Did the Fly Brain Learn to Play Doom? What the Validation Gates Show
By
Samara Garcia
•
No. The fly brain connectome did not learn to play Doom, and the engineer who built the demo published the evidence against his own headline.
DOOMFLY wired the full MaleCNS v1.0 fly connectome into a live Doom arena in September 2026. It reached millions of views. Its README opens with this status line:
Status: live experimental training, not demonstrated learned survival. The current v6 candidate failed its visual, conditioning and survival validation gates.
The author, Coinbase engineer Alex Wormuth, ran three validation gates. All three failed. He published the raw results anyway, alongside the code.
Those failed experiments are more useful than a working demo would have been. They define a five-gate framework for testing any system that produces convincing output from complicated internal state.
What a connectome gives you, and what it leaves out
A connectome is a synapse-resolution wiring diagram. MaleCNS v1.0 maps 166,691 neurons across the central brain, both optic lobes, and the ventral nerve cord of an adult male Drosophila. FlyEM (HHMI Janelia), the University of Cambridge, the MRC Laboratory of Molecular Biology, and Google Research released v1.0 on June 8, 2026 under CC BY 4.0. Twenty-nine proofreaders spent three years on it, an estimated 44 person-years of effort. The Cell paper published September 3, 2026.
The dataset contains edges and synapse counts. It also contains neurotransmitter identity, predicted rather than measured. Eckstein et al. report 87% classifier accuracy per synapse and 94% per neuron overall, but the MaleCNS paper treats serotonin predictions as unclear in its recommended consensus annotation because validation data are comparatively sparse. No neuropeptides are predicted at all.
The dataset does not contain synaptic strength, receptor composition, neuromodulatory state, membrane biophysics, or plasticity.
That gap has one practical consequence. Every fly-brain demo had to invent its own dynamics.
DOOMFLY’s published implementation uses 0.1 ms neural-integration steps with a 35 Hz production game-frame cadence, mapping each ViZDoom frame onto 3,335 R1–R6 brightness inputs and 811 R8 color inputs. Turning and firing are driven by DNp20 and DNpe017. The repository describes the pixel mapping as “inferred proxies” and the control neurons as “engineered controller assignments, not established natural motor functions.”
The wiring is biological. The behavior is a model sitting on top of it. That distinction is why the controls carry the entire weight of the claim.
The conditioning experiment that looked like learning
DOOMFLY’s conditioning gate is a standard associative-learning design. A visual cue is paired with an artificial aversive signal: a 200 ms current injection into two PPL101 dopamine cells, triggered by nonfatal damage. In DOOMFLY’s conditioning model, the adapted plasticity rule operates on 4,184 existing KC→MBON11 connections while the rest of the wiring remains fixed.
Four arms ran, twice each. This is the published output from doom-ui/public/learning-iterations.json:
Conditioning, candidate v5. Development gate: false.
The README summarizes the current candidate as v6. The validation artifact is versioned per gate: conditioning ran on v5, survival on v6. No v6 conditioning replicate has been published.
Rep | Condition | Selectivity | Changed edges | Mean efficacy |
|---|---|---|---|---|
0 | paired | 0.889 | 3,629 | 1.671 |
0 | backward | 0.128 | 6 | 1.001 |
0 | frozen | 0.0 | 0 | 1.000 |
0 | no_imposed_US | 0.889 | 3,566 | 1.621 |
1 | paired | -0.765 | 7 | 0.999 |
1 | backward | 0.236 | 3,655 | 1.441 |
1 | frozen | 0.0 | 0 | 1.000 |
1 | no_imposed_US | -1.463 | 7 | 1.000 |
The no_imposed_US arm receives no punishment signal. It scored 0.889 selectivity, matching the paired condition to three decimals, while moving 3,566 synapses.
The dopamine signal was not driving the weight change. Cue-induced persistent activity was doing it alone.
Replicate 1 then inverted the sign on both headline conditions. Paired dropped to -0.765. The no-punishment arm dropped to -1.463.
The repository states the conclusion directly: “Cue-induced persistent activity and no-punishment effects invalidate the current conditioning interpretation.”
The survival gate closed it out:
Survival, candidate v6. Complete: true.
Condition | Mean survival (s) | Deaths | Changed edges |
|---|---|---|---|
plastic | 3.657 | 2/2 | 1,862 |
frozen | 5.829 | 1/2 | 0 |
shuffled | 3.657 | 2/2 | 1,896 |
Frozen weights outlived plastic weights by 60%. Plastic weights tied exactly with shuffled ones. The sample size was two episodes per group, which the author describes as small rather than treating the result as a finding.
Now remove two rows. Delete no_imposed_US. Delete replicate 1.
What remains is a paired condition at 0.889 against a frozen control at 0.0, with 3,629 synapses moving in response to punishment. That reads as successful aversive conditioning in a simulated connectome.
One control arm separates that claim from a retraction.
Five validation gates: how to validate an ML result
A demo becomes a result when it survives five gates. Each has a cheap version that runs in an afternoon.
Gate 1: Ablation study
Remove the component you are crediting. Hold everything else fixed. Measure.
Fly Dino executed this cleanly. The circuit is small and labeled as small: 80 cells, 1,296 directed edges, 26,029 contacts. Only a 243-parameter readout is trained, on top of a frozen graph.
100 held-out seeds (2100001 to 2100100).
Controller | Completed | Mean survival | Mean score |
|---|---|---|---|
connectome + trained readout | 179.372 s | 2885.75 | |
circuit silenced | 0/100 | 4.509 s | 41.00 |
untrained readout | 0/100 | 4.492 s | 41.00 |
handwritten rule baseline | 0/100 | 46.493 s | 535.27 |
random actions (30 Hz) | 0/100 | 4.662 s | 42.77 |
idle | 0/100 | 4.509 s | 41.00 |
Silencing the circuit while keeping the identical trained readout drops completion from 99 to zero. Performance falls to the idle controller. That establishes a causal path through the biological graph for the cost of one extra run.
