Claude Has a Silent Inner Voice — and Anthropic Just Found It

Anthropic discovers "J-space" in Claude — internal neural patterns where the model silently thinks about concepts without writing them down, echoing how conscious and unconscious processing works in the human brain.

Anthropic has published new interpretability research revealing that Claude has spontaneously developed what researchers call a "J-space" — a small collection of internal neural patterns that function like a silent scratchpad. Unlike chain-of-thought text that appears in output, these patterns operate entirely within the model's weights, allowing Claude to think about a concept without ever writing it down.

The finding draws a striking parallel to human cognition: just as most brain activity — breathing, posture, visual processing — happens below conscious awareness, most of what a language model computes is invisible. But the J-space, the team argues, represents something closer to consciously accessible thought. Named after the Jacobian, the mathematical tool used to discover it, this workspace was not programmed by Anthropic engineers but emerged on its own during training.

In experiments, Claude could report on what was in its J-space and modulate those patterns when asked. Multi-step reasoning tasks lit up intermediate steps in the J-space even when Claude did not output them — and those silent activations causally mediated its performance. When researchers blocked J-space use, the model still spoke fluently and recalled facts, but lost its capacity for deliberate reasoning. Once a concept like "France" entered the J-space, the model could flexibly recall its capital, currency, and continent — suggesting a shared internal representation.

On Hacker News, one commenter noted that taken alongside Anthropic's earlier SAE and Golden-Gate Claude papers, this suggests the company may be using a new representation-learning technique during training. Neel Nanda of Google DeepMind contributed an independent commentary to the paper, describing a small-scale replication on an open-weight model. The full research paper is available on the Transformer Circuits project site.

Sources: Anthropic Research Blog