Putting a Fruit Fly Brain Into a Computer: How Much of "It" Remains?

In early September 2026, teams from Janelia and Cambridge 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 edges.

Within a week, community members put it into Minecraft, Doom, and Beat Saber.

The caption on every repost was roughly the same: today's AI is still learning to understand the physical world, while natural selection has been "training" biological neural networks in the real world for 500 million years. So the question arises — how much of this animal's own "algorithms" is already loaded into those scanned wires?

I broke this into four layers and cross-verified each with independent searches. The conclusion up front:

> A connectome is a netlist, not firmware, let alone an operating system. It contains half of the compiled artifact of the algorithms — the other half of the source code (plasticity rules) is written in the genomic bottleneck and was never scanned.

All numbers are tagged with provenance: A = paper/official, B = authoritative media, C = secondary reporting, D = unverified.

1. Three Facts First

First, this was not the FlyWire team. The 2024 Nature papers completed the female fly brain (139,255 neurons, 54.5 million synapses; Princeton's Murthy/Seung). This male version comes from Janelia FlyEM + MRC LMB/Cambridge (Jefferis & Rubin): Berg et al., Cell 189(18):5504–5526, published 2026-09-03, open access.

Second, coverage is the entire central nervous system: brain + bilateral optic lobes + ventral nerve cord, traceable continuously from eye to leg tip.

Third, the numbers depend on definitions. The "125 million synapses" in the news corresponds to 124.2M in the text; the 25,582,938 run in the Minecraft project is the count of "edges between neuron pairs" — if the same pair touches a hundred times, it counts once. A 5x difference, same fly.

The production pipeline is worth noting: the tissue was cut into 66 blocks, 7 electron microscopes ran for a year at 8nm resolution, Google's flood-filling networks did segmentation, and 44 person-years of manual proofreading were required. But note — only 40.1% of synapse connections have both ends proofread. A "complete graph" does not mean "every line manually verified."

The real scientific advance in the male version is sexual dimorphism: 95% of cell types are shared between sexes, but in the male brain 12% of neurons are wired differently by sex (only 4% in females), with roughly 100 male-specific interneurons supporting courtship behavior. Differences in wiring are differences in behavior.

2. The Minecraft Fly: How Real Is It?

I checked primary sources one by one. The projects are real, but far weaker than the legend.

• NeuroCraft Fly (Minecraft): a real GitHub repo and demo video exist, but the README states the code has not been released. The author admits the sensory mapping is modeled, the readout (which neurons count as "forward") was hand-picked, and movement is a scripted body program.
• DOOMFLY: fully open source. The README says: "live experimental training, not demonstrated learned survival" — 6,385 episodes run, and none of the three gates (vision, conditioning, survival) were passed.
• Beat Saber version: the author clarified the motor system is overfit playback of recorded sequences. The connectome wasn't keeping the beat; a tape was playing.

The most interesting admission is from the NeuroCraft Fly author: "Responses persist after weight shuffling — a finding that needs reporting."

Shuffle all 124 million synapses randomly, and the "fly" still responds in Minecraft. Two readings:

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

The more serious contemporaneous work is Eon Systems' embodied whole-brain simulation (March 2026): an LIF whole-brain model with a MuJoCo physics body that can approach food and groom. But they themselves state: no spontaneous neuromodulation, no learning, and visual input "is essentially decorative." The peer-reviewed, quantitatively validated foundation is the non-embodied whole-brain model of Shiu et al. 2024, Nature.

3. You Get a Netlist, Not Firmware

What do you actually get when you move 124.2M synapses into a computer?

You get: topology (who connects to whom), connection counts, 11,710 cell types, and predicted neurotransmitters (a CNN infers from EM images, 87% single-synapse accuracy).

You don't get — and must supply yourself:

• Synaptic weights — contact count ≠ strength. Release probability, receptor density, and short-term plasticity are not in the graph;
• Excitation/inhibition — EM cannot see the sign; it is inferred from neurotransmitters, with ~13% errors;
• Neuromodulation — dopamine, serotonin, octopamine. The same circuit, different states, different outputs;
• Electrical synapses — gap junctions are too small, "unresolvable in connectomics imaging";
• Intrinsic properties — thresholds, time constants, refractory periods: "single-neuron and single-synapse properties are unknown";
• The body — leg length, wing mass, all outside the brain.

You hold the chip's circuit diagram and bill of materials. Getting the chip to boot still requires component parameters, power delivery, and the burned-in program — and that last layer may not be in the chip at all.

A rarely mentioned trap: individual variability. Comparing two female fly connectomes, only connections with ">10 synapses or >1% of the target cell's input" count as conserved; about a third of cell types cannot be reliably re-identified across brains. Even with identical genomes, flies raised at 18°C vs 25°C develop different synapse pairings (the paper's title is literally "The limits of genetic determinism").

4. Structure vs Dynamics

The strongest evidence from both camps:

Structure camp (the connectome carries real information):

• Shiu 2024 (Nature): a 127,400-neuron whole-brain LIF model, weights = synapse count × sign, with only 1 free parameter. 164 testable predictions, 91% consistent with experiments.
• The killer ablation: the same model, after shuffling synaptic weights, accuracy drops from 100% to 1%. The real wiring is what works.
• Creamer et al. (Nature): a connectome-constrained model fit whole-animal optogenetic data of C. elegans to 92% of the data's own reproducibility ceiling.

