TL;DR

I've read how computer chips work several times. Transistors, gates, clock cycles, the whole stack from silicon to instruction set. I knew the theory, and I didn't understand it. What finally worked was lying in bed, before falling asleep, and building a picture of it: current moving through gates, the gates opening and closing in rhythm, the layers stacking up. Nobody can actually see what happens at nanometre scale, so the picture wasn't something I'd observed. I made it, and once I had it the theory made sense.

Physics and maths work the same way for me. I'm not a maths person, things don't always come easy, and there's a huge difference between reading a textbook and watching an animation of a spline or a Bezier curve being drawn, with the control points pulling the line around. I spent a while just watching those. The understanding arrived through the eyes, and the words came afterwards to label it.

Language isn't always involved. Sometimes it's pure images, sometimes images mixed with verbal thoughts. Dreams, as far as I can tell, are mostly pictures. David Krakauer makes the point that deep insights, the mathematical breakthroughs people write memoirs about, usually start in the unconscious and only get formalised into language afterwards. Which makes sense, given that problem solving was around long before language was.

In an earlier article I wrote that we learn through the senses first and words second, and that language models did it the other way round. I also said there was serious work on AI that doesn't run on language and that I wasn't going there yet. In this article I go there.

Several different machines

The first thing I learned when I started reading about this is that people differ far more than we assume. I'd always taken for granted that everyone pictures things the way I do. They don't.

A study published this year in Cortex asked close to 2,000 adults in the UK about their imagery across eight senses, not just vision. About 14% had reduced or absent voluntary imagery in at least one sense (aphantasia), and about 21% had unusually vivid imagery in at least one (hyperphantasia). The earlier estimates, which only asked about vision, put aphantasia at 3-4%. Taste and smell were the senses most often missing. The imagery people most commonly reported as vivid was sound, and specifically melodies. Some people can hear music in their head the way I can see a picture. (It's self-report, so treat the exact numbers loosely.)

So "how humans think" turns out to be several different machines that happen to produce similar output. Some people solve with pictures, some with inner speech, some with sound or with movement and feel. When we ask whether AI thinks like us, the honest follow-up is: like which of us?

One place the visual machine shows up at work is data. We build charts because the eyes find patterns the numbers hide. I covered that in the earlier article, so one line is enough here: half of data science is admitting that the human visual system is the best pattern detector in the building.

The Swiss army knife

Then where does language fit? Chomsky's position, roughly, is that language is primarily a system for creating and interpreting thought, and only secondarily a tool for talking to others. We're born with the faculty for it. I can see how that could be true, and I don't think it's the whole truth.

There's a strong counterpoint from neuroscience. Fedorenko, Piantadosi and Gibson argued in Nature in 2024 that the brain's language network is separate from the networks that do reasoning, and that people with severe aphasia, who have lost language almost completely, can still do arithmetic, logic and work out what other people are thinking. Their conclusion: language is primarily a tool for communication, and thought gets along without it.

My own position is somewhere in between, and it's about what innate means. We're born with the capacity to learn language, not with language. Babies hear voices from the womb during the last months of pregnancy, and newborns cry with the melody of the language they heard there: French babies with a rising contour, German babies with a falling one, in the first days of life. So the preparation starts early. But the order is still pictures first, then sounds attached to pictures, then meaning, then words.

The way I think of it: language is a Swiss army knife for intelligence. You can do almost anything with it. Explain a chip, a contract, a feeling, a joke, as long as the other side shares the meaning of the words. It's a tool for conveying. It's a weaker tool for discovering. The discovering, in my experience and apparently in the research, happens somewhere else, and language is what you reach for afterwards to carry it across.

Symbols

A dog knows what the leash means. A vervet monkey has different alarm calls for a leopard and an eagle, and the others look in the right direction. Nobody labelled anything. The symbol got its meaning from experience: this sound, then that thing happened.

AI didn't start there. For most of its history, a model learned what a cat was because a human had typed cat next to a few million pictures. ImageNet, the dataset behind the 2012 breakthrough, was built by people tagging images by hand. The problem has a name, the symbol grounding problem, which Stevan Harnad described in 1990: how does a symbol in a machine come to mean anything, rather than just pointing at other symbols?

