AI and the World's Hard Problems · Life sciences · Anthropic · 2026-10-13
A global workspace in language models
The paper reports evidence that Claude has developed a distinction similar to 'consciously accessible' processing in human brains. It has a small set of internal neural patterns, called the J-space and found with a Jacobian-based technique, that play a special role compared with its other processing. Each pattern is tied to a word and signals that the word is on the model's mind, not that it is being output, and the J-space emerged during training rather than being designed.