Reading the brain like a network

For ten years the people who build artificial intelligence told us they could not understand what they had built. In 2026 one of them said the opposite: the networks were readable all along, and the field had been looking at them wrong. I think that essay is the most important thing written this year for anyone who wants to understand a brain. Here is why, what we did to test it, and what it changes.

The mess that wasn't

An artificial neural network is made of millions of simple units, called neurons after the real thing. Each one adds up signals from the units before it and passes a number on. Train it on enough pictures and it learns to tell a cat from a car. The question that has haunted the field is what any single unit inside it is doing.

In 2020 a group of researchers looked inside an image-recognition network and found a unit that fired for cats and also for cars. Nothing connects cats and cars. The conclusion was that networks cram unrelated ideas into the same unit because they run short of units, the way you might store winter coats and tax records in the same closet. The field named this polysemanticity, built a theory called superposition to explain it, and spent years building tools to un-mix the closet. The working assumption became that a trained network is a tangle, and that understanding it means decoding it rather than reading it.

In 2026 Nick Cammarata, one of the people who found the cat-and-car unit, published an essay called Neural Networks Are Already Clean arguing that this was a misreading. Looked at properly, the cat-and-car unit detects a V shape. Cat ears make a V. The front of a car makes a V. The unit had one job all along, and the people studying it had been looking at the wrong level, at the labels on the pictures instead of the shape the unit was built to see.

He then did the audit the field had never done. In a modern language model, about nine in ten units have one main job that a person can state in a sentence. Most of the rest have two closely related jobs. A few percent are not yet understood. Larger networks are not messier than small ones. They are cleaner, because a unit that generalizes to related things wins out over one that packs unrelated things together, and that advantage grows with scale. The reason the old network looked like a mess was that it was an old design that had learned badly, plus tools that were not good enough to read it. He puts his own confidence in the picture at about 85 percent and calls it a working hypothesis. It argues against part of his own field, and the technical paper is still to come. I take it seriously for a reason that has nothing to do with AI.

How they read a network

The method is ordinary science, done by AI agents at a scale no lab could match. For a single unit, an agent collects what makes it fire, what makes it go quiet, and which earlier units feed it. It writes that up as a short paper. Then a second agent, which has never seen the unit, reads only the paper and predicts what the unit will do on new inputs. If the prediction holds, the explanation was real. If it fails, the paper was wrong, and that counts.

For a behavior rather than a unit, the agents trace the path. Feed the network a medical case and ask which units carry the diagnosis from the symptoms to the answer. Remove each one in turn and see whether the answer survives. In one of the essay's examples, four units carry a diagnosis; remove them and it is gone. Draw the units as boxes and the connections as arrows, and you have a map of cause and effect inside the network, not a correlation but a mechanism you can test by cutting it.

The guard against fooling yourself is what the essay calls the near-miss test. An explanation of a unit must predict not only the examples that fire it but the close cousins that do not. A unit described as detecting cat ears should stay quiet for a fox face. Across 183 units the explanations beat the near misses 164 times and lost 4. The point is not any one finding. It is that the whole loop, investigate, write, predict, check, runs without a person in it, so it can be run on every unit in the network.

Every step of that loop depends on one thing: you can do anything you like to the network. Drive a unit, silence it, rerun the input, as often as you want, on an identical copy. The essay calls it the best laboratory in natural science. Keep that in mind, because it is where the brain is different.

Why the brain might be the same, and why nobody knows

Neuroscience has its own version of the cat-and-car neuron, and it has the same reputation. Record from neurons in the front of a monkey's brain while it does a task, and many of them respond to odd combinations: this cue and that reward and which hand is about to move. The field calls it mixed selectivity. It is the brain's polysemanticity, and it carries the same moral: the brain is a tangle, so understanding it means decoding it, not reading it.

The essay's author is not a neuroscientist and says so. But he makes a pointed observation. If the cat-and-car neuron had been in a human brain, seen through a few electrodes or a brain scan, nobody would ever have found out that it detected a V shape. He could not see what it was when it sat in front of him for years with every tool he wanted. So he discounts the tangle reading of the brain almost to zero and guesses that the organizing principle transfers: if a brain learns, it has the same forces that make networks clean, and its neurons probably have jobs you could state in a sentence too. He names the counterexample himself. If neurons in the brain really do respond to many unrelated things, he is wrong.

Here is the catch. Everything in his method depends on being able to run the experiment: drive a neuron, silence it, rerun the input, on an identical copy. A brain in a dish does not allow that. A map of a brain, the complete wiring diagram that is now within reach for a mouse, gives you structure and at most one snapshot of activity. It does not let you ask the neuron what it does.

So the honest version of the question is not whether brains are clean. It is whether, given a map, you can build a model of the brain that runs, and then read the model the way the essay reads a network, and whether the model is right about the real brain. That is three questions, and they can each be tested. We spent a night testing them.

