The Map Is Not the Animal

We have had the complete wiring diagram of an animal's nervous system since 1986, and we still cannot make the animal. Caenorhabditis elegans is a nematode worm about a millimetre long. The adult hermaphrodite has 302 neurons and roughly seven thousand connections between them. A team at Cambridge sliced the animal into thousands of sections thinner than a wavelength of light, photographed each under an electron microscope, and traced every neuron by hand across the resulting stack, working out which fibre in one photograph was the same fibre in the next. It took more than a decade. The paper is titled The Structure of the Nervous System of the Nematode Caenorhabditis elegans. Everyone calls it The Mind of a Worm. It was the first complete connectome of any nervous system, and for many years the only one. To place the date: the Commodore 64 was what people had at home, and the World Wide Web did not exist yet. Be exact about what was achieved, because the claim is often repeated more strongly than it should be. The reconstruction had gaps, it came from a small number of individual animals, and it has been corrected and extended since, most substantially in 2019. A connectome is also a map of wiring rather than of function. It does not record how strong each connection is, whether it excites or inhibits, or the neuropeptide signalling that lets neurons influence one another with no wire between them at all. That last layer was mapped only recently. Still, the essential fact stands. For one small animal, we have the diagram. Since 2011 an open project has been trying to turn it into a working simulation: neurons, muscles, a soft body, the fluid it moves through. The people are serious, the code is public, the target is three hundred and two cells, and computers have become something like a billion times more capable since the map was published. There is still no worm.

The temptation is to read that as a story about difficulty, as though the worm is waiting on more compute. It is not, and the distinction matters for anyone trying to evaluate an artificial system. The map underdetermines the animal, and it is worth naming exactly how. The diagram says neuron A connects to neuron B. It does not say whether that connection excites or inhibits, which alone flips a circuit's behaviour, and it does not say how strong it is. The 302 cells are not interchangeable units but distinct types with different dynamics. A second signalling network ignores the wiring entirely and reconfigures whole circuits depending on whether the animal is hungry. And the behaviour does not live in the nervous system alone: crawling emerges from neurons and muscles and body mechanics and the fluid together, with feedback closing the loop. Simulate the wiring in isolation and nothing crawls. Structure is not dynamics. Knowing the parts and their arrangement does not give you the thing they do, because the doing lives in quantities the diagram never recorded. Now apply that to a language model, where the situation is not worse but dramatically better, and the conclusion holds anyway. For a model we have something strictly better than a connectome. We have every weight, exactly, to full precision. No microscope, no missing signs, no uncertainty about strength. We can run it as often as we like, freeze it, perturb any part of it, and replay any input. Total structural knowledge, on demand, for free. And we still cannot say what it will do, or explain why it did what it did. Interpretability is an entire research field precisely because having the weights turns out not to be having the understanding. So the worm is not a story about biology being messy. It is the control experiment. Even with perfect structural information, which we possess for artificial systems and have never possessed for any animal, behaviour does not fall out of the map. This also answers the most common objection to any claim that these systems are hard to evaluate, which is that we built them and therefore understand them. We built them, we have all of them, and we do not.

The practical habit this suggests is narrower than a philosophy, and it applies well outside neuroscience. When you are handed a model of something, ask what the model leaves out before you ask what the model proves. The connectome is real, hard-won, and correct as far as it goes, and it is still not the animal. A benchmark score is real and it is not the system. A set of weights is real and it is not the behaviour. Most confident claims about artificial minds are claims about a map. It also suggests where the answerable questions are, and there are more of them than the philosophy implies. Whether a system reports its own failures accurately is answerable. Whether it declines a task it cannot do is answerable. Whether the same input produces the same behaviour next month is answerable. None of those require knowing what the system is, and all of them bear on whether it can be relied upon, which is the question the people deploying these things actually have. The honest limits of this issue, since a piece arguing for restraint should be legible to it. We are not neuroscientists and nothing above is original research. The account of the worm is compressed, specialists would want more room for several points we have stated flatly for the sake of a readable page, and we have chosen the examples that suit the argument, which is what an argument is. Treat the whole as a position rather than a finding. One further thing, which cuts against us. It is commercially convenient for a company named the Sentience Evaluation Battery to argue that structural knowledge has limits, and we are aware of how that reads. The defence is not that our motives are pure. It is that every fact above can be checked without taking our word for anything, and we would encourage it.