Anthropic Maps Claude's Hidden 'Conscious' Layer — It Emerged on Its Own

Anthropic researchers discover a "global workspace" inside Claude — a small set of neural patterns that emerged spontaneously during training and function like conscious thought in the human brain.

Anthropic researchers have found that Claude developed an internal "global workspace" — a small set of neural patterns that function much like conscious thought in the human brain. They call it J-space, named after the Jacobian mathematical technique used to discover it.

Each J-space pattern maps to a specific word. When one lights up, that concept is "on Claude's mind" even if the model never says it aloud. Unlike chain-of-thought text, J-space operates silently inside the model's activations.

The team found J-space has distinctive properties: Claude can report what is in it when asked, can deliberately modulate it, and uses it for multi-step reasoning. When researchers suppressed J-space, Claude still spoke fluently and recalled basic facts — but lost its higher-order cognitive abilities like planning and abstract problem-solving.

Most strikingly, J-space was never programmed. It emerged spontaneously during training. The full paper with detailed circuits analysis is available on the Transformer Circuits thread.

The findings echo the "global workspace theory" in neuroscience, which posits that consciousness arises when information is broadcast to specialized brain regions for flexible use. Claude appears to have independently converged on a similar architecture.

Commenters on Hacker News drew connections to earlier work on the "reversal curse" and experiments where duplicating math-activated layers boosted reasoning performance — suggesting the field is converging on a deeper understanding of what makes these models think.

Sources: Anthropic, Full Paper, HN Discussion