That's moved more than I expected. Self-supervised learning got rid of most of the hand labelling. CLIP, the model behind a lot of today's image tools, learned from pictures and the captions people had already written for them on the web, so the labels were still human, just free. And a paper at ICML this year (one of the big machine learning conferences) found that grounding emerges inside language models during ordinary training, with no explicit grounding objective, concentrated in the middle layers, where attention heads gather information about the environment to predict the next word.

So the models got further than I expected: they connect symbols to meaning without anyone labelling the data. The catch is where the meaning comes from. It’s grounded in human language, not in the world. The model knows “leash” is near “walk” because we wrote it down a billion times. The dog knows because it went for the walk.

Can a model think in pictures?

In the earlier article I mentioned that vision models chop an image into patches and treat each patch as a token, so inside the model, seeing and reading have more or less merged. What hasn't changed is that the reasoning still runs through text. The model looks, then it talks to itself in words.

There's research trying to change that. Mirage, a 2025 paper that made it to CVPR this year (the main computer vision conference), lets a model reason with latent visual tokens, internal images it never draws, mixed into its chain of text. The authors call it machine mental imagery, which is a fair description of what I do in bed. A follow-up, Mull-Tokens, goes further and lets the thinking tokens be neither text nor image. And the bigger bet, LeCun's world models, is to learn how the world behaves from video before touching language at all. Senses first, words second, done on purpose this time.

The one place where AI has clearly learned something without our language is games. AlphaGo's move 37 against Lee Sedol in 2016 wasn't in any human game record. The system played itself millions of times and found it. Geoffrey Hinton's point is that this is the real advantage of digital intelligence: it can generate its own experience, at a scale and speed no brain can, and whatever it learns can be copied to a thousand other instances in seconds. Our learning dies with us. Theirs doesn't. That's where the title of this piece comes from, and I'll come back to it.

It's worth keeping Krakauer's objection next to that. Move 37 took thousands of GPUs and megawatts of electricity. A brain runs on 20 watts and finds its insights in bed. He'd call the first capability and the second intelligence, and I think he has a point, even if the gap is closing.

Where it came from

Part of what we know never arrived as information. It came from experience, from dreams, from things absorbed through the senses without noticing. We call it gut feeling, sixth sense, intuition. We don't treat it as knowledge we can hand to someone else, because we can't say how it got there in the first place. It's part of us, and we can't trace it.

In AI everything starts as data we're aware of. Databases, books, the internet, labelled or not. It's digital, traceable, and we call it training data. You can point at every byte that went in, copy it, audit it, remove it.

The traceability stops at the door, though. I wrote earlier about confabulation: models don't store strings, they reconstruct from weights, the same way we rebuild memories rather than replay them. So inside the model it's as untraceable as inside us. The difference is upstream. We can't point at what went into a person. And, until a few days ago, I'd have said there was no unconscious side either: no dreams, no absorbing the room while thinking about something else. Everything a model did was driven by a goal or a prompt.

Then on 27 October Google Research published something they call the sleep framework. The model alternates between a wake phase, where it takes input and stores it in a fast short-term memory, and a sleep phase, where it stops taking input and consolidates. New knowledge goes into pre-allocated capacity so the old weights don't get overwritten; they map that onto deep sleep. The temporary parameters get reset afterwards, which they call synaptic pruning. And there's a stage they call dreaming, mapped onto REM sleep, where the model generates its own samples, keeps the useful ones and trains on them. The results they report are good, with the usual caveat that they're the authors' own: near-zero forgetting when learning two languages it had never seen (Manchu and Kalamang), near-perfect recall at 10 million tokens, better factual retention than fine-tuning.

This is the first serious attempt I've seen at a complaint I made in the earlier article: that the deployed model doesn't learn anything while it thinks, and that our long-term store gets updated every night while we sleep and theirs never does. They even called it dreaming, which tells you where the field is looking for the missing piece.

