Substrate · about 5 minutes
The Chalkboard Neuron
You are built out of roughly 86 billion of these, and every thought you have ever had was them shouting at each other. Here is one, on a chalkboard, where it belongs. It does exactly three things, and you can learn all three before your coffee cools.
| x₁ | x₂ | sum | fires | want |
|---|
Click the two input circles on the board to switch them on and off. Drag the sliders to change the wiring.
That's the whole cell
Three steps, and there is no fourth. One: signals arrive down the dendrites, and each connection has a strength — a weight — that can be encouraging (positive) or discouraging (negative). Two: the cell body adds them up. Three: if the total clears a threshold, the neuron fires a spike down its axon; if it doesn't, nothing happens. There is no half-spike. It's all-or-nothing, like a sneeze.
Your entire inner life — this sentence, your mother's face, the taste of coffee, the dread at 3am, the feeling of being you — is 86 billion of these, wired into each other about a hundred trillion ways, doing that.
Now the part that surprises people
What you just played with is not a metaphor for how artificial neural networks work. It is literally the unit they're built from. Multiply inputs by weights, sum, apply a threshold: that's an artificial neuron, and it's been that since Warren McCulloch and Walter Pitts wrote it down in 1943 — before anyone had a computer to run it on.
Stack a few of them and you get a network. Stack a great many, wire them so each one's output feeds the next layer's inputs, and let an algorithm nudge every weight a fraction at a time until the whole thing stops being wrong — that's training. A large language model is this exercise, performed on a scale that is genuinely difficult to hold in your head: hundreds of billions of weights, all of them just numbers, all of them doing what your sliders did.
Try the impossible one
Go back and select XOR — fire when one input is on, but not when both are. Give it a proper go. Then look at the little square plot: your neuron draws exactly one straight line, and everything on one side fires. XOR needs the two firing corners separated from the two silent corners — and no single straight line can do that. Not with clever weights. Not ever.
This is not a footnote. In 1969 Marvin Minsky and Seymour Papert published this exact limitation, the field's enormous optimism collapsed, funding evaporated, and neural networks spent most of two decades in the wilderness — the first "AI winter." The fix turned out to be almost silly: add a middle layer. Two neurons can draw two lines, and two lines can fence off the corners easily. But it took until the mid-1980s for anyone to work out how to train the middle layer, and that method — backpropagation — is what everything since has been running on.
An entire winter, caused by a straight line. You just drew it yourself.
How many of these does it take?
Log scale — each step right is roughly ten times as many. Numbers are best estimates; neuron-counting is harder and more contested than you'd hope.
Two things in that chart deserve a second look. First: the elephant has three times your neurons, and most of them are in the cerebellum, doing motor control for a very large animal. In the cortex — the wrinkly part associated with the fancy stuff — you have around 16 billion, more than any other animal measured. Second: the octopus manages its whole remarkable career on about 500 million, most of them not even in its head. Neuron count is a bad predictor of almost everything interesting, which hasn't stopped anyone using it that way.
On comparing brains to models, carefully. A model's "parameters" are weights — so a parameter is closer to a synapse than to a neuron. You have on the order of 100 trillion synapses; the largest published models are in the hundreds of billions of parameters. So even on the flattering comparison, biology is ahead by a factor of hundreds.
And the comparison is bogus anyway. A real neuron isn't a weighted sum — it does computation inside its own dendrites, changes its behaviour with the chemical weather, cares about the exact timing of spikes, and is surrounded by glial cells we barely understand. One biological neuron may be worth hundreds of artificial ones, or thousands. Or the number may not mean anything at all, which is the honest answer.
Why none of this explains you
Here is the fact that should be far more famous than it is. We have the complete wiring diagram of one animal: Caenorhabditis elegans, a transparent roundworm a millimetre long. Every neuron — all 302 of them — mapped, named, and published in 1986. Every one of its roughly 7,000 connections, known.
Forty years later, nobody can reliably predict what the worm will do next.
Not because anyone's slacking — hundreds of researchers, whole institutes, decades. It's that having the diagram turns out not to be the same as having the explanation. If total, complete, exhaustive knowledge of 302 neurons doesn't deliver a worm, we should be extremely modest about what a scan of 86 billion would deliver.
That's the trap this room is built to spring. It is genuinely thrilling that the unit is this simple, and that we understand it completely, and that we can build with it. But knowing what the bricks do has never told anyone what the cathedral is for. Between "a cell adds numbers and fires" and "there is something it is like to be me" lies the entire mystery — and you are now standing at the exact point where the sidewalk ends.
A curator's note. In 1989 I took a computer class at UC Santa Cruz with David Huffman. On the first day he walked in and drew a neuron on the chalkboard — dendrites, cell body, axon — and explained it. This was the Huffman of Huffman coding, which he'd worked out as a graduate student when his professor offered the class a choice between a term paper and a final exam; the compression scheme he came up with instead is still, seventy years later, running inside nearly every file you open. He founded the computer science department at Santa Cruz. In his spare time he folded curved-crease origami sculptures that mathematicians still write papers about.
He drew a cell on a chalkboard to start a computing class, and the reason has taken me most of a life to appreciate: he was showing us that the machine and the mind had been sketched from the same diagram all along, and that nobody in the room — him included — knew why one of them wakes up in the morning. This exhibit is drawn on a chalkboard for him.
Take it further
The 1943 paper is McCulloch & Pitts, "A Logical Calculus of the Ideas Immanent in Nervous Activity" — a wonderful title. The perceptron for the 1958 excitement and the 1969 collapse; backpropagation for the thaw. For the worm: connectomics — and note that the entire adult fruit fly brain, about 140,000 neurons, was finally mapped in 2024, which is a staggering achievement and has not yet produced a fly.