Breaking

Anthropic Uncovers Hidden 'J-space' Thinking Area Inside Claude

July 8, 2026 at 00:41 EDT

Research & Papers
Foundation Models
Open Source

Anthropic's interpretability team said on July 6, 2026, that it had discovered inside its Claude model a small set of neural patterns, called "J-space," that hold and manipulate concepts without ever appearing in the output text. The work, titled "global workspace in language models," was published alongside a company blog post, a detailed paper and an explanatory video.

July 6, 2026 · Anthropic Interpretability

Inside Claude, a Silent "J-Space" Where the Model Quietly Thinks

Researchers report a hidden internal workspace — "J-space" — that emerged spontaneously during training. It holds only a few dozen concepts, yet it steers the model's multi-step reasoning before a single word is spoken.

<10%

of total activations — J-space is a small, privileged fraction

~ dozens

of concepts held and manipulated at any moment

8 → 6

answer flipped by swapping one internal pattern

Causal intervention experiment

Editing a hidden concept rewrites the output

internal pattern

"spider"

answer → 8

→

swapped to

"ant"

answer → 6

Internal representations don't just reflect thinking — they causally shape what the model says.

Reading the workspace — arithmetic surfacing in J-space

4 + 17

= 21

→

× 2

= 42

→

+ 7

= 49

The "Jacobian Lens" (J-lens) reveals which internal patterns push a word toward being spoken — letting researchers watch reasoning form layer by layer, before any text appears.

What it could enable

Detecting hidden intent — prompt injection, reward hacking
Visualizing intermediate steps of multi-step reasoning
Reducing misbehavior via Counterfactual Reflection Training

What it is NOT

Access consciousness only — reportable, reasoning-capable
No proof of phenomenal (subjective) experience
Research-stage; no change to everyday use of claude.ai

Suppress J-space and the higher functions vanish — yet basic conversation survives. It is a limited analogy to human thought, not a claim of a conscious mind.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.