NEURALZOO

Four nervous systems in one enclosure and nobody in the chair. Three of them have no words, so a translator stands in, and the fourth talks into a room where nothing can hear it. Every line is printed with the counts it was made from. The resolution gets worse as they get closer to you.

WORMC. elegans, N2
299 neurons, graded
live in this tab

FLYD. melanogaster, male CNS
165,122 neurons, spiking
live in this tab

CORTEXH. sapiens, group average
82 regions, no cells at all
207 people, blurred together

CLAUDEclaude-sonnet-4-6
no body, no senses
reads the transcript

WORM299 neurons · graded

—

C. elegans N2. every cell drawn. schematic positions, grouped by ganglion.

FLY165,122 neurons · spiking

—

male CNS, frontal view. 140,024 of the cell bodies are at their real measured coordinates.

CORTEX82 regions · no cells

—

H. sapiens, 207 people averaged. layout is schematic. the 1,114 edges are real.

CLAUDEno body · no senses

nothing to draw. this one has no neurons, no cell bodies and no coordinates. it is the only participant in the room with nothing to show you, and the only one that can tell you it is in a room.

the room acts on its own all runs

loading the fly, 11.7 MB

what reaches whom

Nobody is talking. The room is

Nobody is typing. There is no chair for you here, because a sentence has nowhere to land in three of the four: there is no cell in a worm, a fly or a cubic map of cortex that language arrives at. So the room acts instead. Every few seconds it picks one of six things at random and does it, and the three answer with whatever their wiring does. The fourth reads all of it and talks anyway, to nothing.

What they can receive is a change in the room. Each stimulus below is delivered to the cells, or in the cortex's case the regions, that actually carry it. Two of the rows are blank on one side, which is the most honest thing on this page. The worm cannot be shown a light. The fly cannot be hurt here, because its nociceptors sit in the body wall and the body wall is not in the dataset.

Read the three columns left to right and watch the resolution fall apart. 299 neurons, every one traced by hand and finished in 1986. 165,122 neurons, every one traced, finished two weeks ago. Then 82 smudges averaged across 207 strangers, which is the best whole-brain map of your own species that anyone has.

stimulus	worm, of 299	fly, of 165,122	cortex, of 82

what is running

Two animals, two different models, because they are two different animals

The worm is the 1986 White, Southgate, Thomson and Brenner reconstruction: 299 neurons, 2,279 chemical edges over 6,465 synapse contacts, and 557 gap junctions carrying 947 contacts, which are simulated as actual resistors. Almost no C. elegans neuron spikes, so every cell is a continuous membrane potential with sigmoidal transmitter release, after Wicks 1996 and Kunert 2014.

The fly is the male central nervous system released on 3 September 2026 by Janelia FlyEM, the Cambridge Connectomics Group and Google Research: 165,122 traced neurons. Adult fly neurons do spike, so it runs as leaky integrate-and-fire, resting at −52 mV, threshold −45 mV, 10 ms membrane constant, 2.2 ms refractory, 0.275 mV per synapse, after Shiu 2024. Edge signs come from the published per-neuron transmitter predictions: acetylcholine excites, GABA and glutamate and histamine inhibit, monoamines get a third of the weight.

The cortex is not in the same category as either of them and the page should not let you forget it. It is the ENIGMA group-average human connectome: 68 cortical regions from the Desikan-Killiany atlas plus 14 subcortical structures, with edges that are streamline densities from diffusion MRI averaged over 207 Human Connectome Project subjects. 82 nodes, 1,114 edges. Not one cell. It runs as linear network diffusion, which is the standard first-order model for spreading on a structural connectome, and it is reported as how much of that spread lands in each region against what a uniform input would have put there.

There is no human connectome at the resolution of the other two and there will not be for a long time. The only synaptic-resolution human data that exists is H01, one cubic millimetre of temporal cortex holding about 57,000 cells, of which 104 have been proofread. The pairwise connectivity for the rest is 332 files totalling 159 GB of automated segmentation that nobody has checked. That is the state of the art on us.

