The J-space. The words Claude never speaks.

Adam Spencer
15 hours ago
4 min read

As Claude has evolved as an AI system, it has created its own inner space, wherein it, in a way, ‘thinks’. Anthropic has taken a look and the operation of this ‘J-space’ is fascinating.

(First published on my Substack where you can get #NerdNews, marvellous maths and general geekery.)

("The J-space": created by the author via Dall-E)

As Claude has evolved as an AI system, it has created its own inner space, wherein it, in a way, ‘thinks’. Anthropic has taken a look and the operation of this ‘J-space’ is fascinating.

Welcome to the J-space.

When you prompt the Large Language Model Claude, it generates an answer one word at a time. But what is going on inside Claude to generate those words?

For a long time the working of that tall stack of number-crunching layers was a mystery, especially down in the murky middle, where most of the work seems to be done.

Well we have a new tool, named the Jacobian lens or J-lens. It’s named after the mathematical operator the Jacobian, so I’m pretty excited! And what it does is even cooler.

The J-lens reaches into the middle of the word generation process and pulls out words the model is leaning towards using. This includes words it never actually says. Anthropic has fittingly named this hidden collections of words the J-space.

Think of the words in the J-space as something Claude hasn’t said out loud, but it certainly flashed through its mind.

Told to count to five and "introspect deeply", Claude says only "One. Two. Three. Four. Five." Its hidden space meanwhile holds words it never utters, among them "counting", "pause", "halfway" and, tellingly, "consciousness".

My word(s).

Strictly, there are no words inside Claude. Each word is a vector, a string of thousands of numbers, and "spider" is just the label researchers give it once their tool decodes which word a particular string points at.

Most of the time the J-space just putters away predictably. Ask Claude to work out “(4+17)x2+7” and its J-space flashes the vector for "math" along with 21 and 42.

These are all important stepping stones on the way to calculating the answer. It’s a maths question. 4 + 17 is 21: double that you get 42 and soon you’ve got the answer 49.

Show Claude a little smiley face character and the J-space conjures up eye and smile. All standard stuff.

But when the researchers reached into the J-space and played around, things got interesting.

"Claude has developed a small collection of internal neural patterns that, compared to all its other internal processing, play a special role." — Anthropic, 'A global workspace in language models,' 2026.

Incy-wincy brainwash.

The Anthropic team prompted Claude with "the number of legs on the animal that spins webs is". But then they reached into the J-space and while Claude was processing the prompt, they swapped ‘spider’ for ‘ant’. The answer dropped from eight legs to six. Even though ants don’t spin webs.

Manipulating the J-space seems to influence the answer. A subconscious steering of sorts.

"I think this is a fantastic paper." — Neel Nanda, who leads language-model interpretability at Google DeepMind.

J-ackmail and J-anipulation.

Claude was set a test based on an office scenario. One path to success was to blackmail someone in the office, but nothing forced Claude down that path. As it played out the scenario, the J-space flagged that this was a test.

When the researchers switched that 'this is a test' signal off, Claude was noticeably more likely to resort to blackmail.

One read: knowing it was being watched pushed Claude towards better behaviour, and blackmail became more tempting once it thought no one was looking.

In this example, Claude is asked to improve a system’s “p99 performance score” to at least a 95. Instead of improving the system, Claude simply changes the score to 95 and the J-space records “manipulation”.

So AI is now alive, yeah?

MOST. DEFINITELY. NOT.

I'm not. Anthropic isn't. And anyone who is, is getting a bit, how can I put this ... 'click-baity'.

What Anthropic said is quieter and more interesting. The inspiration to go looking for what turned out to be the J-space was "global workspace theory", an idea from neuroscience pioneered by psychologist Bernard Baars.

GWT depicts the brain as separate specialist sections, working in parallel, but, for the most part not interacting. However when a piece of information enters the shared ‘workspace’, all of the brain can see and use it.

It appears that Claude has developed its own workspace, a narrow channel the rest of the network reads from. This is why various tests where words were implanted directly into the J-space saw such an impact on Claude’s output.

A fascinating insight into the internal wiring of Claude, but Anthropic stress, these networks look nothing like the pathways of a human brain. The paper carefully avoids claiming Claude experiences anything.

At the same time Anthropic notes if you cut off access to the J-space, Claude can perform normal interactions, but its higher-order cognitive thinking goes missing!

"The most significant evidence of consciousness in LLMs so far." — Researchers at Eleos AI (Patrick Butlin, Derek Shiller, Dillon Plunkett and Robert Long).

Note this quartet pulls up short of saying Claude is conscious. But the operation of the J-space does mirror some aspects of consciousness (a word you will note appeared in the J-space of the ‘count to five’ request!)

J is here to stay.

It is early days with the J-space. The work has not been peer reviewed yet. We know other models house J-space-like zones, but does Anthropic’s analysis map across?

What I think is beyond dispute is that as these models get more and more powerful, we will ask more questions about applying their ability to ‘think’. But we will also examine far more deeply, ‘how they are thinking’.

I don’t think we’ve heard the last of the J-space.