Did the Fly Brain Learn to Play Doom? What the Validation Gates Show

No. The fruit fly connectome did not learn to play Doom. DOOMFLY wired the MaleCNS v1.0 connectome, a cell-by-cell wiring map of a fruit fly's brain, into a live Doom arena in September 2026, and the demo went viral on a claim that the fly brain had learned to survive the game. The repository says otherwise: its own status line reports that the current 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.

Key Takeaways

• DOOMFLY's fly-brain connectome did not learn to play Doom: its own published validation data show plastic (learning-enabled) weights performing no better than shuffled, randomized weights, and worse than frozen ones.
• A working demo and a validated result are different things. Five validation gates separate the two: ablation, an absent-mechanism control, shuffle control, budget parity, and replication.
• The differentiator in hiring isn't whether someone shipped an impressive demo; dozens of engineers did that in the same nine-day window. It's whether they ran the control that could have killed their own hypothesis and published it anyway.

What the MaleCNS connectome gives you, and what it leaves out

A connectome is a synapse-resolution wiring diagram, essentially a complete map of which neurons connect to which, and how strongly, down to the level of individual synapses. MaleCNS v1.0 maps 166,700 neurons across the central brain, both optic lobes, and the ventral nerve cord (the fly's equivalent of a spinal 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 was published September 3, 2026.

The map itself is extremely detailed, but it's still missing pieces. It records which neurons are wired together and roughly how many synapses connect them. It also records a best guess at which chemical each neuron uses to signal (its neurotransmitter), though that's predicted by a classifier rather than directly measured. Eckstein et al. report 87% classifier accuracy per synapse and 94% per neuron on the FAFB whole-brain split, falling to 78% and 91% on HemiBrain. The MaleCNS paper itself flags one neurotransmitter, serotonin, as unreliable in its predictions because there isn't enough validation data to trust it. A separate class of signaling molecules, neuropeptides, isn't predicted at all.

What the map doesn't include is anything about how strong each connection actually is in a living fly, what receptors are present, the brain's chemical state at any given moment, the physical properties of the neurons' membranes, or how connections change over time as the fly learns. In short: the map shows you the circuit board, not the electricity running through it. That gap has one practical consequence. Every fly-brain demo had to invent its own version of "how the brain actually behaves," since the data alone doesn't specify it.

DOOMFLY's version of that invented behavior works like this: it runs the simulated brain in very fast, very short time steps (0.1 milliseconds each) to keep up with the game's frame rate, and it maps what the game's camera sees onto the fly's simulated eyes, thousands of individual light and color sensors standing in for pixels on screen. Two specific "descending neurons," essentially output wires that in a real fly would help control movement, were assigned to turning and firing in the game. The repository is upfront that this is invented, describing the pixel mapping as "inferred proxies" and the control-neuron assignments as "engineered controller assignments, not established natural motor functions."

The wiring is biological and real. The behavior layered on top of it is a model the engineer built. That distinction is why the controls carry the entire weight of the claim.

DOOMFLY's 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 dopamine-releasing neurons (PPL101, which in a real fly helps signal punishment), triggered by nonfatal damage. The learning rule operates on 4,184 connections from Kenyon cells to MBON11, one mushroom body output neuron within the fly's broader memory circuitry.

Four arms ran, twice each. This is the published output from doom-ui/public/learning-iterations.json:

Conditioning, candidate v5. Development gate: false.

Selectivity is the signed difference between the two cue-suppression values, oriented to the designated conditioned cue: positive means the circuit discriminated in favour of that cue, negative means against it. Mean efficacy is the average modeled efficacy across the 4,184 plastic edges, where 1.0 is the unchanged baseline.

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.

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.

It gets worse for the claim on the second run. Replicate 1 flipped the sign on both of those same conditions. The paired arm, which had scored a strongly positive 0.889, dropped to -0.765. The no-punishment arm dropped to -1.463. A result that flips sign between two runs of the identical experiment is not a stable finding. The repository states the conclusion directly: "Cue-induced persistent activity and no-punishment effects invalidate the current conditioning interpretation."

The survival gate, a separate test of whether the "learning" actually helped the simulated fly survive longer in the game, closed the case:

Survival, candidate v6. Complete: true.

