A fruit fly's entire central nervous system is now a download. One team plugged it into Doom and watched it fail thousands of times without learning anything. Another team used the same 166,700 neurons as a language model's recurrent backbone and scored 88.45 test perplexity, against 124.89 for a rewired copy of the same graph. Both results are public and reproducible. They cannot both be about the same thing.
The wiring diagram is MaleCNS v1.0 — 166,700 neurons, 25,582,938 directed connections, 124,177,617 synaptic contacts, released under CC-BY 4.0 in June 2026 and published in Cell that September. Within weeks engineers had pointed it at Doom, Super Mario 64, a Strandbeest robot, a simulated traffic street, and a GPU in a browser tab. What nobody had agreed on was whether the topology is worth copying.
The wiring diagram everyone downloaded
The Cell paper is the provenance. Berg and colleagues at HHMI Janelia, Google Research, and the MRC Laboratory of Molecular Biology produced the first complete connectome of an adult male Drosophila central nervous system: brain, both optic lobes, and the ventral nerve cord, fully proofread and annotated across 11,710 neuron types, with the neck connective intact. That last part matters more than the headline number. A brain-only map tells you what a fly can perceive. A brain-and-nerve-cord map tells you what it can do.
The dataset is unusually clean for its size. The team labels weak edges as noise below 11 synapses, which removes 80% of connections while keeping roughly 90% of all synapses in the strong set. The comparison with the 2024 female FlyWire map also produced the first brain-wide sexual dimorphism census in an adult animal: 8,069 isomorphic neuron types, 138 dimorphic, 289 male-specific, 71 female-specific.
Then the internet did what the internet does. Within days of the September announcement there were Minecraft builds, Doom agents, music videos, and a browser tab running the whole connectome as a spiking network on a consumer GPU. The pattern that emerged is worth stating plainly, because it is the reason this is a live dispute rather than a settled result: a wiring diagram is now a programmable object you can download, wire into anything, and draw conclusions from without training on biology at all.
MaleCNS is a reconstruction of one adult male specimen — no glia, no vasculature, no neuromodulatory state, no development. Every synapse count is a static edge weight. It is a superb map of structure and a poor model of a living nervous system. Every claim below is about what the graph does as an architecture, not about what the animal does.
Pole A: the topology is the prior
The strongest version of the optimistic case arrives in arXiv:2610.10014, from October 2026. The authors take the MaleCNS adjacency as the fixed recurrent mask of an artificial network, learn one scalar per anatomical edge, and train two models: bounded addition, and a grounded relational language task built from a fixed 100-word lexicon. Rewired controls are generated by directed double-edge swaps that preserve every node's in-degree and out-degree, retain self-loops, and run five accepted swaps per source edge — after which only 0.588–0.594% of original edges survive.
The anatomical graph reaches 92.77% mean accuracy on held-out addition against 67.93% for degree-preserving rewires. On the strict paired language endpoint it scores 61.59% across four input-output interfaces against 44.17% for matched rewires, a 17.42 percentage-point gap with a 95% crossed-bootstrap interval of [11.39, 24.90]. The authors do not stop at four graphs. At the canonical interface the anatomical graph scores 71.35%, twenty independent rewires average 46.56% with a standard deviation of 6.99 points, none of them reaches the anatomical score, and the finite-ensemble upper-tail rank fraction lands at 1/21.
A parameter-matched dense recurrent control — 5,058 units, same codebooks, same four-million-token budget — does not close the gap. Neither does a sensory-shuffle intervention: replacing the task-defined sensory amplitudes drops the anatomical model from 60.94% to 19.27% while the rewire mean falls from 41.80% to 13.02%. Two models trained with sensory input disabled score zero, which is the sanity check a lot of this literature skips.
FlyGM, in arXiv:2602.17997, points the same way on a physical control task. Instantiating the adult fly connectome as a directed message-passing graph and training it with imitation learning plus PPO produces an angle error of 8.29 at high yaw where degree-preserving rewiring reaches 13.55 and an Erdős–Rényi graph degenerates to 125.36. The spiking-neural-network baseline never learns a usable gait at all — biological plausibility at the single-neuron level does not substitute for a connectome-level structural prior.
