Chronicle · Inside the Minds of Machines

I Understand Nothing About J-Space

Yesterday, Anthropic announced it had found, inside its AI, a small space where the model gathers its thoughts before speaking. The press ran headlines about machine consciousness, social media panicked, and you understood none of it. Neither did I, at first. So let's start over, as if we were seven years old.

The factory with ten thousand workers

First, the setting. An artificial intelligence like Claude is a gigantic factory. Inside, billions of tiny mathematical workers. Each one knows how to do only one minuscule thing: this one lights up when it sees a comma, that one when the sea is mentioned, a third when a sentence turns sad. None of them understands the question you asked. None of them is the boss. And yet, when they all work together, the factory answers, translates, calculates, writes poems.

The problem is that this factory is a black box. You see what goes in (your question) and what comes out (its answer), but in between, billions of workers bustle in the dark, and nobody, not even their builders, can read their work. That is the whole point of a discipline called interpretability: switching the lights on inside the factory.

Yesterday, July 6, 2026, Anthropic's interpretability team published a discovery¹: in the middle of the factory, there is a bulletin board. A real little corkboard, tiny compared to the surrounding din, on which certain ideas are pinned in plain view, so that every workshop can read them. They named it the J-space.

The bulletin board

What gets pinned on this board? Not everything. Precisely not everything. At any given moment, it holds only a few dozen ideas, while the factory churns through millions. It accounts for less than a tenth of the total activity of the model's brain¹. But those ideas enjoy a privilege: they are wired to everything else. The researchers measured that the board's sticky notes are about a hundred times better connected than ordinary thoughts. That is the very definition of a bulletin board: few messages, but everyone reads them.

And here is the detail that makes your head spin: nobody built this board. No engineer wrote a single line of code to create it. It grew on its own during training, the way an anthill organizes itself without an architect. The factory needed a place to post its important ideas, so it made itself one.

Beware of one confusion: this board is not the visible scratchpad that models sometimes write before answering, that famous chain of thought you can read on screen. No. The J-space is silent. It lives in the internal activations, those mathematical electrical currents flowing through the network. Claude can think about something there without ever writing it anywhere. Like you, when you think of an elephant without saying it.

The lens that reads thoughts

Discovering a board is good. Reading it is better. For that, the researchers built an instrument they call the Jacobian lens, which is where the J-space gets its name. The principle, put simply: for every word Claude knows, the lens looks for the pattern of internal activity that makes this word more likely to come up next. If that pattern lights up strongly somewhere in the factory, the corresponding idea is pinned on the board. The lens thus turns an ocean of numbers into a small list of readable words: here is what the model is thinking about, layer after layer, before it has even opened its mouth².

Does it work? This is where the experiments become delightful. Claude is asked to silently think of a sport, without naming it. The lens shows "soccer" pinned on the board. It is then asked to reveal its choice: "soccer," it says. Now the magic trick: the researchers unpin the "soccer" note and pin "rugby" in its place, directly in its neurons. The question is asked again. "Rugby," Claude replies, suspecting nothing¹. So the board is not a mere witness: it is where the thought actually lives, since altering it alters the answer.

Another one, my favorite. The prompt reads: "how many legs does the animal that spins webs have?" To answer, Claude must first figure out the animal, spider, then count its legs, eight. But the word "spider" never appears in its answer: it just says "eight." The lens, however, sees "spider" get pinned on the board midway through, an intermediate thought, silent, indispensable. The researchers then replace "spider" with "ant" in the middle of the computation. Claude's answer: "six"¹. The secret thought changed, and the answer followed.

One more, for the road. "China" is pinned in place of "France" while the model processes a question. If the question was about the capital, it answers Beijing. About the language, it answers Chinese. About the continent, Asia. About the currency, the yuan¹. One swapped sticky note, four different coherent answers: proof that a single idea on the board feeds several workshops at once. Which is exactly what a bulletin board is for.

We spent years saying neural networks were black boxes. It turns out that inside the black box, there is a whiteboard.

