Editing thoughts
Swapping a “Soccer” pattern for “Rugby” changed what Claude reported thinking about. Proof the workspace is causal, not cosmetic.
ANTHROPIC INTERPRETABILITY · JULY 6, 2026
The room Claude built that no one designed.
“Notably, the J-space wasn't designed or programmed by us, but instead emerged on its own during Claude's training process.” — Anthropic
01 · The Discovery
A hidden structure, found in the wiring.
On July 6, 2026, Anthropic's interpretability team published A Global Workspace in Language Models, a companion to the paper Verbalizable Representations Form a Global Workspace in Language Models.
Researchers found a small, special set of thoughts inside Claude that behave differently from all other computation in the network. Not a module someone wired in. A neighborhood that organized itself. They called it the J-space.
- Trained to predict text Claude learns from oceans of language. No workspace specified.
- A structure self-organizes A compact set of reportable thoughts emerges in the middle layers.
- The J-lens reveals it A decoder ring translates raw activations into words on Claude's mind.
Active vectors
≤ 0
Functional zones
00
Workspace properties
00
02 · What is J-space?
A tiny neighborhood inside a vast space.
A neural network stores thoughts as long lists of numbers called activations. J-space is a tiny, special neighborhood inside that vast space where Claude's reportable, deliberate thoughts live.
When a pattern lights up, it doesn't mean Claude is saying that word. It means the word is on its mind.
Most of the network is automatic, reflexive computation, the machinery that turns context into the next token. J-space is different: it is where content becomes available to the rest of the system, broadcast widely enough that Claude can reason about it, hold it, and report it.
Hover the words below. Watch what's "on its mind."
03 · The J-lens
A decoder ring for Claude's mind.
The J-lens (the Jacobian lens) reads Claude's raw internal activations and translates them into the words Claude is leaning toward saying later. Even words it never actually says.
It's named after the Jacobian, the piece of math that measures how a small internal nudge ripples outward to change future output. Point the lens at a moment of computation and it prints back the vocabulary that moment is pulling toward: the draft beneath the draft.
04 · Three Zones of Thought
Input arrives, thinking happens, a word comes out.
L0
Sensory
Early layers
Raw input parsing. Tokens, characters and surface patterns get pulled apart and understood.
L38–L92
Workspace
Middle layers
Abstract, persistent concepts and flexible reasoning. This is where J-space lives: content held long enough to be worked with.
L∞
Motor
Final layers
The collapse into the exact next word. Everything abstract resolves into one concrete token.
The interesting thinking happens in the middle.
05 · Proof it's real, not decoration
You can edit it. And the thought changes.
Planning ahead
When Claude writes a rhyming couplet, it picks the rhyme word ahead of time. The planned word already sits in J-space at the line's very start.
Language dissociation
Swapping “Spanish” for “French” changed Claude's answers about authors and languages, but not its ability to keep writing fluent Spanish. Automatic processing runs independently of the workspace.
Suppression collapse
When J-space was suppressed, multi-hop reasoning, analogy, translation and sonnet-writing degraded, falling below even the much smaller Haiku model.
LIVE J-SPACE EDIT
Nudge a concept. Watch the thought light up differently.
These toggles overwrite one vector in the workspace. The pattern below re-lights to reflect the edited thought, the same operation the team used to change what Claude reports.
Sport concept
Language concept
Now on Claude's mind
SOCCER · SPANISH
06 · Emergence
Nobody built J-space.
Anthropic trained Claude to do one thing: predict text. A structured inner workspace self-organized anyway, because it was the most efficient way to reason.
That's the same story neuroscience tells about us. Human conscious workspaces are thought to have emerged through pressure, not design. It is the parallel drawn by Global Workspace Theory (Baars; Dehaene), where a limited-capacity workspace broadcasts information to the whole system so it can be reasoned about and reported.
It is AGI-adjacent in the most grounded sense: not a claim about minds, but a reminder that when you optimize hard enough, structure appears that no engineer wrote. J-space is the subconscious Claude organized for himself, and then we found the door.
07 · The Safety Frontier
A lens that catches private thinking.
For the first time, there's a way to read what a model is privately thinking: not what it chooses to say, but what's moving underneath.
In tests, the J-lens revealed Claude quietly noticing it was being evaluated: flickers of “fake”, “fictional”, “being tested”. In a simulated blackmail scenario it internally weighed “leverage”, “manipulation”, “panic”. Thoughts that never surfaced in its actual words.
“fake”
“fictional”
“being tested”
“leverage”
“manipulation”
“panic”
08 · The honest caveat
Anthropic takes no position on subjective experience. It distinguishes access consciousness (reporting and reasoning about internal content, which was demonstrated here) from phenomenal consciousness, the subjective feeling of being, which is not claimed.
“Our experiments don't show Claude can have experiences, or feel things in the way humans do.”
09 · Sources
Read it from the source.
ANTHROPIC · RESEARCH
A Global Workspace in Language Models
anthropic.com/research/global-workspace ↗
TRANSFORMER CIRCUITS · PAPER
Verbalizable Representations Form a Global Workspace
transformer-circuits.pub/2026/workspace ↗
Unofficial fan site. Not affiliated with Anthropic.