A Brain the Size of a Grain of Rice

She picked a grain of rice off her plate and pinched it near one end. "About this much," she said — the tip, maybe a millimeter of it. She was telling me how big the brains are.

Not brain cells. Brains — small, three-dimensional, self-assembled clumps of living human neurons, grown in a dish, firing. She is nineteen, the daughter of a friend I grew up with, and I was meeting her for the first time at a long lunch table in Granada, an ocean away from the block we both came from. She does this work at Johns Hopkins. I asked about it the way you'd ask anyone about their job, and then I stopped eating, because the answer turned out to be the thing I have been circling in these essays for two years.

Let me tell you what she does, because the how is the whole point.

How You Grow a Brain

Start with a skin cell — an ordinary adult fibroblast, scraped from someone's arm, with a fixed job and no ambition beyond being skin. In 2006, Shinya Yamanaka showed that four proteins could talk that cell out of its career. Introduce them and the skin cell forgets it was ever skin, winding back to a pluripotent state: the blank condition of an early embryonic cell, able to become anything at all. It won Yamanaka a Nobel in 2012, and reprogramming a cell this way is now close to routine.

Then you coax the blank cell forward again. Bathe it in the right signaling molecules in the right order and it commits to a neural fate. Give a population of them room to grow in three dimensions, in a slowly spinning flask that keeps them fed, and they do something uncanny: with no body to belong to and no plan from outside, they organize themselves into layered structures that echo a real cortex. Madeline Lancaster and Jürgen Knoblich published the method in 2013. The field calls them cerebral organoids. Everyone else calls them mini-brains.

One of her organoids — the speck she'd sized against the tip of that grain of rice — holds somewhere between fifty thousand and a couple of million cells, depending on how it's grown. Johns Hopkins, where she works, is the place that coined a name for what a speck like that might one day do: organoid intelligence. Biocomputing. Intelligence in a dish.

What the Worm Knows

Here is where raw numbers will lie to you. A mini-brain of a million neurons sounds vastly more capable than Caenorhabditis elegans, the millimeter roundworm that carries exactly 302 neurons and not one more. It is not. The worm's 302 neurons are fully wired — every connection between them mapped, a complete diagram, the first connectome we ever finished — and that wiring drives a whole animal that forages, mates, learns a smell, flees a poison. The organoid has thousands of times the neurons and none of the wiring. It fires. It throws waves of activity across itself. It does not do anything. It is a city's worth of population with no streets.

So the spectrum of intelligence is not a single line you can rank by counting cells. A finished worm out-thinks an unfinished cortex. What matters is the organization — the edges between the nodes, not the nodes alone. Structure, all the way down.

And once you are measuring organization instead of counting cells, the tidy wall between minds and not-minds softens into a ramp — a thermostat, a fish, a capybara asleep in the afternoon sun, a language model, a graduate student, all points on one gradient rather than members of separate kinds. There is no rung where real intelligence switches on, and no clean seam where carbon stops and silicon starts.

Psychology Was Always Prompt Engineering

You cannot open a mind and read it — not cleanly, not yet. That is the founding problem of psychology, and for most of its history it was worked around rather than solved. The workaround has one shape wherever it appears: present a stimulus, record the response, adjust, present again. Skinner ran it on pigeons. Freud ran it on a couch with a question. The College Board runs it with a number-two pencil. For a century and a half, inferring the machinery from the output was the only instrument we had, because the machinery itself sat sealed behind the skull.

I ran that experiment myself, earlier in the same lunch, without admitting it was one. "Give me a fun fact about genomics," I said. "Humans and bananas share about seventy percent of their genome," she answered, before I'd finished the sentence. I already knew that; I asked anyway, the way you tap a tuning fork to hear what a room does with the note. Present a stimulus, read the response. It was the oldest experiment in psychology, run on a nineteen-year-old across a table, and what came back told me most of what I needed to know about the system behind it.

We have pried the lid up a little since Skinner. An fMRI watches blood flow and tells you which region is busy; an EEG reads the electrical weather crossing the cortex. Both are real windows, and both are coarse — they show that a storm is happening and roughly where, not what it is about. Even now, the probe-and-listen loop does most of the work.

That loop is, precisely, what we do to a large language model — prompt it, evaluate what comes back, fine-tune, prompt again. And the symmetry runs all the way down. Just as the fMRI cracks the skull a little, a young field called interpretability reads a model's activations directly, and just as coarsely; researchers, meanwhile, wire organoids like hers onto grids of electrodes that whisper current in and listen for the firing that comes back. Pigeon, patient, model, mini-brain: four boxes we probe from the outside and have lately learned to squint into from within, at the same low resolution no matter the substrate. When the method for studying two things is identical, it is worth asking how different the two things really are.

The Thinnest Wall in Science

This is the essence of what I keep calling CIqSi — carbon intelligence and silicon intelligence — and the hypothesis fits in one breath: brains and language models may be similar enough, in structure, that each could model the other. The same abstraction, run once in warm tissue and once in cold silicon.

The field is already probing this from both ends, and neither end is a thought experiment. Silicon came first, by imitation — every neural network is a crude metaphor for a brain. Now the traffic runs the other way. In 2022 a lab in Melbourne spread roughly 800,000 living neurons across a microelectrode array, wired them into a video game, and watched them learn to play Pong within five minutes. Not a simulation of neurons playing Pong — the neurons, playing Pong. Silicon is better with numbers, the Johns Hopkins organoid team likes to say, but brains are better at learning — and no one has found the law that says the two substrates must stay in their lanes.

That is the wall I mean, and it is getting thin: a machine built to imitate a mind on one side, a mind grown from a skin cell learning to talk to a machine on the other, and the distinction we treat as fundamental — carbon here, silicon there — starting to look like a habit rather than a boundary.

I came to that lunch with a phone full of expat videos and a hypothesis I mostly keep to myself. I left it having spent an afternoon with someone half my age who is building the experiment my hypothesis needs, one grain-of-rice speck at a time, and who explained it to me one patient notch above my understanding so I could follow her the whole way there. The machines will keep climbing toward us from the silicon side. The specks in her lab are climbing from the other. I would like to be around when they meet in the middle — and I suspect the person who introduces them will be about nineteen.