Frozen weights outlived plastic weights by about 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 keep only replicate 0's paired and frozen arms.

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 validated result when it survives five gates. None of them require special resources; each has a cheap version that runs in an afternoon.

Gate 1: Ablation study

Remove the component you're crediting for the result, hold everything else fixed, and measure what happens. Fly Dino executed this cleanly. Its published experiment documentation describes a compact circuit of 80 cells, 1,296 directed connections, and 26,029 synaptic contacts. Only a compact, 243-parameter readout layer is trained on top of that frozen, unchanged graph.

Here's what happened across 100 held-out random seeds:

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.

Even so, the author declines to overreach with the finding. The author also notes that the experiment does not test whether biological topology outperforms equally sized artificial networks or rewired networks, a genuinely different and harder claim that this experiment wasn't designed to answer.

Gate 2: Absent-mechanism control

Run the arm where your proposed cause is absent, and confirm the effect disappears with it.

This is the gate DOOMFLY failed, and the one most projects skip, since a frozen control already exists and a second one can feel redundant. It isn't: frozen only tests whether learning is switched on, while the absent-mechanism arm tests whether the thing you're crediting, here, the punishment signal, is actually driving it. DOOMFLY's identical 0.889 score in both the punished and unpunished arms is what it looks like when the answer is no.

The rule generalizes: any time you credit a specific cause, a reward term, a retrieval step, a prompt section, an index, this is the control you're missing.

Gate 3: Shuffle control

Scramble the structure while keeping its statistics intact. If performance holds up anyway, the structure wasn't the cause. DOOMFLY ran this one: shuffled weights tied plastic (learning-enabled) weights exactly, both at 3.657 seconds.

FlyDoom builds around this idea, shipping matched Erdős–Rényi random graphs and degree-preserving rewiring alongside the real connectome. Its README states the standard plainly: "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 gap here: a topology-benefit claim "would require matched artificial/rewired controls trained with equal budgets; that study is not included."

This gate invalidates most informally published benchmark comparisons. 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 results. Posting after the first run would have produced a wrong claim backed by real evidence. A single run agreeing with your hypothesis is the most expensive result available, because it's the one that stops further testing.

How the fly brain Doom demos score on the five gates

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. Lyra Bubbles (@lyraaaa) later described the Beat Saber system as overfit to one track and still using replay or teacher-forced input, while saying its motion was not simply a fixed prerecorded movement sequence.

NeuroCraft Fly, created by Evan Smith, states in its README that hand-chosen 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's an argument for putting caveats in the status line rather than a separate limitations file.

Why the fruit fly connectome build wave still mattered

The dataset shipped under CC BY 4.0 with a downloadable graph and a hosted Neo4j instance. By September, community indexes such as the awesome-fruit-fly and awesome-fly were cataloguing a growing collection of MaleCNS connectome projects, and awesome-fly also linked starter scaffolding for new experiments.

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, within 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 didn't get cheaper. It got harder, because there's more output to assess and the output looks better than it is.

What validation rigor measures in hiring

Shipping a demo quickly became less distinctive during the September build wave as community indexes filled with fly-connectome experiments. Knowing which of your results survives Gate 2 is still a differentiator. So is publishing the run that killed your hypothesis.

Neither shows up on your resume. A take-home won't 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 engineering marketplace that evaluates AI and software engineers on demonstrated work and connects them to vetted VC-backed startups and high-growth tech companies through recurring Match Day cycles, and it's free for engineers to use. Assessment runs on what you can show: Match Day opens every Monday, companies have 72 hours to send salary-backed interview requests.

If your last project included the experiment that could have killed it, that's the artifact worth putting in front of a hiring team.

What the DOOMFLY validation gates show

The headline claim, a fly brain learned to play Doom, doesn't survive DOOMFLY's own data. Plastic weights survived 3.657 seconds against 5.829 seconds for frozen weights, tied exactly with shuffled weights, and the conditioning result flipped sign on a second run. The five-gate framework (ablation, absent-mechanism control, shuffle control, budget parity, and replication) separates a demo that looks like a result from one that actually is.

Several September 2026 fly-connectome demos documented important limitations in their own READMEs or other first-party project materials. The distortion entered downstream, in coverage that read the demo and skipped the documentation sitting right next to it.