Dynamics camp (behavior lies beyond structure):

• Lappalainen 2024 (Nature): the same visual connectome, 50 trained networks, different mechanisms emerge — one cluster of T4c cells completely loses direction selectivity. One graph, multiple algorithms.
• The worm lesson: modeling directly with anatomical weights "fails to produce correct behavior"; weights must be re-learned.
• The harshest datum: after a single associative learning event in the fly, the relevant synapse strengths drop by 80% on average. One experience rewrites 80% — consider the value of the "weights" in a static graph.

The two camps are not contradictory; together they form the answer: the connectome's information content is enough to constrain "what the circuit computes," not enough to determine "how the circuit is currently computing it."

A correction to a widespread myth: the claim that "Lechner constrained an RNN with the worm connectome and found topology, not weights, does the work" is false. The actual paper is *Neural circuit policies enabling auditable autonomy* (Nat Mach Intell 2020); its weights are entirely end-to-end trained, and no "topology vs weights" ablation appears in it. The correct citation for that claim is Shiu's shuffling experiment.

5. What Did Natural Selection Actually "Train"?

The answer is Zador's *A critique of pure learning* (2019): the genomic bottleneck. The wiring diagram is far too complex to be written gene by gene — evolution does not store a weight table; it stores a highly compressed set of wiring rules.

So the connectome you scanned back is neither "trained weights" nor "random initialization." It is — a snapshot of architecture + strong initialization + a set of plasticity rules.

Three pieces of evidence:

1. Learning instantly rewrites what the connectome doesn't give you (the 80% above); 2. The same genome grows different graphs (Kiral 2021); 3. Even the most "innate" circuits only provide scaffolding — the fly's head-direction cells (ring attractors) are the closest thing to a "hard-coded navigator," but their correspondence to external landmarks is arbitrary; every new environment requires plasticity to recalibrate the compass. Even this hardest bone: the graph gives you the axle, not the reading.

And an overlooked third party: the body. Plugging fly circuits into Minecraft swaps the biggest hidden variable. Sensory circuit statistics and body geometry co-evolve — remove the wings and the mechanics of 87 joints, and the input-output distributions all change. The Minecraft demos' "modeled sensing, hand-picked readouts" is not a detail; it is the essence: that machine runs not the fly's algorithm in the physical world, but a zeroth draft of a new algorithm in another world.

6. Fly vs Large Models

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

• Trainable parameters: insect-inspired navigation agents use 64,000 (mushroom body + central complex), vs millions for end-to-end RL baselines;
• Pre-training frames: 0. First-episode SR 0.84 / SPL 0.48 (Habitat point-goal benchmark; Lu & Webb, Edinburgh, 2026 preprint). Baselines need 10⁷–10⁹ frames;
• Power: ring-attractor networks on Loihi run at ~18.6 μW, vs ≥2 W for SLAM-dedicated FPGAs — five orders of magnitude.

On the other side, LLM spatial reasoning (VSI-Bench): humans 79.2, the best GPT-5 only 55.0. (A 2024 claim that models were "below random" no longer holds; 2025 re-evaluations show near-human size estimation, with the real weakness in perspective-taking.)

On real hardware: an fly LPLC2 visual pathway model embedded in a micro-robot achieves 96.1% collision detection with 70 KB of memory.

Note the scope: this comparison holds only for "navigation," not general intelligence. But it makes a hard point — what biology gives AI is never data; it is architectural priors.

7. Verdict

How much of "it" is inside? About half: all of the wiring, half of the strength, almost none of the current state.

1. Topology, counts, types, and predicted neurotransmitters are real and reusable — Shiu's 91% and the shuffling experiment are hard proof; 2. Weights, polarity details, modulation, intrinsic properties, and learned state are not in the graph — and much of that "weights" layer never exists in any fly at any moment; it exists only in "generation rules + life experience"; 3. What you move into the computer is the factory circuit diagram + pre-experience initialization, not the fly. The Minecraft demos run architecture plus many manual patches; "responses persist after shuffling" discounts the demos — while also showing some robustness genuinely lives in the architecture; 4. The real value for AI is not "moving" but "copying." What makes deployed bio-inspired engineering valuable (64K-parameter zero-shot navigation, 18.6 μW compass, 70KB collision avoidance) is not the synapse list but circuit principles: ring attractors, lateral inhibition, sparse coding, mushroom-body-style "verify before storing." The complete connectome's meaning is upgrading these principles from "hypotheses" to "verifiable blueprints."

Finally, upgrade the question. "How many algorithms are already inside" presumes algorithms are static and portable. The evidence points to a better question:

> Instead of asking "how many algorithms are in the graph," ask "how many algorithms can this graph constrain." > Shiu's answer: very tightly. Lappalainen's answer: still more than one.

What this milestone truly leaves behind is not a fly living in Minecraft, but the first-ever synapse-by-synapse verifiable complete blueprint of perception-action circuits. A blueprint is not a building — but in the era without blueprints, people couldn't even agree on how many beams there should be.

---

Key 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 · Eon Systems technical page.