The author still declines the larger claim. The work “does not test whether biological topology is better than equally sized artificial or rewired network.”
Gate 2: Absent-mechanism control
Run the arm where your proposed mechanism is absent. Confirm the effect disappears with it.
This is the gate DOOMFLY failed, and the one most projects skip. A frozen control already exists, so a second control feels redundant.
It is not. Frozen tests whether plasticity is switched on. The absent-mechanism arm tests whether your specific mechanism is what drives it. Those are different questions. Selectivity of 0.889 in both arms is what the difference looks like when it goes wrong.
The rule generalizes. Any time you attribute behavior to a specific cause (a reward term, a retrieval step, a prompt section, or an index), this is the arm you are missing.
Gate 3: Shuffle control
Break the structure while preserving its statistics. If performance holds, the structure was not the cause.
DOOMFLY ran this one. Shuffled weights tied plastic weights at 3.657 seconds.
FlyDoom is built around it. The project ships matched Erdős–Rényi random graphs and degree-preserving rewiring alongside the real connectome. Its README sets the standard: “a null or negative result is a valid answer; this project does not assume that biological topology helps.”
Outside connectomics, shuffle your labels, embeddings, graph edges, or retrieved chunks. A number that survives the shuffle describes your pipeline, not your data.
Gate 4: Budget parity comparison
Compare only at equal training budget, equal parameter count, and equal data.
Fly Dino names its own violation here. A topology-benefit claim “would require matched artificial/rewired controls trained with equal budgets; that study is not included.”
This gate invalidates most benchmark comparisons published informally. Two models trained to different budgets on one chart are two unrelated runs.
Gate 5: Replication and negative results
Run it again on a different seed before publishing.
DOOMFLY’s second replicate inverted the sign on both headline conditions. Posting after replicate 0 would have produced a wrong claim supported by real evidence.
A single run agreeing with your hypothesis is the most expensive result available, because it is the one that stops further testing.
How the viral demos score
Project | Ablation | Absent mech. | Shuffle | Budget parity | Replication |
|---|---|---|---|---|---|
DOOMFLY | Not reported | Ran, failed | Ran | Not reported | Ran, failed |
Fly Dino | Passed | Not reported | Not reported | Declined, stated | |
FlyDoom | Not reported | Not reported | By design | By design | Not reported |
Beat Saber demo | Not reported | Not reported | Not reported | Not reported | Not reported |
NeuroCraft (Minecraft) | Not reported | Not reported | Not reported | Not reported | Not reported |
Note: “Not reported” means the project published no result for that gate.
The Beat Saber fly-brain demo became one of the most widely circulated examples in the September 2026 connectome-demo wave. The creator later clarified that the Beat Saber system was overfit to one track and still relied on replay input, but was not simply executing a fixed prerecorded movement sequence.
The Minecraft project states in its README that neural readouts “select and modulate scripted body programs.” The visible motion is authored animation triggered by network activity.
Jessica Paquette was explicit about Fly64, her Super Mario 64 version: “There is no training, reward, or goal to collect stars. Mario can walk into a wall and stay there.”
Every one of these builders documented the limits accurately in their own repository. The distortion happened downstream, in coverage that read the demo and skipped the README. That is an argument for putting caveats in the status line rather than a separate limitations file.
Why the build wave still mattered
The dataset shipped under CC BY 4.0 with a downloadable graph and a hosted Neo4j instance. The awesome-fruit-fly index catalogued roughly 80 fly-connectome projects in the weeks after release, along with starter scaffolding for new ones.
Three habits appeared across that corpus that usually require years of institutional pressure.
Provenance pinning. DOOMFLY pins SHA-256 hashes for three MaleCNS source files in source.lock.json and raises an error if a downloaded file does not match its recorded digest. Fly Dino rejects unexpected source hashes in its data builder. Unrelated projects converged on dataset checksum verification independently, inside a week.
Disclosure of AI assistance. Paquette recorded hers in the README: "this code is literally 100% vibe coded with GPT Astra, and I did it just for fun. I have not reviewed the code." Other builders did the same. Stating how code was produced now functions as a credibility signal.
Negative results as a shipped artifact. DOOMFLY published failed validation results, while other projects such as FlyDoom explicitly treated null or negative findings as valid experimental outcomes.
Building was the correct instinct. Current tooling produces a connectome-driven game loop over a weekend, and roughly 80 projects demonstrated that.
Evaluation did not get cheaper. It got harder, because there is more output to assess and the output looks better than it is.
What this measures in hiring
Shipping a demo stopped being a differentiator in September, when 80 projects appeared in nine days.
Knowing which of your results survives Gate 2 is still a differentiator. So is publishing the run that killed your hypothesis.
Neither is visible on a resume. A take-home will not surface either. Both are obvious in artifacts: an ablation nobody requested, a replicate that inverted, a status line reporting failed gates.
Fonzi is a curated AI engineering marketplace that evaluates engineers on demonstrated work and connects them to vetted VC-backed startups and high-growth tech companies through recurring Match Day cycles. It is free for engineers, and assessment runs on what you can show.
If your last project included the experiment that could have killed it, that is the artifact worth putting in front of a hiring team. Match Day opens every Monday. Companies have 72 hours to send salary-backed interview requests, and you then have 24 hours to decide which to take.
FAQ
Did the simulated fly brain learn to play Doom?
What is the MaleCNS v1.0 connectome?
Can a connectome be used to simulate a brain?
What is an absent-mechanism control?
