This paper gave us the first complete wiring diagram of any nervous system. The organism is a 1mm worm with 302 neurons and ~7,000 synapses. The question it posed — if you know every connection, can you predict behavior? — is still the central question of computational neuroscience and the ultimate test of whether neural network models capture anything real about intelligence.
Why C. elegans
Sydney Brenner chose C. elegans in 1963 because it was the smallest organism with a nervous system complex enough to exhibit behavior: chemotaxis, mechanosensation, thermotaxis, associative learning. The adult has exactly 959 somatic cells across all individuals — the cell lineage is completely invariant. This made it the only organism where “map the entire nervous system” was a realistically finite project.
The work took over a decade of serial-section electron microscopy: cut the worm into ~50nm slices, photograph each slice under an electron microscope, trace every neuron by hand through thousands of images. White et al. completed this in 1986. It remains one of the most labor-intensive datasets in biology.
What the connectome showed
Hierarchical topology: Sensory neurons → interneurons → motor neurons, with a small rich-club of hub interneurons (AVA, AVB, AVD, AVE, PVC) receiving convergent sensory input and projecting broadly to motor circuits. The same three-layer, hub-and-spoke topology appears in every complex nervous system subsequently mapped — it may be universal.
Stereotypy: The wiring is nearly identical across individuals. You can label any neuron by its name (e.g., AVAL, AVBR) and its connectivity will be the same. This means the worm’s behavior is essentially deterministic from the genome, not learned from experience. There are ~20 behaviors, all encoded in 302 neurons.
Circuit motifs: Reciprocal inhibition between forward (AVB→B motor) and backward (AVA→A motor) circuits implements winner-take-all switching between locomotion directions. This same motif — two circuits that inhibit each other, with only one active at a time — appears in vertebrate locomotor circuits, insect escape circuits, and presumably anywhere a binary behavioral choice is needed.
OpenWorm: the artificial C. elegans
The OpenWorm project (2011–present) implemented the C. elegans connectome as a simulated neural network, combined it with a physics model of the worm’s body, and embedded it in a virtual environment. The simulated worm’s locomotion and chemotaxis emerge without additional programming — behavior falls out of the wiring diagram and the body-environment interaction.
This is ALife’s most complete achievement: a fully simulated organism whose behavior emerges from a biologically accurate circuit, not from hand-coded rules. OpenWorm has also been loaded onto small robots (Lego Mindstorms), which then exhibit worm-like phototaxis and obstacle avoidance. The worm’s nervous system runs on silicon.
The implication: for an organism with 302 neurons, we understand the relationship between wiring and behavior well enough to reproduce it. The question is whether this understanding scales — and what changes when you go from 302 to 3,016 (winding-2023-the-connectome-of-an-insect) to 86 billion.
Connection to Brooks’ embodied AI
Brooks argued that intelligence is not in the head alone but in the coupling of body, brain, and environment (brooks-1991-intelligence-without-representation). C. elegans is the biological proof: the worm’s chemotaxis behavior requires the chemical gradient in the environment, the sensory neurons that read it, the interneurons that integrate it, and the muscles that respond. Remove any component and the behavior disappears. The “intelligence” is distributed across all three.
Minimal sufficient neuroscience
C. elegans establishes a lower bound: 302 neurons is enough for 20 behaviors, sensory integration, and rudimentary associative learning. This is the minimum complexity threshold von Neumann predicted (neumann-1966-theory-of-self-reproducing-automata) — below this, systems degenerate; above it, they can support open-ended behavior.
For AI, this means: whatever computational principles are at work in 302 neurons (sparse connectivity, hub topology, reciprocal inhibition, winner-take-all) are sufficient for a minimal behavioral repertoire. Any artificial system that implements these principles should exhibit similar capabilities.