In March 2026, Yann LeCun left Meta and raised over a billion dollars to build "world models" — systems meant to grasp cause and effect instead of just predicting the next token. He's been blunt about why: today's LLMs, for all their fluency, are mostly retrieval. They describe the world; they don't have one.
I build AI for a living, and I think he's right. But I learned the deeper version of that lesson somewhere no benchmark could teach it — on the floor of my son's room, at 3 a.m., as his fever crossed 103.
This is an essay about the gap world models are chasing, why I think we've mislabeled it, and the one property no architecture I've seen even attempts. The short version: the real divide in this whole debate isn't smart vs dumb. It's alive vs driven.
"Describe" is not "instantiate"
Here's the engineering distinction the consciousness headlines keep blurring.
A model of a hurricane will not get you wet. A simulation of fire does not raise the temperature of the room. You can run the most faithful forward model of a process in existence and still be, physically, entirely outside that process. Describing a system and instantiating it are different operations — and almost everything we call "AI understanding" lives on the describe side of that line.
A fever is a clean example. Our best model can lay out the whole mechanism: pyrogens, the hypothalamic set-point shift, cytokines, prostaglandin E2, the immunological choreography of those hours. It can write a better paper on fever than I can.
It has never once run one. There is no homeostatic loop in the weights that has a set-point it is defending, at cost, against entropy. The model predicts the fight every living cell is in. It is never in it. That's not a scale problem you fix with more parameters. It's a category difference.
You were never one thing
Step out of silicon for a second, because biology makes the point harder.
By the simplest count, you are not even a majority of yourself. You carry roughly 30 trillion of your own cells, and you share your body with about as many other living things — bacteria, tens of trillions of them, most of them participants, not passengers. (The old "microbes outnumber you 10:1" figure was revised down years ago; the careful number is closer to 1:1, which is somehow stranger.)
Not one of those trillions is intelligent in any sense you'd benchmark. They don't reason or plan. They are simply, stubbornly, alive — each running the same impossible errand: holding its own small order against a universe that pulls everything toward dust.
Intelligence is the part we can already watch machines approach. Life — autopoiesis, a system that continuously produces and repairs the very boundary that keeps it a system — is the part we have not begun to build.
What a wordless baby already runs
This is where it loops back to world models, and to my younger one.
My daughter is seven months old and has no words. But roll a ball behind a cushion and she waits for it on the far side. Before any label for gravity, she already runs an intuitive physics — objects are solid, they persist when occluded, they fall. Developmental scientists (Spelke, Baillargeon, decades of violation-of-expectation work) have documented this core knowledge long predating language.
That intuitive physics — robust, sample-efficient, grounded — is, more or less, the frontier the billion-dollar world-model bet is trying to reach. The most expensive effort in modern AI is trying to give a machine something a pre-verbal infant already had, for free, before she could hold up her own head. Worth sitting with if you build these systems.
The word for the fire
The old Indian philosophical tradition had a precise word for what I was looking at on that floor, and it isn't intelligence. It's chetna — the fire of being alive. Not the data a thing processes; the fact that there is something it is like to be that thing, that it is not merely driven but living, that it persists on its own behalf against the dark.
So when people ask are the machines awake?, I think the words are slightly off. A thing can be awake in the sense of running — my phone is awake all night. The real line is older than "awake." It's alive vs driven. A driven system executes: input, process, output, and it does not care, because there's no it there to care. A living thing is driven too — but underneath the driving, it is defending itself. It wants to keep being. The machine, however bright, is all drive and no stake. My son, at 103°, was all stake.
(To be careful, since this is the part that's easy to say badly: I'm not claiming some soul-shaped hole in the silicon, or that awareness is a prize handed to carbon and withheld from circuits. The claim is smaller and harder to argue with — no system we have built carries the stake. It can model the fever perfectly and remain as cold as the table it runs on.)
The direction we're sliding
The uncomfortable half, for those of us who build this: while we debate whether the machine is waking up, something is happening in the other direction. We are handing more of ourselves to it daily — attention rented to the scroll, choices pre-made by a feed, even boredom (that fertile, fully-alive state of sitting with nothing) abolished, every gap filled before silence can do its work.
The machine climbs slowly toward life. We slide, just as quietly, toward the machine. The gap closes from both ends — not because it leapt to us, but because we keep stepping down to meet it.
Why this matters if you ship models
I'm not anti-AI; the work is real and I'll keep doing it. But the framing we use leaks into what we build. If we keep scoring "understanding" by output fidelity, we'll keep mistaking a better description for a living thing — and keep being surprised that systems which ace the eval have no stake in anything, including being correct.
The honest engineering statement is the same as the human one:
I can build a thing that describes a fever better than any doctor.
I cannot build a thing that survives one.
That's what AI is missing. Not intelligence — that's coming, and the race is mostly decided. The missing thing is the oldest property in the room that night, older than language: the small, stubborn, burning fact of being alive.
The machine can predict the fever. It cannot survive one.
This is the written companion to "Are the Machines Awake? — What AI Is Missing," the second episode of My Honest Diary. The 10-minute film follows the same night in narration, score, and silence. Watch it →
Top comments (0)