A Thousand Switches to a Neuron

For years the picture in everyone’s head was a tidy correspondence: an artificial neuron is a stripped-down model of a real one. The real neuron collects inputs, sums them, and fires if the total crosses a threshold. The artificial neuron collects inputs, sums them with weights, and passes the total through a function. Same cartoon, one in wetware, one in math. It made the brain feel like a biological version of the networks we build, just larger and slower and harder to inspect.

Then, in 2021, David Beniaguev, Idan Segev, and Michael London measured how good that cartoon actually is, and the answer was: not good at all. They took a detailed biophysical model of a single cortical pyramidal neuron, the workhorse cell of the thinking parts of the brain, and asked what size of artificial network you’d need to reproduce its input-output behavior faithfully, down to millisecond timing. Not a whole brain. One cell. The answer was a deep neural network of five to eight layers, which worked out in their setup to somewhere around a thousand artificial neurons to stand in for one biological one.

A thousand to one. The unit we’d been treating as the atom of the brain turns out to be a small deep network in its own right.

Where does all that hidden complexity come from? Mostly the dendrites, the branching tree of fibers that carries inputs into the cell. In the cartoon, dendrites are passive wires that dump signals onto the cell body to be summed. In reality they’re active. Different branches perform their own local computations, and a particular receptor, the NMDA receptor, lets a branch respond nonlinearly, in a way that depends on timing and location, not just on the raw total of what arrived. Tellingly, when the researchers stripped the NMDA receptors out of the model, the required artificial network collapsed from eight layers to one. Almost all the depth, almost all the thousandfold gap, lived in that one mechanism. The neuron is deep because its dendrites are doing pattern recognition before the cell body ever “decides” anything.

You can read this two ways, and the first way is a wet blanket on everything I’ve been arguing. If one biological neuron is worth a thousand artificial ones, then comparisons like “this model has as many parameters as the brain has synapses” are off by orders of magnitude, and the brain is vastly, humiliatingly more computer than our biggest networks. On that reading, biology isn’t encouragement for the inevitability thesis. It’s a measure of how far we still are.

But that’s not the reading I land on, and here’s the turn. The thousand-to-one gap isn’t a wall. It’s a receipt. It’s proof that architectures exist which pack the work of a thousand crude units into a single cell, and, more to the point, that such architectures are findable, because a process with no foresight at all already found them. Evolution had no blueprint and no goal. It had variation, selection, and deep time, and out of that blind loop came a component so computationally dense that our best deliberate engineering needs a thousand parts to match one of them. The efficient design was reachable by pure undirected search. That’s the fourth pillar: not that our networks are close to the brain, but that the brain is standing proof of how much better networks can get, and that the improvement is available to a process that isn’t even trying.

That’s the optimistic frame. Now let me be fair to the pessimistic one, because it has real force.

The first problem is that the thousand-to-one number is a measurement of a specific model, not a law. It’s how many artificial neurons it took to fit that biophysical simulation to that accuracy at that time resolution, using the network architectures the researchers happened to try. Push the required accuracy and the number climbs. Relax it and the number falls. Simplify the neuron model you’re imitating and it falls further. The ratio isn’t a fundamental constant of biology. It’s a number with a lot of dials behind it, and quoting it as though it were the exchange rate between carbon and silicon flatters a single careful experiment into a universal law it never claimed to be.

The deeper problem is that raw computational density might be beside the point. The whole spirit of this series is that intelligence is emergent, that it lives in the interactions of many simple units, not in the sophistication of any one of them. If that’s true, then discovering that neurons are individually complicated cuts against the clean emergence story rather than for it. Maybe the brain is powerful not because it wires up simple parts but because its parts aren’t simple, in which case “just add scale” is the wrong lesson and the magic is partly down in the cell, in the biochemistry, in machinery we can’t get by stacking layers. The neuron-as-deep-network result can be spun as support for artificial networks, but read honestly it’s at least as much a warning that we’ve been underestimating the substrate.

So here’s the residue I’ll actually stand behind. Evolution, with no designer and no plan, produced computational elements far denser than anything we build, which proves that highly efficient architectures exist and can be reached by blind search. That much is real, and it’s genuine reason to think there’s enormous headroom above where our systems sit today. What I won’t claim is that this headroom is close, or that scaling our current crude units is the road to it, because the same result that shows how far a blind process can go also shows how much of the work might be hiding inside the parts, exactly where a theory built on simple parts doesn’t want it to be.

Which is the tension I can’t dissolve, and won’t pretend to. Biology proves the ceiling is high. It also hints the ceiling is high for reasons that live in the cell, not just in the wiring, and if that’s where the intelligence partly lives, then the most important thing the brain has to teach us is the one thing our architectures were built to ignore.