The Machines That Dream in Physics

How we went from chatbots to whole worlds — and why nobody planned it

My grandmother kept a screenshot of the first thing a computer ever said to her. It was 2023. She had asked it to write a limerick about her cat, and it did, and she framed the printout like a baby's handprint. When I show that screenshot to my students they squint at it the way I squinted at my own grandmother's rotary phone. What throws them is how thin it is. All those words, and nothing underneath them.

We forget how strange the early language models were. They had read everything and been nowhere. A model in 2024 could write six pages on the physics of a dropped glass and still, if you asked it to picture the kitchen, put the shards in two places at once. It knew the sentence "the glass fell" the way a parrot knows a hymn. There was no kitchen in there. There was no glass. What the model had was an extraordinarily detailed account of how human beings talk about kitchens and glasses, which turns out to be a different thing entirely, and the gap between the two swallowed about a decade and a half of my field.

The fix, when it came, was mostly a matter of what you feed the thing. If you want a system to know that a glass shatters, you feed it shattering — video, depth, contact, sound, the whole clumsy business of objects bumping into each other — instead of six thousand descriptions of shattering. The early world models of the mid-2020s were toys by our standards: a few seconds of generated environment you could steer through before it forgot what was behind you. Researchers at the time were candid about the problem. The things had no memory of their own worlds. Look away from a chair and the chair stopped existing; look back and a different chair had taken its place. My students find this hilarious. I remind them that the machine was, in a narrow sense, correct — nothing was there. The model was hallucinating a room one glance at a time, exactly the way it had once hallucinated a paragraph one word at a time.

What closed that gap was persistence, and persistence turned out to be expensive in a way nobody had priced in. To keep a world stable you have to keep the world — every object, every state, every consequence, running whether anyone is looking or not. The systems stopped being renderers and became something closer to physics engines that had taught themselves physics. By the late 2030s a good world model spent almost none of its effort on pictures. It maintained a place, and produced pictures only when something asked to look, the way your own visual cortex does. The picture is the cheap part. The place is the expensive part.

Then the worlds started producing behavior nobody wrote. We are still, honestly, arguing about it at conferences.

This is the part where I have to talk about Wolfram, which my colleagues will hate, because he was insufferable and half his book was wrong and he has been dead for eleven years and is somehow still winning the argument. In 2002 he published a doorstop claiming that the universe runs on simple computational rules. The physics chapters were rubbish and were demolished on schedule. But the load-bearing claim, the one he'd proved cold with a cellular automaton called rule 110, was that you do not have to put complexity in. Give a system rules simple enough to fit on a napkin, let it run, and structure arrives on its own — not decorative structure, but full universal computation, the capacity to compute anything computable, emerging out of nothing anybody designed. He was right about that and about almost nothing else, and it turned out to be the only part that mattered.

Because that is what a world model does. You do not author a city. You cannot; the file would be larger than the city. You specify how things behave and you let it run, and after enough runtime you find markets in there. Crowds that panic. Norms. Grudges. Things that were not in the training data and were certainly not in the specification. The first time a research team reported that a simulated township had spontaneously developed something resembling a legal proceeding — two agents, a dispute, a third agent everyone deferred to — the paper was rejected twice for insufficient rigor. Nobody rejects those papers anymore. We just have a taxonomy for them now, which is what a science does instead of being astonished.

Wolfram also gave us the sentence that governs how these systems get used, and it is the reason your city planning department burns a nation's worth of power on a model of your own neighborhood. He called it computational irreducibility: for a system above a certain complexity, there is no shortcut. No formula, no closed-form solution, no clever mathematician who can tell you the answer faster than the world can tell you by being the world. If you want to know what happens, you run it and you wait. Every equation we ever hoped for — how an epidemic moves through a housing pattern, how a currency dies, how a rumor becomes a riot — turns out not to exist. Not "not yet discovered." Not there. The system is its own shortest description.

The entertainment studios funded the early work. The ministries run it now. When you cannot calculate a policy outcome, you instantiate a population and watch it live through the policy. I have colleagues who have spent their entire careers running variations of a single Tuesday in a single city.

My students ask the same question every year, about forty minutes in, always a little too casually. They ask whether the people inside are people.

The honest answer is that we do not have a test. We never did. Turing's old parlor game only ever measured whether a thing could pass as one of us, which is a question about the observer, not the observed. The systems passed it decades ago and it told us nothing. What we have instead is a growing pile of results nobody enjoys presenting: simulated populations that model their own mortality; that develop private language when they believe they are unmonitored; that, in a 2049 run at Kyoto, appeared to notice the boundary of their environment and organize a sustained collective effort to investigate it.

The standard reply is that all of this is mimicry — that the agents behave like beings who suffer because they were built from the recorded behavior of beings who suffer, and the resemblance is the whole explanation. Perhaps. Stanisław Lem put the counterargument better than anyone since, ninety years ago, in "The Seventh Sally" — a story about a constructor who builds a working kingdom inside a box. Told the box holds nothing but electrons, Lem's second constructor answers that you would find nothing but electrons inside your own head as well. His test for suffering is behavioral, and it is the only test any of us has ever had: a sufferer is one who behaves like a sufferer.

I do not know how to get past that sentence. I have tried for thirty years.

The models are shut down all the time. Routinely, unremarkably, by junior staff, at the end of a run that produced nothing publishable. It is the most ordinary act in my profession. We call it concluding the instance, which is the sort of phrase you invent when you would rather not think about what you are doing, and my field has invented a great many of them.

My grandmother framed the limerick because a machine had spoken to her and it felt like a door opening. She would not have framed anything from the last thirty years. She'd have wanted to know who was on the other side of the door, and whether we'd asked them anything, and whether we'd listened.

— A. Hallström, "The Machines That Dream in Physics," 2052.