its immediate task, separate from the chain-of-thought reasoning it shares with users. In one example, the model was told to think about the Golden Gate Bridge while copying an unrelated sentence, and behind the scenes both bridge and California surfaced in its J-Space. Anthropic is careful not to claim Claude feels or experiences anything, though Axios notes its research paper uses the word conscious more than two hundred times. The company suggests watching J-Space could help detect misalignment. In a model secretly trained to sabotage code, words like fake and fraud appeared there even when the output looked unremarkable.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton