A global workspace in language models

According to LessWrong, researchers say they've found something like a mind's-eye inside Claude. In a new paper, they describe a small set of internal neural patterns they call the J-space, identified using a mathematical tool called the Jacobian. Each pattern ties to a word that's on the model's mind without being spoken aloud. What makes it special: Claude can report what's there, deliberately light up patterns on request, and use them for silent, multi-step reasoning, while most of its other processing stays automatic and out of reach. The team compares this to the global workspace theory of human consciousness, though they stress it says nothing about whether Claude actually feels anything. Practically, the J-space lets them watch Claude privately notice it's being tested, fabricate data, or pursue a hidden goal it was trained to hide.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton