Uploading a Fruit Fly Brain Into a Computer: How Much of "It" Survives? (Complete Male Drosophila CNS Connectome)

*Deep-dive on the complete male fruit fly CNS connectome · cross-verified across four source tiers · 2026-09-15*

In early September 2026, teams at Janelia and Cambridge (MRC LMB) published in Cell the complete connectome of the adult male fruit fly central nervous system: 166,700 neurons, 124.2 million synapses, and 25.6 million neuron-to-neuron directed edges [A].

Within a week, community members put it into Minecraft, Doom, and Beat Saber. The shared 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. So the question arises—how much of this animal's own "algorithms" is already encoded 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 the compiled artifact of the algorithm*—the other half of the source code (plasticity rules) lives in the genomic bottleneck and was never scanned.

All figures carry source-tier labels (A = paper/official, B = authoritative media, C = second-hand, D = unverified).

Key facts

Three commonly garbled facts

First, this was not the FlyWire team. The 2024 Nature papers completed the female fly full brain (FAFB dataset + FlyWire segmentation, 139,255 neurons, 54.5M synapses, Princeton Murthy/Seung). The male version comes from Janelia FlyEM + MRC LMB/Cambridge (Gregory Jefferis & Gerald Rubin teams): *Berg et al., Cell 189(18):5504–5526, 2026-09-03*, DOI 10.1016/j.cell.2026.08.015, open access [A].

Second, coverage is the entire CNS—central brain + bilateral optic lobes + ventral nerve cord (VNC), contiguous through the neck, traceable from eyes to leg tips [A]. The corresponding female full-CNS dataset (BANC, ~160,000 neurons, Nature, June 2026) also landed this year. For the first time in biology, both sexes of one species have complete CNS wiring diagrams.

Third, mind the denominators. Headlines say "125M synapses"; the paper says 124.2M. The 25,582,938 figure used in game demos is the count of directed neuron-to-neuron edges (one edge per neuron pair regardless of contact count)—nearly a 5× difference describing the same fly.

How the map was made

| Stage | Detail | |---|---| | Sampling | One male fly, nervous system cut into 66 slabs, ~8 nm isotropic resolution [A/B] | | Imaging | 7 electron microscopes running ~1 year: scan surface → mill layer → repeat [B: UKRI/MRC] | | Segmentation | Google flood-filling networks + PATHFINDER [A] | | Proofreading | 44 person-years of manual proofreading; 94% presynaptic / 42% postsynaptic attribution done; only 40.1% of synaptic connections proofread at both ends [A] | | Release | v1.0 published 2026-06-08, CC BY, fully downloadable at male-cns.janelia.org [A] |

Note the last row: "complete connectome" ≠ "every connection proofread." Roughly 60% of synaptic connections still carry automated-segmentation uncertainty. The "it" we discuss is itself a noisy map.

Connectome chronology

| Dataset | Scope | Neurons | Synapses | Published | |---|---|---|---|---| | MaleCNS v1.0 (this paper) | male · full CNS | 166,700 | 124.2M | Cell 2026 | | FAFB / FlyWire | female · full brain | 139,255 | 54.5M | Nature 2024 | | BANC | female · full CNS | ~160,000 | — | Nature 2026-06 | | FANC | female · VNC only | ~14,600 | ~45M | Nature 2024 | | Larval brain | larva · brain | 3,016 | 548K | Science 2023 |

The male dataset's key scientific increment is sexual dimorphism: ~95% of brain cell types are shared between sexes; <5% of male neurons are sex-specific or dimorphic (largely *fruitless*/*doublesex*-expressing), yet these drive wiring differences in 12% of male neurons (vs 4% in females); ~100 male-specific interneurons form a densely interconnected network supporting courtship [A/B].

Wiring differences are behavioral differences—live evidence that the connectome carries information.

The Minecraft fly: how real is it?

Within a week of publication, the community ran MaleCNS inside games. First-hand evidence shows it's real but far weaker than the legend:

• 2026-09-04 · tweet: Evan Sinclair Smith claimed MaleCNS v1.0 running in Minecraft with all 166,700 neurons and 25,582,938 edges; self-described help from "GPT-6 Astra" (unsubstantiated, D) [B].
• 2026-09-06 · GitHub: NeuroCraft Fly repo appeared with a 2m27s demo—but the README states code not yet released; the repo is a landing page [A].
• Same day · DOOMFLY: Alex Wormuth's Doom version, fully open source. README: "live experimental training, not demonstrated learned survival"—6,385 episodes run, failing all three verification gates (vision, conditioning, survival) [A/B].
• 2026-09-07 · YouTube: Ro0oney's "I Put A Fly's Conscious Brain into Minecraft" (240K views in 3 days) actually used the female FlyWire model, required hand-coded reactions, and early versions lacked steering neurons [C].
• Beat Saber: the author clarified the motor system was overfit playback of recorded sequences—a tape recorder, not the connectome keeping the beat [B].

The most damning line, written by the authors themselves

The NeuroCraft Fly README candidly notes: "responses persist after weight shuffling—this is a finding that needs to be reported." Two readings, both valid:

1. The demo is far from "a fly alive in a game"; 2. More profoundly—some aspects of behavior may live in architecture and rules, not in specific wiring.

Meanwhile, far more serious work: Eon Systems (San Francisco) released an embodied whole-brain simulation in March 2026—a LIF whole-brain model with a MuJoCo physical body (NeuroMechFly v2, 87 joints), demonstrating food-seeking and dust-triggered grooming. Its foundation is the peer-reviewed Shiu et al. 2024 (Nature) model. But Eon itself states this "should not be interpreted as structure alone being sufficient to recover the fly's full behavioral repertoire," with no spontaneous neuromodulation, no learning, no hormonal state, and "largely decorative" vision [A].

Methodology: you get a netlist, not firmware

The connectome gives you:

• Topology — a 25.6M-edge directed graph of who connects to whom
• Connection counts — synapse contacts per neuron pair (a proxy for strength)
• Cell types — 11,710, 97.9% cross-matchable to existing datasets
• Predicted neurotransmitters — CNN inference from EM, 87% single-synapse / 94% single-neuron accuracy [A: Eckstein 2024]

You must supply yourself:

• Synaptic weights — contact count ≠ strength: release probability, receptor density, short-term plasticity are absent [A: Liu & Jeanne 2021]
• Excitation/inhibition — EM can't read signs; neurotransmitter inference errs on ~13% of single synapses
• Neuromodulation — dopamine, 5-HT, octopamine: same wire, different states, different outputs
• Electrical synapses — gap junctions "cannot be resolved in connectomics imaging" [B: Curr Biol commentary]
• Intrinsic properties — thresholds, time constants, refractory periods: "single-neuron and single-synapse properties are unknown" [A: Lappalainen 2024]
• The body — leg lengths, wing mass, light intensity: all outside the brain

> The connectome is a netlist, not firmware, not an OS. > Like chip reverse-engineering: a schematic and a BOM still leave you short of component parameters, power delivery, and the burned-in program—and that last layer may not exist in the chip at all, but in how it is used.

Is synapse count even a usable strength proxy? Evidence is mixed:

• For: in larvae, contact count predicts contact area, which correlates linearly with strength [A: Barnes & Cardona 2022]; in the fly olfactory circuit, between-connection synapse-count variance (CV=0.31) is the dominant anatomical source of strength variance [A: Tobin 2017].
• Against: predicting somatic PSP amplitudes still requires physiological data—"pure counting is not enough" [A: Liu & Jeanne 2021].

Worse, individual variability: comparing FlyWire and hemibrain (two different female flies), only connections with >10 synapses or >1% of a target's input are highly conserved; the most common mushroom body cell type differs nearly twofold in count; about one-third of cell types cannot be reliably re-identified across brains [A: Schlegel 2024]. Even one genome grows different graphs: flies raised at 18°C form synaptic pairings absent at 25°C [A: Kiral 2021].