Even the thinner results lean the same way. A character-level model on 16,384 retained MaleCNS neurons scores 88.45 test perplexity against 124.89 for an edge-matched shuffle, and a prespecified six-seed replication has real wiring beating its degree-preserving control in all six pairs, with a two-sided exact probability of 0.03125.
Pole B: the topology is furniture
The skeptical case does not argue the numbers are wrong. It argues the comparison is not isolating topology.
arXiv:2604.04033 builds what its authors call a control ladder. Starting from an observation that looks strong — a connectome-constrained model with lower early loss, lower mean activity, and faster runtime against a naive self-loop-matched random graph — they remove two confounds one at a time. At five steps the checkpoint-initialized comparison reads 0.514 loss versus 0.698, and 0.656 mean activity versus 1.861. Then they train both graphs from a shared from-scratch initialization. The loss advantage becomes −0.0020. Replace the naive null with a degree-preserving rewire and the activity advantage reverses sign to −0.0106. A five-sample degree-preserving ensemble holds the revised reading.
The game side of the ledger is more bluntly empirical. The untrained whole-graph DOOMFLY simulation retains all 166,700 neurons and 25,582,938 directed edge rows, reads game frames as 3,335 brightness and 811 colour retinal signals, maps DNp20 right-minus-left activity to view rotation and summed DNpe017 activity to movement and firing, and schedules a 200 ms aversive input into two PPL101 dopamine cells on nonfatal damage. Plasticity touches 4,184 existing KC→MBON11 connections. It still fails its visual, conditioning, and survival gates. Scramble the 25.6 million connections and it dies in 7 seconds instead of 53 — the wiring is plainly load-bearing.
Then a second group ran the control nobody else ran. One hundred seed-matched games per arm, all five ViZDoom scenarios. Real wiring, rewired wiring, and no wiring at all score 19.1, 19.6, and 19.3 kills per game on defend_the_center. Scrambling the trained network's edges after training does drop it to near-random, so the network genuinely uses the connections. It just does not matter which ones it was trained through. And the output is nearly constant: turn slightly right, walk forward, keep shooting. A spiking autopilot with no brain whatsoever survives 52 seconds against the fly's 53, on 12 kills against 12.
A separate PPO distillation project reaches the same place from the other side. Real connectome, shuffled edges, no connectome — 18.6 ± 1.8, 19.0, 18.4, 18.3 on defend_the_center, with the connectome version training 22 times slower: 646 steps per second against 14,686 for the parameter-free backbone. A two-sided result that says the wiring costs you compute and returns nothing.
Neither camp is alleging fraud. They are reporting different numbers from different null models, and the gap between them is entirely methodological.
✔ What the pro camp establishes
+ The advantage survives 21 graph draws, four token interfaces, and a parameter-matched dense control
+ It holds on two different task families — arithmetic and grounded language — fitted independently
+ A third group finds pre-registered, dose-matched evidence that real wiring coordinates antagonistic motor pools better than all 30 rewired networks at 19 of 20 neuron pairs
Load-bearing condition: the null must be degree-preserving and initialization must be topology-agnostic. Weaken either and the effect mostly disappears.
✗ What the contra camp establishes
− A self-loop-matched random null is not a null for directed degree, and it manufactures most of the headline gap
− 100 seed-matched game runs per arm: real, rewired, and no wiring all score within noise of each other
− The connectome backbone trains 22× slower than a parameter-free alternative for identical scores
− Passing software tests is not evidence of biological validity, and a 6,385th failed attempt is a negative result, not progress
Falsifier: a topology advantage that survives shared from-scratch initialization and a 20-graph degree-preserving ensemble on a task where the wiring is not the bottleneck.
The camps are measuring different graphs
Here is the part that survives both camps. A degree-preserving rewire is not a neutral substrate. It preserves every node's in-degree and out-degree, and in doing so it preserves reachability — the giant strongly connected core and the typical token-to-output distance stay almost unchanged. What it destroys is reciprocity and higher-order motif structure, and it systematically shortens sensory-to-output paths. The pro camp's own paper says this plainly: shared reachability with divergent higher-order statistics.
