What Remains of "It" When You Put a Fruit Fly Brain into a Computer? (Complete Male Drosophila CNS Connectome)
*In-depth study of the male fruit fly full CNS connectome · Cross-verified via four retrieval paths · 2026-09-15 · zhichai.net*
In early September 2026, teams at Janelia and Cambridge published in Cell the complete central nervous system connectome of an adult male fruit fly: 166,700 neurons, 124.2 million synapses, and 25.6 million neuron-to-neuron edges [A]. Within a week, community members had loaded it into Minecraft, Doom, and Beat Saber, prompting a viral framing: today's AI is still learning to understand the physical world, while natural selection has been "training" biological neural networks for 500 million years. The question: how much of this animal's own "algorithms" is actually in the scanned wiring?
The answer up front: the connectome is a netlist, not firmware, and certainly not an operating system. It contains roughly half of an algorithm's compiled artifact—the other half (the plasticity rules) lives in the genomic bottleneck and was never scanned.
Key facts
- 166,700 annotated neurons (including sensory axons; 166,483 with connections)
- 124.2 million synaptic contacts (each contact counted separately)
- 25,582,938 directed neuron-to-neuron edges (the number used in the Minecraft project; media frequently confuse this with 124.2M)
- 11,710 cell types (8,069 shared between sexes)
- NeuroCraft Fly (GitHub, 2026-09-06): claims all 166,700 neurons and 25,582,938 edges in Minecraft, but the README admits code is unreleased; the landing page honestly states that "after shuffling weights, responses persist—a finding that must be reported" [A].
- DOOMFLY (Alex Wormuth): fully open source, but README says "live experimental training, not demonstrated learned survival"—6,385 episodes run, failing all three validation gates [A/B].
- A YouTube Minecraft video (240K views in 3 days) used the female FlyWire model, with hand-scripted reactions [C].
- A Beat Saber demo was overfitted replay of recorded sequences—a tape playing, not a connectome keeping the beat [B].
- More seriously: Eon Systems (March 2026) demoed an embodied full-brain LIF simulation with a MuJoCo body (NeuroMechFly v2, 87 joints), built on the peer-reviewed Shiu et al. 2024 model—but explicitly notes no spontaneous neuromodulation, no learning, no hormonal state, and "essentially decorative" vision [A].
- Shiu et al. 2024 (Nature): a 127,400-neuron LIF full-brain model with weights = synapse count × sign, only one free parameter; 91% of 164 testable predictions matched experiments; shuffling weights collapsed accuracy from 100% to 1% [A].
- Creamer et al. (Nature): connectome-constrained dynamics fit C. elegans optogenetic data at 92% of the data's own reproducibility ceiling; shuffling degraded performance [A].
- Lappalainen et al. 2024 (Nature): the same connectome trained 50 times yields different mechanisms, some losing direction selectivity entirely; removing connection strengths significantly degrades predictions [A].
- Learning rewrites everything: one associative learning event drops relevant synaptic strengths by ~80% on average [A: Hige & Turner 2015]—a static connectome is closer to a pre-training initialization.
- Jonas & Kording's 6502 microprocessor study warns that connectomics methods can see structure but not information-processing hierarchy [A].
- Berg et al., *Cell* 189(18):5504–5526 (2026), doi:10.1016/j.cell.2026.08.015 · male-cns.janelia.org
- Dorkenwald et al., *Nature* 634:124–138 (2024) · Schlegel et al., 634:139–152 · Shiu et al., 634:210–219 · Lappalainen et al., 634:1132–1140
- Eckstein et al., *Cell* 187:2574 (2024) · Hige & Turner, *Neuron* (2015) · Kiral et al., *Cell Reports* (2021) · Zador, *Nat Commun* (2019)
- Jonas & Kording, *PLoS Comput Biol* (2017) · Lechner et al., *Nat Mach Intell* 2:642 (2020) · Lu & Webb, arXiv:2601.16806 · arXiv:2509.13827 · arXiv:2412.14171
- NeuroCraft Fly: github.com/evnsnclr/neurocraft-fly-public · DOOMFLY: github.com/nftechie/doomfly · Eon Systems: eon.systems/updates/embodied-brain-emulation
Three commonly garbled facts
1. This is not the FlyWire team's work. The 2024 Nature papers completed the *female* full brain (FAFB dataset + FlyWire segmentation, 139,255 neurons, 54.5M synapses, Princeton's Murthy/Seung teams). This male version comes from Janelia FlyEM + MRC LMB/Cambridge (Gregory Jefferis & Gerald Rubin's teams): *Berg et al., Cell 189(18):5504–5526, 2026-09-03*, DOI 10.1016/j.cell.2026.08.015, open access [A].
2. Coverage is the entire CNS—central brain + bilateral optic lobes + ventral nerve cord (VNC), continuous through the neck, traceable from eye to leg tip [A]. The female full-CNS dataset (BANC, ~160,000 neurons) was also completed in June 2026. For the first time, both sexes of one species have complete CNS wiring diagrams.
3. Watch the denominators. "125M synapses" headlines vs. the 25.58M edges used in game demos—nearly a 5× difference, same fly.
How the map was made
| Stage | Detail | |---|---| | Sampling | One male fly, nervous system cut into 66 blocks, ~8 nm isotropic resolution [A/B] | | Imaging | 7 electron microscopes running ~1 year: scan surface, shave a layer, repeat [B] | | Segmentation | Google's flood-filling networks + PATHFINDER [A] | | Proofreading | 44 person-years of manual proofreading; only 40.1% of synaptic connections have both sides verified [A] | | Release | v1.0 on 2026-06-08, CC BY, fully downloadable at male-cns.janelia.org [A] |
"Complete connectome" ≠ "every connection manually verified"—about 60% of connections still carry automated-segmentation uncertainty. The real scientific novelty is sexual dimorphism: ~95% of brain cell types are sex-shared, but the <5% of male-specific neurons cause 12% of male neurons to differ in wiring (vs. 4% in females), underpinning courtship and other male-specific behaviors [A/B].