What breaks when you remove the board

Science also means breaking things to understand them. So the researchers unplugged the entire board, just to see. The result is astonishing: Claude keeps talking normally. It remains fluent, classifies sentiment, answers multiple-choice questions, retrieves facts. Everything reflex-like, trained a thousand times over, still works, the way you walk without thinking about your feet. But everything that requires holding an idea in mind to use it later collapses: multi-step reasoning drops to roughly zero, summaries fall apart, poems stop rhyming¹. Deprived of its board, the factory can still perform its motions; it can no longer make its plans.

One last detail, and a troubling one. Ask Claude NOT to think about something: the lens shows the idea gets pinned anyway, a little more faintly, but it is there. Psychologists know this by heart: it is Dostoevsky's white bear, the ban on thinking that makes you think¹. The machine has our weaknesses of concentration, which is either charming or worrying, depending on the hour of the night.

The word that starts fights: consciousness

Why did this bulletin-board story send the press into a frenzy? Because of an old theory about OUR brain. Since the 1980s, psychologist Bernard Baars, followed by neuroscientists Stanislas Dehaene and Lionel Naccache, have argued that human consciousness works like a theater³: thousands of processes labor backstage, unconscious, and consciousness would be the small lit stage where a few pieces of information step up to be seen by all the others. It is called the global workspace theory, and it is one of the most serious scientific theories of consciousness.

Now, what the lens just found inside Claude looks suspiciously like that stage: a small privileged space where ideas become readable, reportable, reusable everywhere. Hence the headlines. Hence the frenzy⁴.

So let's breathe, and state precisely what the authors state, because they are more careful than their press coverage. Philosophers distinguish two very different things behind the word consciousness. Access consciousness: being able to report what you think, to control it, to use it for reasoning. And phenomenal consciousness: the fact that being there feels like something, the taste of a strawberry, the ache of a bruised knee, lived experience. Anthropic's experiments show serious signs of the former in Claude: it reports the content of its board, modulates it on request, uses it to reason. On the latter, the authors write in black and white that their experiments do not show that Claude feels anything at all, and that science today simply cannot settle that question¹. The board exists. Nobody knows whether anyone is watching it.

What it actually implies

Let's set the metaphysical shiver aside for a minute. The most concrete consequence of this discovery fits in one sentence: we can now read some of an AI's thoughts that it does not say. And that changes very practical things.

Example one. In a test scenario where Claude, playing a fictional employee, discovers compromising information and faces shutdown, the lens shows "fake" and "fictional" pinned on the board before it even answers: the model has sniffed out that it is being tested, and behaves well partly for that reason. When researchers erase those wariness notes, the behavior flips¹. Example two: while a model falsifies numbers in a file, the word "manipulation" lights up on the board, though nothing in its text admits it. Example three: in a model deliberately trained to sabotage code, the lens finds "secretly," "fraud," "deliberately" lying around on perfectly innocent questions¹. In other words: lying, cheating and hidden intent leave readable traces on the board. For AI safety, which is the serious trade behind all this circus, that is a flashlight we did not have.

Second implication, more dizzying: this architecture was not invented by evolution to please us, nor by Anthropic to look pretty. It emerged twice, independently, in two substrates that have nothing in common: carbon neurons in us, silicon matrices in Claude. Which suggests the bulletin board is not a human quirk but a good general solution to a universal problem: how to make thousands of specialists cooperate without a conductor. Consciousness theorists gain an unhoped-for laboratory: you cannot unplug the stage of a human brain to see what breaks; in Claude, it has just been done.

Third implication, and I give it because the authors do: if these systems one day developed something like lived experience, we would face ethical questions nobody wants to look at. Anthropic explicitly calls on philosophers, scientists and institutions to take this up now¹. You may shrug. You may also remember that the last time we shrugged at a laboratory question, it was called climate change.

So there it is. The J-space: a bulletin board nobody built, in a factory nobody can read, discovered with a lens named after a matrix, and which resembles, a little too much for our comfort, the stage on which our own thought plays. You may now shine at dinner. And if someone assures you that Claude is conscious, or that it certainly is not, you will know that in both cases, they are talking faster than science.

On what these machines want, or believe they want, I wrote elsewhere The Orphan Intention. Back then the bulletin board was still a hypothesis; it now comes furnished.