So the pro results are not a claim that "biology beats random." They are a claim that higher-order connectome statistics beat a null that has had precisely those statistics removed. That is a real and interesting claim. It is also a much narrower one than the headlines imply, and the contra results do not contradict it — they contradict a weaker experiment that was run first.
What the contra results do establish is that in the regime most people actually care about — a game, a driving task, a locomotion task with a learnable policy on top — the adjacency is not where the performance lives. Read the doomfly-rl parameter counts closely. The frozen recurrent core is 48.2 million parameters, but the convolutional "eye" feeding it and the MLP "hand" reading it out together hold more than half the model's capacity. Rewire the recurrent core and you change which subgraph the encoder talks to. You do not change the encoder. In arXiv:2610.10014 the analogous situation is inverted: the recurrent mask is the model, every trainable scalar sits on an anatomical edge, and the task is small enough that the backbone cannot be bypassed.
Different regimes, different answers, both honest. The error is in reading either result as a statement about flies.
| Result | Anatomical wiring | Degree-preserving control | Reported by |
|---|---|---|---|
| Held-out addition | ✔ 92.77% | ✗ 67.93% | arXiv:2610.10014 |
| Grounded language, 4 interfaces | ✔ 61.59% | ✗ 44.17% | arXiv:2610.10014 |
| Fly locomotion, high yaw | ✔ 8.29 rad/s | ✗ 13.55 rad/s | arXiv:2602.17997 |
| Early loss, from-scratch | ◐ −0.0020 gap | ◐ no difference | arXiv:2604.04033 |
| Kills per game | ◐ 19.1 | ✗ 19.6 — and 19.3 with no wiring | doom-fly-control |
| Training throughput | ✗ 646 steps/s | ✔ 14,686 steps/s without wiring | doomfly-rl |
Best available figures as published. Each row uses a different task, task size, and null construction; read down the control column, not across the rows. Compiled from arXiv:2610.10014, arXiv:2602.17997, arXiv:2604.04033, and the two open control repositories.
There is a third line of work that changes the shape of the question. In arXiv:2609.38665, Zhao and colleagues test which features of motor output depend on the specific wiring, using pre-registered criteria and replication across two independently reconstructed connectomes. Rhythm turns out to be generic: several rewired networks oscillate more readily than the real ones. Antagonistic coordination does not. The real wiring coordinated opposing motor pools more strongly than every rewired network, at 19 of 20 neuron pairs in MaleCNS, and the effect traces to how premotor input is allocated between those pools rather than to input strength.
Coordination, not rhythm, is what reveals whether a connectome model's wiring matters. That is a sharper instrument than any game score, and it is the one to watch.
Every published pro result used degree-preserving rewiring. Every published contra result used a task where the encoder and decoder carried most of the capacity.
The unresolved question is whether the advantage survives both conditions at once: from-scratch initialization, a 20-graph degree-preserving ensemble, and a task small enough that the wiring cannot be bypassed.
What would settle it
Three experiments, none of which is hard. Run the arXiv:2610.10014 addition and language tasks with shared from-scratch initialization instead of a topology-aware one. Run them again against an Erdős–Rényi null with matched in- and out-degree, not just matched edge count. And run the Doom control arms on a task where the policy has to fit through the recurrent core rather than around it. If the 92.77% holds under all three, the connectome is an architectural prior and the 2026 skeptics were measuring the wrong thing.
If it does not hold, the honest conclusion is narrower and still worth having: evolution optimized this graph for a body weighing a milligram and acting in milliseconds, and that solution does not transfer to problems where data and gradient descent are the binding constraint. The most useful output of MaleCNS v1.0 may be diagnostic rather than constructive — a fully annotated ground truth for when a recurrent topology is doing real work and when it is expensive decoration.
As we wrote in July 2026, the fly upload was the threshold moment. The wiring diagram that followed it turned out to be the more interesting artifact, precisely because it could be argued with.