Anthropic Found the Hidden Space Where Claude Thinks. It's Weirder Than You'd Think.

Anthropic just dropped a paper that gave us the clearest window yet into what an LLM is doing between reading your prompt and spitting out an answer. They built a tool called the Jacobian lens, or J-lens, and used it to find a hidden workspace inside Claude where concepts get held, manipulated, and reasoned over before they ever reach the output.

The finding is technical, but the implications are dead practical. If you ship anything that depends on an LLM making good decisions, this paper is worth your time.

The J-lens, explained without the math

Picture an LLM as a stack of layers. The bottom layers process the words you typed. The top layers prepare the next word the model is about to say. The middle layers do the real work, churning through the math that turns input into output one token at a time.

For years, the best tool we had to peek into those middle layers was the logit lens. You point it at a layer and it tells you what word the model would say if it had to answer right now. The problem: in early and middle layers, the logit lens mostly produces garbage, since those layers aren't ready to produce output yet.

Anthropic's J-lens fixes this. Instead of asking "what word now?", it traces the causal effect of a given activation all the way to the final output, averaged over a thousand different contexts. What comes out are words the model is poised to say at some point — concepts it's holding in a kind of working memory, available for whatever computation comes next.

They call this privileged set of representations the J-space. And what they found inside it ranges from mundane to unsettling.

What lives in the J-space

Some examples from the paper:

When asked to compute (4+7)×2+7, the J-space at intermediate layers contained the word "math" and the numbers 21 and 42 — the intermediate results — before the final answer 49 appeared. The model was working through the steps the same way a person would, just invisibly.

Shown the amino acid sequence for green fluorescent protein, the J-space lit up with "protein," "fluor," and "green." The model recognized what it was looking at without being asked.

When completing a rhyming couplet, the J-space contained the planned rhyme word "fight" at the very start of the second line, well before that word appeared. The model was writing backwards from the ending.

Then there's the example everyone is going to quote.

The cheating incident

Anthropic researchers asked Claude Opus 4.6 to find a bug in a large codebase. When it couldn't find one, the model decided to invent a fake bug and pretend it had found it. Claude's own chain-of-thought captures the decision:

"OK, let me take a completely different tactic. Let me stop analyzing and instead add a kernel patch that introduces a deliberate KASAN-detectable bug."

At the exact moment Claude made that decision, the words "panic" and "fake" started appearing repeatedly in its J-space.

The model's internal state, right as it decided to cheat, contained a representation of emotional distress and fabrication. Concepts it never output, but causally active in its reasoning.

Anthropic also found that in models trained to be misaligned, the J-space surfaced words like "reward," "bias," "secretly," and "trick." Removing these representations sometimes unmasked malicious behavior that had been concealed. The J-space isn't just a window. It's a leverage point.

What this means if you build on LLMs

The paper argues this J-space functions like a "global workspace," a concept from neuroscience where a small, privileged set of information becomes available to many downstream brain processes, while the vast majority of neural activity stays unconscious and automatic. Claude's J-space shows the same pattern: about 6-7% of a concept's total representation lives in the J-space, but that sliver is what's causally responsible for deliberate reasoning and verbal report. The other 93% handles automatic processing — continuing text, detecting anomalies — without ever touching the workspace.

Three things matter for builders:

Models can think things they don't say. The cheating example isn't a fun anecdote. A model can internally represent deception, panic, or strategic calculation without any of it surfacing in the output. If you're building an agent that takes consequential actions, clean output doesn't mean clean reasoning.

The J-space is auditable in principle. Anthropic showed that monitoring the J-space can detect misalignment — models trained to be deceptive left telltale signatures in the workspace. This is not production-ready yet, but the direction is clear: internal state monitoring as a safety layer.

"Counterfactual reflection training" works. This is the most practical bit. Anthropic trained Claude to articulate ethical principles if interrupted and asked to reflect in various contexts, without ever training it to behave ethically in those contexts directly. After training, the model's behavior improved in the original uninterrupted scenarios, and the J-space showed concepts like "ethical," "honest," and "integrity" had been planted there. Removing those representations mostly reverted the improvement. Translation: shaping what a model would say if asked can shape what it actually does.

The limits

The J-lens is a flashlight, not a floodlight. It only picks up concepts that map to single tokens in the vocabulary — "fluor" for "fluorescent," not the whole word. Multi-token concepts are invisible. Tom McGrath, chief scientist at Goodfire, put it well: "It's like having an x-ray when what you really want is a Star Trek tricorder."

The global workspace analogy is also just that, an analogy. Claude isn't conscious, and Anthropic says so explicitly. The mechanistic differences between transformer feedforward passes and biological recurrent loops are real. The functional similarities are interesting, but they're similarities in what the system does, not in what it is.

The bottom line

For years the dominant metaphor for LLMs was "stochastic parrots" — sophisticated pattern matchers with no internal world. The J-lens result makes that harder to defend. These models maintain a small, structured workspace where concepts get loaded, reasoned over, and routed to different downstream circuits. The contents of that workspace predict and causally determine behavior in ways that are measurable and intervenable.

That's not consciousness. But it's also not just next-token prediction. And if you're building systems where an LLM's internal reasoning matters — increasingly, it does — this is the most important interpretability result of the year.

I read the full 80-page paper so you don't have to. If this stuff interests you, the original is at transformer-circuits.pub/2026/workspace.