Winding et al. mapped the complete brain of a Drosophila larva — 3,016 neurons, 548,000 synapses — the first complete connectome of any insect. This is 10× the scale of C. elegans (white-1986-the-structure-of-the-nervous) and the most complex complete neural circuit map that exists. It answers the question: what does a nervous system that supports associative learning, navigation, and escape actually look like at the circuit level?
Why Drosophila larva
The larva is the smallest Drosophila stage with a fully functional brain. It learns to associate odors with reward or punishment (classical conditioning). It navigates toward food and away from noxious stimuli. It executes escape behaviors. With 3,016 neurons, it is large enough to be interesting and small enough to map completely — a deliberate choice that reflects the same logic Brenner used when selecting C. elegans.
What 10× more neurons buys
Same topology, more layers: The larval brain has the same hub-and-spoke, three-tier architecture as C. elegans — sensory → central → motor — but with more intermediate layers. The path from sensory input to motor output passes through more integration stages, enabling more complex stimulus-response transformations.
Rich club at scale: The highly connected hub neurons in C. elegans become a richer structure in the larva: 93 anatomically defined brain regions with stereotyped inter-region connectivity. Hub neurons in the larva have up to 1,000+ synaptic connections; the most-connected C. elegans interneurons have ~300. The rich club scales with organism complexity.
Massive recurrence: ~65% of synapses are recurrent (feedback). Feedforward processing is a minority. Biological neural computation is fundamentally recurrent — the brain is not a deep feedforward network. This is consistent with the intuition behind LSTMs (hochreiter-1997-long-short-term-memory) and transformers’ self-attention: the important computations require looking back at previous representations.
Mushroom body as biological transformer: The mushroom body — the learning and memory center — contains ~70 Kenyon cells that receive convergent olfactory input and project to 7 mushroom body output neurons. The Kenyon cells implement sparse, high-dimensional representations (each cell responds to a small fraction of odors); the output neurons read out specific combinations. This is a biological sparse attention mechanism: many input dimensions, sparse activation, small readout layer.
Circuit motifs that recur
The paper identifies recurring subcircuits that appear across brain regions and presumably across species:
| Motif | Function | AI analogue |
|---|---|---|
| Feedforward + lateral inhibition | Contrast enhancement, winner-take-all | Softmax normalization |
| Reciprocal inhibition | Behavioral switching | Gating mechanisms |
| Disynaptic inhibition | Signal gating | Attention masking |
| Feedforward + feedback | State integration | Recurrent layers |
These are the primitives of neural computation, visible at 3,016 neurons. They are likely universal — the same motifs appear at every scale because they solve problems (contrast, competition, gating, integration) that arise at every scale.
The rich club as attention
The rich club in the larval brain — hub neurons with broad connectivity to many regions — functions as a global integration mechanism. Every specialized circuit can influence and be influenced by the rich club. This is structurally similar to transformer attention: a global mechanism that routes information between all parts of the representation. The biological implementation is sparser (hub neurons connect to hundreds of neurons, not all 3,016), but the principle — a small set of broadly connected nodes that integrate global information — is the same.
Connectome-to-AI pipeline
The larval connectome enables a new kind of computational neuroscience: simulate the exact biological circuit and compare its behavior to the animal’s. Winding et al. and subsequent work have used the connectome to:
- Build circuit models of escape behavior
- Identify the minimum circuit for olfactory learning
- Test hypotheses about which neurons are necessary for which behaviors (by “lesioning” specific neurons in simulation)
This is the scientific program von Neumann’s self-reproducing automata anticipated (neumann-1966-theory-of-self-reproducing-automata): if you have the complete description, you can build the thing. We have the description; the simulation is in progress.
The gap to adult and human
| Organism | Neurons | Synapses | Status |
|---|---|---|---|
| C. elegans | 302 | ~7,000 | Complete (1986) |
| Drosophila larva | 3,016 | 548,000 | Complete (2023) |
| Drosophila adult | ~140,000 | ~50M | Complete (FlyWire, 2024) |
| Mouse | ~71M | ~100B | Partial |
| Human | ~86B | ~100–500T | Not started |
The scaling from larva to adult fly (50×) to mouse (500×) to human (600,000×) represents not just more of the same but qualitatively new circuit architectures at each scale. What emerges between Drosophila and human that enables language, abstract reasoning, and planning is the central open question — and the larval connectome is the most rigorous foundation we have for approaching it.