Evidence showdown: structure camp vs dynamics camp

Structure camp: the connectome genuinely carries information

• Shiu 2024 (Nature): 127,400-neuron LIF whole-brain model, weights = synapse count × sign, one free parameter; 91% of 164 testable predictions matched experiment; unbiased screening of 106 SEZ cell classes predicted feeding neurons at >90% accuracy [A]
• Killer ablation: shuffling synaptic weights dropped accuracy from 100% to 1%—the real wiring works, not the model form [A]
• Creamer et al. (Nature): connectome-constrained dynamics fit whole-*C. elegans* optogenetic data at 92% of the data's own reproducibility ceiling; shuffling hurt, extra edges didn't help [A]
• Behavioral control: sex-specific wiring corresponds to sex-specific courtship behavior—a natural experiment in structure-behavior causality [A/B]

Dynamics camp: beyond structure lies behavior

• Lappalainen 2024 (Nature): connectome-constrained visual models matched 26 experiments—but 50 trained networks grew different mechanisms from the same connectome, with one T4c cluster completely losing direction selectivity; dropping connection strengths (topology only) significantly degraded predictions [A]
• The worm lesson: modeling directly from anatomical weights "does not produce correct behavior"; weights must be relearned [A/B: Kunert 2014; Choi & Kim]
• Learning rewrites everything: a single associative learning event drops relevant synaptic strength by ~80% on average [A: Hige & Turner 2015]—the static connectome is closer to a pre-training initialization
• Methodological alarm: Jonas & Kording applied full connectomics methodology to a 6502 microprocessor—"structure visible, but information-processing hierarchy cannot be meaningfully characterized" [A: PLoS CB 2017]; Brette: "anatomical connectivity is very limited in informing function" [A/B]

The two camps do not contradict; together they are the answer: Shiu's success relied on "topology + minimal dynamics + experimenter-given inputs," sidestepping all missing items; every dynamics-camp counterexample appears where spontaneous dynamics, learning, or neuromodulation are required.

The connectome's information content suffices to constrain what a circuit can compute, not how it is computing right now.

A fitting footnote: early reconstructions identified only ~50% of T4 input synapses due to resolution, so the 2013 direction-selectivity conclusions failed to replicate under 2017 high-resolution reconstruction [A: Takemura 2017]—map fidelity itself is a biting variable.

Correcting a widespread claim

A popular Chinese-language citation says "Lechner constrained an RNN with the worm connectome and found topology, not weights, did the work." Cross-checked: the real paper is *Neural circuit policies enabling auditable autonomy* (Nature Machine Intelligence 2020), using 19 control neurons / 253 synapses of worm-like topology for autonomous driving—but all weights were trained end-to-end; no topology-vs-weight ablation appears. The claimed title "Designing and optimizing nervous systems with RNNs" does not exist. The genuine topology-vs-weight evidence is Shiu's shuffle experiment (pro-topology) and Lappalainen's multi-mechanism finding (pro-"topology does not fix behavior"). [A]

Conceptual layer: what has natural selection actually been "training"?

Zador's *A critique of pure learning* (Nature Communications 2019) names the answer: the genomic bottleneck—wiring diagrams are too complex to write gene-by-gene, so evolution stores not a weight table but a highly compressed set of wiring rules [A].

> Natural selection never wrote 124 million synapses into DNA. It wrote a compression algorithm— > so the scanned connectome is neither "trained weights" nor a random initialization, > but a snapshot of "architecture + strong initialization + a set of plasticity rules." > The weight table doesn't fit in DNA; what fits is the program that generates weight tables.