Three caveats. The results are Google's own and nobody has reproduced them yet. The authors say themselves that sleep and dreaming are borrowed metaphors, and that the whole thing runs on "a fixed model architecture and human-defined training rules". And it still starts from digital data you can point at. The dreams are synthetic samples, which is a long way from a day of being in the world. It's closer to us than last year, and that's about it.

If you’re a tech-savvy person, the comparison that will come to mind is defragmenting a hard disk: both run offline and move things to where they belong. The difference is that defrag doesn’t change a single byte, the same files come out, just placed better. Sleep changes the content. It throws most of the day away, keeps what’s useful, and the dreaming stage adds things that were never there.

At what point is there knowledge, or intelligence, that’s genuine? A model comes up with an idea and a user is impressed. One user, impressed, by something generated from millennia of archived human knowledge. But we do the same thing: we gather from many sources and call the result our own idea. So that can’t be what makes ours genuine. What’s different is that we can’t trace where our knowledge came from, and a good part of it never went through our conscious mind at all. Whether that’s enough to call it genuine, I don’t know. What I do know is that the models don’t have the unconscious part yet.

What I'd say now

In 2019, on a panel at Microsoft's UK office in Reading, I said that computers are linear, and that empathy, imagination, creativity and seeing the bigger picture are ours. Hand them the repetitive work, keep the interesting problems. The first article in this series already retracted half of that: the machine turned out to be the fuzzy, creative part, and we had to build the linear part around it ourselves. Here's the rest of the scorecard, as of writing (autumn 2026).

Creativity: wrong. The model is the System 1 part. It recombines everything ever written, which is where most of our creativity comes from too. There’s a saying in Serbian, probably shared across the Balkans: “zajebi to što ne liči na ništa”, roughly “forget the thing that doesn’t resemble anything”. It’s not a compliment. If something resembles nothing you’ve seen before, it’s usually ugly. We’ve always known that good new things are made of old ones.

Imagination: half wrong. Models can produce, and they're getting latent images of their own. What they don't have yet is the wordless kind. The picture you build in bed of something nobody can see, the chip at nanometre scale, the curve pulled by its control points. That's the part this whole article has been about.

Empathy: the hard one. Empathy isn't only feeling bad for someone or understanding how they feel; it changes our decisions. Models can reproduce the part other people read, and the persuasion studies I wrote about last time show they're good at it. Whether it's genuine we can't tell. The easy answer is "it can't be", but that only holds if you still think of AI as a complex autocorrect, and I don't think that's true anymore. I'd rather say we can't tell than pretend to know.

The bigger picture: still ours, for a different reason than I thought in 2019. Back then I assumed the machine couldn’t see wide enough. It can, it sees more than any of us. What it doesn’t have is a reason to look. Seeing the bigger picture means knowing what you want from it, and wanting, as I argued last time, is something we put into models by hand. It doesn’t come with the data.

So the 2026 version of what's ours is shorter: the things we know without words, the questions nobody has asked yet, and deciding what the answer is for. Krakauer says the human job is moving from calculating to asking, and I think that's right.

And there's a shift in what the list is for. In 2019 it was about abilities, what we can do that they can't. In 2026 it's about ownership, what we shouldn't hand over even if they could. The last article asked that about skills, like finding your way or writing an email. This one asks it about deciding what the skills are for.

Hinton's framing closes it for me. Their weights copy, save and transfer in seconds. Ours don't. Everything a person learns in a lifetime dies with them, and it's also the only part that's genuinely theirs. The part that doesn't copy is the part that's ours. I find that mildly uncomfortable, and I'm OK with it.

Steering

Any prediction I make now has the same odds as the one from 2019. What was true then isn't; what's true now will change. I can reason and predict as much as I like, but that's not the point. We shouldn't be predicting a future we're still in control of. We should be steering it.

Two questions, one easy, one easier. Do we want a powerful intelligence that helps push the scientific frontier? I think most of us would say yes. Do we want a system that works against the well-being of our civilisation? A hard no. Everything else is in between, and that's where the work is.

If you think in pictures, or in sound, or in something I haven't thought of, I'd like to hear how it works for you.

Further reading