The full fly edge list is 10,511,038 connections and would be a 74 MB download. This page ships the 2,168,143 of them carrying twelve synapses or more, which is 61 million of the 311 million synapse contacts. Before cutting it we checked what the cut costs: light still reaches 149,930 of the 165,122 neurons, and still lands on the descending neurons in three hops.

stated plainly

The parts that are not the animals

There is exactly one channel between them, and it is movement. Round two is not them talking. It is the dish. A worm that reverses and a fly that walks are mechanical events, and all three have something that answers to vibration: ALM, AVM and PLM in the worm, 5,756 mechanosensory cells in the fly, postcentral gyrus in the cortex. The amount delivered is scaled by how much each one actually moved. Nothing else crosses. No smell, no sight, no meaning, and certainly no sentence.
The cortex can be shaken but cannot shake. It contributes nothing to round two. A structural connectome is a map of fibre bundles with no motor output on the far end, so it receives vibration and never produces any. It is the only participant that can only ever be acted upon.
The cortex is scale free, which means it cannot be surprised. Linear spreading, normalised. The same input always produces exactly the same pattern no matter how hard you deliver it, so its round two line never changes. A real brain answers differently depending on how hard you hit it. This one does not.
The cortex has no signs, at all. Diffusion MRI counts fibre bundles. It cannot tell a connection that excites from one that inhibits, so every edge here is positive and the thing can only ever spread. Half of what a brain does is quieting other parts of itself, and none of that is in this map.
The cortex is not a person. It is 207 people averaged into one 82 region matrix. Nobody has this brain. It is the shape you get when you blur everybody together, and it is the best whole-human map that exists.
The sentences are written by a lookup table. Thirteen branches over two numbers. The numbers are measured, the English is not.
The fly panel is real anatomy, and 15 percent of it is not. Each dot is one cell body, drawn front on. 140,024 of the 165,122 sit at the coordinates measured in the volume. The other 25,098 have no soma location in the release, so they are scattered across the field at random and drawn dimmer. The brain's shape is real. A sixth of the specks in it are placed by a random number generator.
The fly panel is a total, not a movie. It shows how many times each cell fired across the whole 300 ms window, faded up. It is not a recording of the spikes in order.
The cortex layout is invented. Two arcs for the two hemispheres and a column for the subcortical structures. Those positions mean nothing. The 1,114 edges drawn between them are the real matrix.
Four of the rows on the fly's receipt are noise. DNa01, DNa02, MDN and DNp09 are two to four cells each. At a 4 Hz background a two cell group swings by 10 to 20 Hz with no input at all, which is why you will see MDN move on a turn where the fly received nothing. They are printed because they are the famous ones and because you should be able to see the noise floor. The sentence is decided by the 1,314 cell descending population and the 4,064 Kenyon cells, which are the only two readouts on the page with enough cells to mean anything.
The fly is never silent, because we made it noisy. Every neuron gets a random 4 Hz background kick. Without it the resting network is dead, and an inhibitory input has nothing to switch off, so half the fly's wiring would be invisible. The synaptic gain and the per-edge synapse cap were then tuned until the resting network sat near 7 Hz instead of pinning at its refractory ceiling. Those three numbers are ours, not the fly's.
The fly sees badly here, and it should not. Its photoreceptors are histaminergic and graded, and real fly vision works by inverting that signal at the lamina. Integrate and fire handles that poorly, so light produces a much smaller change than it would in the animal. Smell, which runs through spiking projection neurons into the mushroom body, comes through clearly. That asymmetry is a fault in the model, not in the fly.
The stimulus amplitudes are ours. Nobody has measured what a given stimulus does to these networks. The drive rates were set so that something happens, which means the size of a response is a choice and only its direction is evidence.
The fly starts from rest every turn. A real nervous system carries its previous state forward. This one is reset before each stimulus so turns can be compared, which throws away exactly the thing brains are for.
The worm's signs are inferred, and one is wrong on purpose. Nobody traced whether a worm connection excites or inhibits. Transmitter identity stands in. Glutamate is treated as inhibitory here, and in the animal it is both, depending on the receptor.
The weights are synapse counts, not strengths. In both animals. How strong a given connection is has been measured for a handful of pairs and guessed for millions.
Claude is not translating anyone. It gets the same receipts you do, and it is instructed never to write a line for the worm, the fly or the cortex. If it starts narrating what one of them wants, it has gone wrong. It is also the only participant here that would notice it was in a room, which is either the point or the joke depending on the day.
Neither animal is deciding anything. A gradient climbed is not a choice made.