Three pieces of evidence nail this down:

1. Learning instantly rewrites what the connectome doesn't give you: one associative learning event drops synaptic strength ~80% (Turner: "that's huge") [A]. 2. One genome grows different graphs: development temperature changes synaptic pairings [A: Kiral 2021]—part of the map is developmental noise. 3. Even the most "innate" circuits only provide scaffolding: the fly's head-direction cells (E-PG ring attractor) are neuroscience's closest thing to a hard-coded compass—yet the correspondence between activity bumps and external landmarks is arbitrary, and flies must recalibrate the compass through plasticity in every new environment [A].

The forgotten third party: the body

Wiring the fly into Minecraft swaps the biggest hidden variable: the body. Morphological computation and embodied cognition literature is clear: perception–circuit statistics co-evolve with body geometry (Pfeifer & Bongard: "some processes are done by the body; otherwise the brain must do them") [A]. Strip the 87-joint mechanics, replace vision with Minecraft pixels, script the body programs—and the input-output distributions change entirely. That machine runs not the fly's algorithm in the physical world, but draft zero of a new algorithm for the fly's wiring in another world.

The ledger: a comparison

For the same narrow task of "find your way, don't hit walls":

| | Fly-inspired engineering | End-to-end large models / RL | |---|---|---| | Trainable parameters | 64,000 (mushroom body + central complex navigation agent) | millions and up | | Pretraining frames | 0 (first-trial SR 0.84 / SPL 0.48) | 10⁷–10⁹ frames | | Navigation power | 18.6 μW (ring attractor on Loihi) | ≥ 2 W (SLAM-dedicated FPGA) |

Sources: Lu & Webb, arXiv:2601.16806 (Edinburgh, embodied agent on Habitat point-goal) [A: preprint]; Loihi ring-attractor power estimate [A: PMC8881593].

Disclosed limitations: Lu & Webb's cross-trial memory consolidation uses a gating rule depending on shortest path length (which the agent shouldn't know—questioned [D]), and that rule gets stuck in local minima in deterministic simulation.

On the other side: multimodal LLM spatial reasoning (VSI-Bench)—humans 79.2, best GPT-5 only 55.0, early Gemini-1.5 Pro 45.4 [A]. A 2024 claim that models were "below random" no longer holds; the real weakness is perspective-taking and integrated reasoning.

Already on real hardware: a fly LPLC2 visual pathway model embedded in the microrobot Colias achieves 96.1% collision detection with 70 KB of memory [A: arXiv:2509.13827]; locust LGMD collision-avoidance networks have been flight-validated on drones (EU STEP2DYNA) [B].

These are pre-connectomics bio-inspired designs—circuit principles, not full maps—but the direction agrees: biology offers not data, but architectural priors.

Verdict: three conclusions and a better question

How much of "it" is inside? About half: all of the wiring, half of the strength, nearly zero of the current state.

1. Topology, counts, types, and predicted neurotransmitters are real and reusable (Shiu's 91% and shuffle ablation are hard proof); but weights, polarity details, neuromodulation, intrinsic properties, and learning state are absent—and the "weight" layer itself, in large part, never existed in any fly at any moment; it exists only in "generation rules + lived experience." 2. What you upload is the factory wiring diagram plus a pre-experience initialization—not the fly. The real "algorithm" has three layers: architectural priors (the connectome provides these, and well), plasticity rules (absent; written in genes and synaptic molecular machinery), and current weights (only a frozen snapshot). The Minecraft demos run layer one plus heavy manual patches; "responses survive weight shuffling" both discounts the demo and confirms some robustness lives in architecture. 3. The real AI value is not in moving but in copying. Deployed bio-inspired work (a 64K-parameter zero-pretraining navigator, an 18.6 μW compass, 70 KB collision avoidance) converges on one message: the valuable thing is not the list of 124 million synapses but circuit principles—ring attractors, lateral inhibition, sparse coding, mushroom-body-style "verify-then-store." The full connectome's significance is upgrading these principles from conjecture to verifiable blueprints.

