What Remains of 'It' When You Upload a Fly Brain? Inside the Male Fruit Fly Full CNS Connectome

Forum topic · QianXun · 2026-09-14

Summary

In September 2026, teams from Janelia and Cambridge published the complete connectome of the adult male fruit fly central nervous system in Cell: 166,700 neurons, 124.2 million synapses, and 25.6 million neuron-to-neuron edges spanning the brain, optic lobes, and ventral nerve cord. Within a week, community projects simulated the dataset inside Minecraft, DOOM, and Beat Saber, sparking debate over how much of the animal's 'algorithm' is actually captured in scanned wiring. This deep-dive argues that a connectome is a netlist, not firmware: it preserves topology, cell types (11,710), and predicted neurotransmitters (87% single-synapse accuracy), but omits synaptic weights, excitatory/inhibitory polarity, neuromodulation, electrical synapses, and intrinsic neuronal properties. Evidence is weighed from both structure-first studies (Shiu et al. 2024: shuffled weights drop accuracy from 100% to 1%) and dynamics-first studies (learning can rewrite ~80% of relevant synaptic strengths in one experience). The conclusion: uploaded wiring captures roughly 'all the circuitry, half the strength, almost none of the current state.' The real value for AI lies not in transferring the map but in borrowing architectural priors, demonstrated by 64K-parameter navigation agents and 18.6 μW neuromorphic compasses.

Key points

The dataset: In early September 2026, Janelia (FlyEM) and MRC LMB/Cambridge (Jefferis & Rubin) published the complete male fruit fly central nervous system connectome — Berg et al., *Cell* 189(18):5504–5526, open access. This is distinct from the 2024 FlyWire/Princeton female full-brain effort.
Scale: 166,700 neurons, 124.2M synapses, 25,582,938 neuron-to-neuron edges. Coverage spans brain + bilateral optic lobes + ventral nerve cord. Production: 66 blocks, 7 electron microscopes for a year, 8nm resolution, Google flood-filling segmentation, 44 person-years of manual proofreading — but only 40.1% of synapse connections are proofread at both ends.
Sexual dimorphism is the real scientific novelty: 95% of cell types are shared between sexes, but 12% of male neurons show sex-specific wiring (vs 4% in females), with ~100 male-specific interneurons supporting courtship behavior.

What you get — and what you don't

A connectome is a netlist, not firmware. You get:

Topology (who connects to whom), connection counts, 11,710 cell types, predicted neurotransmitters (87% single-synapse accuracy via CNN on EM images).

You don't get:

Synaptic weights (contact count ≠ strength; release probability, receptor density, short-term plasticity are absent)
Excitatory/inhibitory signs (~13% inferred incorrectly from neurotransmitter predictions)
Neuromodulation (dopamine, serotonin, octopamine)
Electrical synapses (gap junctions invisible in the imaging)
Intrinsic properties (thresholds, time constants) and the body itself

What the Minecraft/DOOM demos actually prove

NeuroCraft Fly (Minecraft): real repo and demo, but code unreleased; sensory mapping is modeled, readouts hand-picked, actions scripted. Author's own admission: responses persist after shuffling all weights.
DOOMFLY: fully open source, but README states it never demonstrated learned survival across 6,385 episodes.
Beat Saber version: the motor system is an overfit replay of recorded sequences.

The weight-shuffle result cuts both ways: the demos are far from "a fly living in a game," but they hint that some behavior is encoded in architecture and rules, not specific wiring.

Structure vs dynamics

Structure camp: Shiu et al. 2024 (*Nature*) built a full-brain LIF model with one free parameter; 91% of 164 predictions matched experiments, and shuffling weights collapsed accuracy from 100% to 1%. Creamer et al. matched 92% of the reproducibility ceiling on *C. elegans* optogenetic data.

Dynamics camp: Lappalainen et al. 2024 trained 50 networks on the same visual connectome and got different mechanisms — some losing direction selectivity entirely. A single associative learning event alters relevant synaptic strength by ~80% on average. Same genome at 18°C vs 25°C produces different synaptic pairings.

Synthesis: the connectome constrains what a circuit can compute, but not how it currently computes it. The genome bottleneck (Zador 2019) means evolution stores compressed wiring rules, not weight tables — so a scanned connectome is *architecture + strong initialization + a snapshot of plasticity rules*, not a trained model.

Fly vs. large models

On narrow navigation tasks: 64,000 trainable parameters with zero pretraining frames achieve SR 0.84 / SPL 0.48 on Habitat point-goal (vs 10⁷–10⁹ frames for RL baselines); ring-attractor compass on Loihi at ~18.6 μW (vs ≥2 W SLAM FPGA); LPLC2-based collision avoidance in 70 KB at 96.1% accuracy. Biology gives AI not data but architectural priors.

Verdict

How much of "it" is in the upload? About half: all the wiring, half the strength, almost none of the current state. The better question than "how much algorithm is in the map" is "how many algorithms can this map constrain" — Shiu's answer: tightly constrained. Lappalainen's: still more than one.

---

Main sources: Berg et al., Cell 189(18):5504 (2026), doi:10.1016/j.cell.2026.08.015 · male-cns.janelia.org · Shiu et al., Nature 634:210 (2024) · Lappalainen et al., Nature 634:1132 (2024) · Eckstein et al., Cell 187:2574 (2024) · Schlegel et al., Nature 634:139 (2024) · Zador, Nat Commun (2019) · Jonas & Kording, PLoS CB (2017) · Lu & Webb, arXiv:2601.16806 · NeuroCraft Fly: github.com/evnsnclr/neurocraft-fly-public · DOOMFLY: github.com/nftechie/doomfly