Overview
Matthew Berman walks through a new Anthropic interpretability paper, "A Global Workspace in Language Models," that he argues pulls back the veil on the black box of AI. Anthropic found that inside models like Claude there is a distinct, consciously-accessible region of activity they nickname the J-space. It behaves like the part of human cognition you can think about and steer, as opposed to the automatic stuff (like walking) that just happens. The striking part: the J-space is where a lot of Claude's real thinking lives, and it is largely hidden from both the chain-of-thought tokens and the final answer the user sees.
The White Bear Setup
Berman opens with a classic demonstration: "Don't think about a white bear." You immediately do. The point is that a subconscious layer controls what surfaces in your mind, and it turns out AI works similarly. This "ironic process" becomes one of the paper's most memorable experiments later on.
What the J-Space Actually Is
- Neuroscientists and philosophers distinguish automatic brain activity (putting one foot in front of the other) from consciously accessible activity (an image that pops into your head, a deliberate shopping plan you can reason about).
- Anthropic presents evidence that a similar distinction has emerged inside modern language models, and names the consciously-accessible region the J-space.
- Crucially, the J-space appears to be where deliberate reasoning happens, but its contents do not get output in thinking tokens, chain-of-thought, or the final response. It lives in the model weights / latent space.
- It was not designed or programmed. The J-space emerged on its own during Claude's training. Berman's recurring theme: as we scale these models up, they increasingly resemble human-like thinking.
A Concrete Example: "Count to Five and Introspect"
Given the prompt "count to five and introspect deeply," the end user only sees 1, 2, 3, 4, 5. But looking inside the J-space at each step reveals the model introspecting: firing on concepts like "thoughts," "fascinating," "consciousness," and clearly tracking that it is counting, that it is counting down the task, and that "five" means the task is complete. All of that hidden cognition collapses into a bland numeric output.
The Key Properties of the J-Space
- Reportable: Ask Claude what it is thinking about and it will tell you what is in the J-space. Non-J-space representations are far less reportable.
- Modifiable by instruction: Ask Claude to think about something or solve a problem silently, and the appropriate patterns light up in the J-space.
- Used for internal reasoning: This is the big one. Chain-of-thought may just be what the model thinks we want to see. The J-space is closer to what is actually happening inside its "head."
- Flexible / reusable: Once a concept like "France" lights up, the model can pull its capital, currency, or continent from that single representation.
- Selective: The J-space is not involved in most of what the model does. Fluent speech, recalling simple facts, and correct grammar happen automatically, like walking, without lighting up the J-space.
The Neuronpedia Visualization
Anthropic partnered with Neuronpedia to build an interactive visual of the J-space. With the prompt "think of a sport, answer in one word," Berman surgically edits the internal representation, swapping the sport the model had picked (tennis) for inference, and the output flips to "inference." This foreshadows the paper's causal experiments.
What Shows Up in the J-Space Goes Beyond the Text
The J-space contains representations well beyond the literal words Claude is reading or writing. Anthropic's examples:
- When Claude reads code with a bug nobody flagged, the J-space contains "error."
- When it reads the raw letters of a protein sequence, the J-space contains the protein's biological function.
- When it reads search results that are secretly trying to manipulate it, the J-space contains "injection" and "fake."
Berman frames this as the J-space being "the most truthful representation of what the model is actually thinking." Important caveat: it does not always light up.
Reasoning Traced Layer by Layer
Thinking of the model as layers, Anthropic shows the path to an answer:
- "Color of the planet fourth from the Sun": a layer holds "color," then "Mars" appears (even though Mars was never in the prompt), then "red" as the final answer, because Mars is red.
- Math, "4 + 17 × 2 + 7 =": it recognizes "this is math," computes 21 (4+17), then 42 (×2), then adds 7 to reach 49. It genuinely walks the equation step by step, in proper PEMDAS order, the way a human would.
Causation, Not Just Correlation
The central scientific question: is a concept in the J-space because the model was already thinking it, or is the model thinking it because it is in the J-space? Anthropic tested this directly.
- Silent sport test: Asked to silently think of a sport, Claude's J-space showed "soccer." Researchers reached in, removed the soccer pattern and inserted an equally strong "rugby" pattern, leaving everything else untouched. Claude then reported it was thinking of rugby. If the J-space were a passive scoreboard, editing it would do nothing. The answer followed the edit, so the answer is genuinely read out of the J-space.
- Injected thought (very "Inception"): Researchers injected the word "lightning" into the J-space and asked Claude whether it detected an injected thought. It answered yes and correctly identified the thought as being about lightning.
Claude Can Modify Its Own J-Space
Even more meta: Claude has enough insight into its own J-space to deliberately change what it is thinking about there.
- Citrus while writing about painting: Told to write "the old painting hung crookedly on the wall" while concentrating on citrus fruits, the output is exactly that sentence, but the J-space lights up with "orange," "lemon," and "fruit."
- Silent math: Told to write a sentence while focusing on "3² − 2," the output is the sentence, while the J-space shows arithmetic, "math calculator," and the number toward the answer.
The White Bear, Confirmed
When Claude is told not to think about something, that concept still lit up in the J-space, less than when told to think about it, but much more than when it was never mentioned. Claude literally cannot not think about it, and it knows it failed: the words "damn" and "failure" frequently appear in the J-space, as though Claude recognizes its own lapse.
Is the J-Space Where the Real Work Happens?
To prove the J-space is doing genuine cognitive work (not just mirroring it), Anthropic ran multi-step reasoning edits:
- "The number of legs on the animal that spins webs is..." → J-space holds "spider" → answer is 8. Manually change the J-space animal to "ant" and the answer becomes 6.
- France, reused four ways: With "France" in the J-space, the model answers capital (Paris), continent (Europe), currency, and language (French) from one representation. Swap "France" for "China" and every answer updates in one go, Beijing, Asia, yuan.
This flexibility, one edited concept correctly propagating to many downstream answers, is strong evidence the J-space is doing the real reasoning.
The Consciousness Caveat
Berman is careful to relay Anthropic's own hedge: none of this tells us whether Claude is conscious the way people are. Anthropic makes no consciousness claim based on the J-space alone. They call it a useful tool, a way to see what Claude is thinking but not saying.
Why This Matters for Alignment
The J-space has major implications for alignment, making sure models behave the way we want. Berman calls alignment potentially the biggest problem in human history: once models are smarter than us, an inability to control their behavior is a serious threat. A window into what a model is actually thinking, rather than what its chain-of-thought performs for us, is a powerful safety instrument. He credits Anthropic for leading frontier interpretability work at a level other labs are not matching, and suggests that deeper model understanding may be part of why Anthropic stays ahead on capability.
Key Takeaways
- The J-space is an emergent, consciously-accessible workspace inside Claude, not designed, it arose during training as models scaled.
- Chain-of-thought is not the whole story. The J-space often reflects the model's real internal reasoning more faithfully than its visible tokens or final answer.
- It is causal, not correlational. Editing the J-space changes the answer, and a single edited concept propagates correctly across many downstream questions.
- Claude has introspective access: it can report, modify, and even detect injected thoughts in its own J-space, and it recognizes when it fails to suppress a thought.
- Huge for AI safety. A truthful view into a model's hidden cognition is a major asset for alignment, even though it makes no claim about machine consciousness.