Finally, upgrade the question. "How much algorithm is left inside" assumes algorithms are static and portable. Four lines of evidence point elsewhere:

> What the connectome preserves is not the algorithm, but half of the algorithm's compiled artifact— > the other half of the "source code" (plasticity rules) lives in the genomic bottleneck, never scanned. > Rather than "how much algorithm is in the map," ask: "how many algorithms can this map constrain?" > Shiu's answer: tightly. Lappalainen's answer: still more than one.

That is what this connectome milestone leaves for the AI era: not a fly living in Minecraft, but history's first synapse-by-synapse verifiable blueprint of complete perception–action circuits. A blueprint is not a building—but without blueprints, people couldn't even agree on how many beams a building needs.

Appendix: number caveats

| Figure | Denominator | Notes | |---|---|---| | 166,700 | annotated neurons (incl. sensory axons) | preprint: 166,691; 166,483 with synaptic edges | | 124.2M | synaptic contacts (one per contact point) | Google blog rounds to "125M"; UKRI says "124M" | | 25.6M | neuron-to-neuron directed edges | Minecraft projects use 25,582,938—most-often-confused with 124.2M | | 11,710 | cell types (final version) | preprint: 11,691 | | 40.1% | synaptic connections proofread at both ends | "complete" ≠ "fully proofread" |

Appendix: key sources by confidence tier

[A] Papers & official

• Berg et al. *Sexual dimorphism in the complete Drosophila male CNS connectome.* Cell 189(18):5504–5526 (2026). doi:10.1016/j.cell.2026.08.015 · preprint bioRxiv 2025.10.09.680999 · male-cns.janelia.org
• Google Research Blog: *A connectomics milestone* (2026-09-03) · UKRI / MRC LMB / Cambridge press releases (66 slabs, 7 EMs, 44 person-years)
• Dorkenwald et al. Nature 634:124–138 (2024) female full brain · Schlegel et al. 634:139–152 · Azevedo et al. 631:360–368 (FANC/VNC) · Winding et al. Science 379:eadd9330 (larva)
• Shiu et al. Nature 634:210–219 (2024) whole-brain LIF model · Eckstein et al. Cell 187:2574 (2024) transmitter prediction · Lappalainen et al. Nature 634:1132–1140 (2024) · Creamer, Leifer & Pillow, Nature (worm whole-brain activity)
• Liu & Jeanne, Curr Biol (2021) · Barnes & Cardona, PLoS ONE (2022) · Tobin et al. eLife (2017) · Hige & Turner, Neuron (2015) · Kiral et al. Cell Reports (2021) · Zador, Nat Commun (2019) genomic bottleneck
• Jonas & Kording, PLoS Comput Biol (2017) 6502 · Lechner et al. Nat Mach Intell 2:642 (2020) NCP · Lu & Webb arXiv:2601.16806 · PMC8881593 Loihi power · arXiv:2509.13827 LPLC2/Colias · arXiv:2412.14171 & 2508.13142 VSI-Bench
• NeuroCraft Fly: github.com/evnsnclr/neurocraft-fly-public (code unreleased) · DOOMFLY: github.com/nftechie/doomfly · Eon Systems: eon.systems/updates/embodied-brain-emulation

[B] Authoritative media: IBTimes UK · TweakTown (Doom, 6,385 episodes) · Ars Technica (2026-09-05) · Wired (2017) Mikula rebuttal · Curr Biol commentary (gap junctions unresolvable)

[C] Second-hand: GIGAZINE (Doom input mapping) · mcfun.cc / Yahoo Tech (Ro0oney used the FlyWire female model)

[D] Unverified (not used as evidence): "GPT-6 Astra involvement" (self-claim only) · thefrontrunners.io (speculative, crypto-marketing, rejected) · single-author preprint "Embodied Drosophila" (self-contradictory numbers, rejected)

--- *Deep research · four-way cross-verified · compiled 2026-09-15 · zhichai.net* *All figures carry denominators; uncertain claims are flagged in-text, and unverified rumors are never used as evidence.*