We tried it on the one brain you can run

There is exactly one animal whose brain exists as a complete, runnable model: the fruit fly. Its wiring diagram, about 130,000 neurons and every connection between them, was finished in 2024 by the FlyWire project. The same year a team at Berkeley turned it into a simulation: a simple model of each neuron, wired exactly as in the real fly. It is crude, and it is the best laboratory in neuroscience, because you can drive or silence any neuron in it and rerun the experiment.

We gave AI agents the wiring and the simulator and asked them to do what the essay's agents do. We froze the rules and the pass bar before the first run, and we report every result pass or fail. In plain terms, this is what happened.

The agents beat the simulation. Then we found out why. We collected 259 published fly experiments, each of the form: drive these neurons, does that neuron respond? The raw simulation got the right answer about 83 percent of the time, on a scale where a coin flip scores 50. Agents that investigated each experiment, ran the model, silenced neurons and wrote up the mechanism scored 90. That passed the bar we had set. Then we reran the agents with every neuron's name, chemistry and brain region replaced by a code, so they could still read the wiring and run the model but could not recognize the published paper. Most of the gain disappeared. On the experiments the model's own authors had used to tune it, the blind agents gained nothing. On the experiments nobody had used, they still gained, about 9 points, and that is the honest size of what the map and the model add. The rest had been memory: the AI knew the fly literature and was quietly recalling the answer. We tried to escape this with papers published in 2026, after the AI's training, and could not. The 2026 results were guessable from the experiment's design alone. The full story of the hiding is its own post.

Neuron papers work. We had the agents write a paper on each of 350 types of neuron around the fly's feeding circuit, from the model alone. A fresh agent that could see only the papers, no wiring and no simulator, predicted the experiments as well as the agent with every tool. Each paper carried its own neuron's knowledge; papers about neighboring neurons did not help yet. An agent with both the tools and the papers did best of all. This is the part of the essay that transferred most cleanly: an explanation written once from the model is as useful as the model itself.

The mouse was different. The one piece of mammal brain mapped at this detail is a cubic millimeter of mouse visual cortex, about 200,000 cells, finished in 2025 by a $100 million program. For it we have the wiring and one snapshot of what each neuron responded to, but no model that runs. We asked agents to read 3,000 of its neurons' wiring and predict what each neuron cared about. They did no better than a 20-line rule that guesses from where the neuron sits. Reading the wiring alone, with no location, did worse. Memory was not the explanation here: an agent given only names scored at chance. In a map you cannot run, the readable wiring did not tell us what a neuron does. That result has its own post too.

The product artifact exists. With the fly model we also ran what a drug company would want: a modelled compound that blocks one kind of inhibitory receptor, across 18 standardized tests at four doses, with the agents tracing every effect to a handful of named neurons and writing out what to measure in real tissue to prove them wrong. At full dose the model brain ran away into seizure-like activity, which is what that class of compound does to real animals.

All of this ran in one night, on subscription accounts and one computer with a borrowed graphics card. The full write-ups, with every number, every caveat and the figures, are on braininsights.ai.

What this means for the brain

Three things, in order of how sure I am.

The method transfers in form. Neuron papers, cause-and-effect graphs, agents that investigate and a second agent that checks: all of it ran on a brain, as long as the brain was a model you could run. The agents found real mechanisms in the fly, traced three quarters of a feeding response through one interneuron, and caught the simulation making mistakes. Nothing about the essay's playbook is specific to artificial networks. It is specific to networks you can experiment on.

The bottleneck is the model, not the understanding. The mouse result is the one I keep coming back to. We had the most detailed map of mammal brain ever made and the agents could not read a neuron's job out of its wiring, because they could not ask it. Whether cortex is clean in the essay's sense is still open, and it will stay open until someone turns a map into something that runs. That is the whole gap between a connectome and an understanding, and it is now measured rather than argued. It is also the product: a brain map is worth what the model built on it can predict.

The only honest test is an experiment nobody has run. Our agents knew the literature better than we expected, and newer papers did not help, because biology is predictable from the design of the experiment. The AI field learned the same lesson about benchmarks years ago. For brains, the clean test is a compound or a perturbation that has never been published, predicted by the model before the lab runs it, with the prediction written down first. That is a higher bar than most neuroscience is held to, and it is the right one.

What comes next

The next test writes itself. Take a compound with no published fly data. Have the agents predict, from the fly model, which neurons change and by how much, and write the predictions down before anyone touches a fly. Then run the experiment in a real fly and score the prediction. If the model is right more often than the literature-free baseline, the map-to-model loop works on a brain and the business of building it is real. If it is wrong, we learn exactly where the model fails, which is the next thing to fix. Either way the result goes up, pass or fail, the way these did.

That is what this lab is for. The question this essay asks, whether a brain can be read the way a network is read, is the question a complete map of a mammal brain will answer. The fly says the method works where a model exists. The mouse says the model is the missing piece. We would rather test the hypothesis than assume it, and we are looking for the people who want to run that test with us.

If you work with flies and have a compound or a perturbation nobody has published, or if you read networks for a living and want to try it on a brain, apply to be a fellow. I would love to hear from you!

Dan