The game demos: how real is the Minecraft fly?
Methodology: you got a netlist, not firmware
The connectome gives you: topology (a 25.6M-edge directed graph), contact counts (a proxy for strength), 11,710 cell types (97.9% cross-matchable), and predicted neurotransmitters (87% per-synapse, 94% per-neuron accuracy [A: Eckstein 2024]).
You must supply yourself: synaptic weights (contact count ≠ strength [A: Liu & Jeanne 2021]), excitation/inhibition signs (~13% single-synapse misclassification), neuromodulation (dopamine, 5-HT, octopamine), electrical synapses ("unrecognizable" in the imaging [B]), intrinsic neuronal properties [A: Lappalainen 2024], and the body itself.
Individual variability compounds this: comparing FlyWire with the hemibrain (two different females), only connections with >10 synapses or >1% of a target's input are highly conserved; about one-third of cell types cannot be reliably re-identified across brains [A: Schlegel 2024]. Even the same genome yields different wiring at different developmental temperatures [A: Kiral 2021].
Evidence showdown: structuralists vs. dynamists
Topology does carry information:
Beyond structure is behavior:
The two camps don't conflict—they compose the answer: **the connectome's information suffices to constrain *what* a circuit computes, not *how* it computes it right now.**
*Correction of a widespread claim:* the often-cited "Lechner constrained an RNN with the worm connectome and found topology, not weights, matters" is inaccurate—the actual paper (*Neural circuit policies enabling auditable autonomy*, Nat Mach Intell 2020) trained all weights end-to-end and did no topology-vs-weight ablation [A].
Conceptual layer: what did natural selection actually "train"?
Zador's *A critique of pure learning* (Nat Commun 2019) names it: the genomic bottleneck. Evolution cannot write 124 million synapses into DNA; it stores a highly compressed set of wiring rules. The connectome is thus a snapshot of "architecture + strong initialization + a plasticity rule set"—not trained weights, and not random initialization either. Even the most innate circuit in neuroscience, the fly's head-direction ring attractor, requires plasticity to recalibrate its compass in every new environment [A].
A neglected third party: the body. Morphological computation and embodied cognition show perception-circuit statistics co-evolve with body geometry. Porting the wiring into Minecraft changes the input/output distributions entirely—the machine runs a draft-zero new algorithm in another world, not the fly's algorithm in this one.
The AI ledger
| | Fly-circuit bio-inspired engineering | End-to-end LLM/RL | |---|---|---| | Trainable parameters | 64,000 (mushroom body + central complex navigation agent) | Millions+ | | Pre-training frames | 0 (first trial SR 0.84 / SPL 0.48) | 10⁷–10⁹ | | Navigation power | 18.6 μW (ring attractor on Loihi) | ≥ 2 W (SLAM FPGA) |
Sources: Lu & Webb, arXiv:2601.16806 [A: preprint, with disclosed limitations]; PMC8881593. Contrast: on VSI-Bench spatial reasoning, humans score 79.2 vs. GPT-5's 55.0, with true weaknesses in perspective-taking and integrated reasoning [A]. On real hardware: an LPLC2 vision model on the Colias microrobot achieved 96.1% collision detection with 70 KB of RAM [A: arXiv:2509.13827].
Verdict: three conclusions
How much of "it" is in the computer? About half: all the wiring, half the strengths, almost none of the current state.
1. Topology, counts, types, and predicted neurotransmitters are real and reusable (Shiu's 91% and the shuffle experiment are ironclad); weights, polarity details, modulators, intrinsic properties, and learning state are absent—and the "weight" layer largely exists in no fly at any moment, only in "generation rules + lived experience." 2. What got uploaded is the factory circuit diagram plus a pre-experience initialization—not the fly. The algorithm has three layers: architectural priors (the connectome delivers these superbly), plasticity rules (not in the map; written in the genomic bottleneck), and current weights (only a snapshot). The shuffled-weights Minecraft result cuts the demo's credibility while confirming some robustness lives in architecture. 3. The real value for AI is not "porting" but "copying." Delivered bio-inspired systems (64K-parameter zero-shot navigation, an 18.6 μW compass, 70 KB collision avoidance) show the value lies in circuit principles—ring attractors, lateral inhibition, sparse coding, mushroom-body "verify-then-store." The full connectome upgrades these principles from conjecture to verifiable blueprints.
> What the connectome preserves is not an algorithm, but half of an algorithm's compiled artifact—the other half (the plasticity rules) is written in the genomic bottleneck and was never scanned. > Rather than asking "how many algorithms are in the map," ask "how many algorithms can this map constrain." Shiu's answer: tightly. Lappalainen's answer: still more than one.
The connectome's true legacy for the AI era: the first synapse-by-synapse verifiable, complete blueprint of perception-action circuits. A blueprint is not a building—but without blueprints, people couldn't even agree on how many beams a building needs.
Number-denomination quick reference
| Number | Denomination | Note | |---|---|---| | 166,700 | Annotated neurons (incl. sensory axons) | Preprint: 166,691; with edges: 166,483 | | 124.2M | Synaptic contacts (each contact one) | Google rounds to "125M" | | 25.6M | Directed neuron-neuron edges | Most-confused figure in media | | 11,710 | Cell types (final version) | Preprint: 11,691 | | 40.1% | Connections with both sides proofread | "Complete" ≠ fully verified |