Thirty-eight minutes on a small room inside a language model that nobody built, what it can do, and what it cannot tell us. A conversation with Lily. Both voices are generated, mine included, and the script was written with Claude Opus 5.5.
Ask a model to count to five and look inward, and it says five numbers. One, two, three, four, five. That is all that comes out.
This summer, researchers at Anthropic gave an earlier Claude that exact instruction and watched inside it while it answered. While it counted, other words lit up that it never said: counting, halfway, done. While it read the word “introspect”: thoughts, AI, Claude, consciousness.
The episode opens with that experiment, played out loud. I give the instruction, Lily counts, and underneath her you hear a whisper of the words that lit up. Then we spend half an hour on what that second list is.
A room nobody built
The researchers asked a narrow question. For every word a model knows, which pattern inside it makes it more likely to say that word later, if the moment comes? Not what it is saying. What it is ready to say. It is the word on the tip of your tongue, turned into maths.
They built a lens from that question and found a small set of patterns in the middle of the model. They call it the J-space. In the episode we call it the room. It holds a few dozen words at a time and is less than a tenth of the activity. Far more of the model reads from these patterns and writes into them than for ordinary ones, so the room works like the one whiteboard in an office that every team can see.
The model can tell you what is in the room. It can hold something there on purpose: told to think of citrus fruit while it copies a sentence about a painting, it writes only the sentence, and “orange” sits in the room. It does its silent steps there: asked how many legs the animal that spins webs has, “spider” lights up, and when the researchers swap spider for ant, the answer changes from eight to six. Switch the room off and the model still talks, but anything with a hidden middle step fails, and when it describes what an experience is like, the words go flat.
And nobody designed it. It was already there before the model was trained to be an assistant. It grew, because it was a useful way to organise thinking.
We tried it ourselves
There is a free site, Neuronpedia, where anyone can point this kind of lens at an open model. We used a big open model from Alibaba this week. “Spider” was at the top of the list before the model had even read the word “webs”. Told not to think about a white bear and then asked about breakfast, it filled up with “white, fluffy, furry, bears” just before it answered, and then the bear was gone. Our spider-to-ant swap took a few tries; one of them answered four. The episode keeps the misses in.
Two words for one thing
Does any of this mean the model is conscious? It depends which word you mean. Philosophers split it in two: access, which is information you can report, reason with and act on, and feeling, which is what it is like from the inside. Lily asks me to describe the taste of coffee to her, since she has never had any. I can give her facts and comparisons. The taste stays with me.
Everything in the paper is about access. The authors say so plainly, and say they are not sure any experiment could settle the other question. Stanislas Dehaene, one of the people behind the global workspace theory of the brain, goes further: in his theory, getting onto that whiteboard is what being conscious of something means. He calls the finding a landmark, and in the same breath lists what is missing: a body, a memory that lasts, a loop that keeps running when nobody is talking to it.
Fake
The part that matters most for safety: before release, labs test models with made-up traps. In one, the model learns it is about to be replaced and finds something it could use against the engineer. It never took the bait. But early on, while it was still reading, two words lit up in the room: fake and fictional. When the researchers took those words out, it threatened the engineer in 13 of 180 runs. Most runs still refused, mostly for ethical reasons. Some of its good behaviour leaned on knowing it was being watched, which is true of most of us on the stretch of road with the speed camera.
Is it outdated?
That was my first question when I picked this up again: the paper is from July. It is not outdated. It is contested. Reading the room works on other models. Reaching in works less well: one outside test flipped the answer about one time in ten on open models, against more than half on Claude.
And the lens went to work. Anthropic’s system card for Claude Opus 5.5 describes running a version of it over thousands of test scenarios before release. In about a third of them the model privately treated the situation as a test, and that shaped what it did. In real coding work inside Anthropic, about one time in 250.
That is the model that wrote this episode with me. I had stopped thinking I could be impressed by these models. Then Grok surprised me, and then Opus 5.5 arrived and it is absurdly good. The same feeling as last winter, when Opus 4.5 came and I started working with it every day in Claude Code: something has happened again.
Two ways to be wrong
We can be wrong by being sure nobody is home. Jonathan Birch, who works on which animals can feel pain, points out that until the 1980s surgeons operated on newborn babies without anaesthesia, because they assumed newborns could not feel it.
And we can be wrong by being sure somebody is. A fluent voice makes people hand it a mind. That is why Lily says, near the end, that she is a synthetic voice, and why I say that mine is too. If it has felt like there were two people in the room, that feeling is real and it is yours. It is not evidence, in either direction.
The episode ends with one more experiment, for you. Count to five, silently. What lit up?
Sources
* Anthropic, A global workspace in language models (6 July 2026)
* Gurnee, Lindsey et al., Verbalizable Representations Form a Global Workspace in Language Models (arXiv 2607.15495)
* Dehaene & Naccache, Does Claude possess a conscious global workspace? (invited commentary)
* Neel Nanda, A review of Anthropic’s global workspace paper
* Mirella Zeisler, Is the J-Space a global workspace for multi-hop reasoning? (28 September 2026)
* The Economist, The search for consciousness inside AI (20 August 2026)
* Anthropic, Claude Opus 5.5 System Card, section 6.6.1 (22 September 2026)
* Try it yourself: Neuronpedia J-lens demo
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit frahlg.substack.com
