DOOMFLY: A Complete Fruit Fly Brain, Simulated and Trained to Play Doom

Ten days ago, the most complete wiring diagram of any brain in history landed in public: the MaleCNS v1.0 connectome, a full map of an adult male fruit fly’s central nervous system — more than 166,000 neurons and roughly 125 million synapses, published in Cell by a collaboration led by HHMI’s Janelia Research Campus with Cambridge and Google Research. It was the culmination of an 18-year effort and the largest brain map ever completed.

This week, the internet did what the internet does. A software engineer took the entire simulated fly brain and taught it to play Doom.

Meet DOOMFLY

The project is called DOOMFLY, and its author is Alex Wormuth, a software engineer at Coinbase. Announced in stages over the past week and now circulating across PC Gamer, PCMag, Tom’s Hardware, and Slashdot, the build is exactly what the name promises: the complete MaleCNS v1.0 connectome — 166,700 neurons of simulated fruit fly nervous system — wired into the 1993 shooter as the game’s controller.

The architecture is surprisingly legible. Each rendered Doom frame is converted into visual signals and fed into the simulated brain’s sensory neurons. Neural activity propagates through the fly’s biological wiring — the actual circuitry mapped from electron microscopy, not a hand-designed network — and motor output is then mapped to the game’s controls: turning, moving, firing. The loop runs continuously. The fly sees; the fly acts.

The most elegant detail is the reinforcement signal. When the simulated fly takes damage, the harness delivers a stimulus to two PPL101 dopamine neurons — real, named cells in the fly’s actual reward circuitry. Punishment flows through the same channels evolution built for it. “Each Doom frame stimulates sensory neurons,” Wormuth explained. “Neural activity is mapped to game controls. Damage triggers a stimulus to two PPL101 dopamine cells as reinforcement.”

It is, in other words, not a neural network inspired by a fly. It is the fly — the whole connectome, running as a live simulation, with the game juicing its dopamine exactly where a biologist would.

Then came Mario and Beat Saber

DOOMFLY turned out to be the opening move of a genre. Within days, others picked up the same published connectome and pointed it at different games. A second team built a live neural simulation of more than 165,000 neurons and connected it to Beat Saber, so the fly brain now swings virtual lightsabers to music. Super Mario 64 followed — Tom’s Hardware’s headline simply reads that the “AI-powered 3D model of over 166,000 neurons can also play Super Mario 64.”

The pattern is now unmistakable: publish a complete nervous system, and hobbyists will have it playing classic video games within a week. When Doom — the eternal benchmark of portability jokes, having run on pregnancy tests, ATMs, and Minecraft — is now being played by a brain simulation, the “it runs Doom” meme has officially closed the loop: the game has finally found a player small enough to fit its reputation.

Why this matters more than the memes

It is easy to file this under internet whimsy, and plenty of the coverage has. But underneath the jokes sit three genuinely serious developments.

First, this is the first full-brain simulation in the wild. Before MaleCNS, complete connectomes existed only for the 302-neuron C. elegans worm and the roughly 140,000-neuron female fly brain. Now the male fly’s entire central nervous system has been mapped, packaged, and — critically — made runnable by people outside the original consortium. A complete brain became an open-source dependency this week. The distance from Nature-paper-to-side-project has collapsed from years to days.

Second, the simulation is credibly biological. Reporting on the broader effort notes that the virtual fly can walk, orient toward food, groom itself when dusted, and begin feeding — behaviors emerging from the reconstructed circuitry rather than from scripted animation. When a simulation of the mapped brain generates recognizable, ethologically correct behavior, the connectome stops being an atlas and starts being a program. Game harnesses are simply the most entertaining way to stress-test that program’s plasticity — feeding it sensory streams it never evolved for and asking whether its wiring adapts.

Third, the reinforcement loop is real neuroscience. Wiring damage to PPL101 dopamine cells isn’t a hack; it’s a faithful use of the fly’s actual punishment pathway. That makes DOOMFLY an accidental experiment in cross-embodiment learning: can a biological reward architecture shaped by 600 million years of evolution find gradients in a first-person shooter? The question sounds absurd, but it is the same question serious labs ask when they train neural networks in novel environments: what is innate structure worth?

From connectome to compute

The backdrop matters too. The connectome itself was only completable because of AI — Google Research’s flood-filling neural networks segmented each neuron from terabytes of electron microscopy data, driving a thousandfold efficiency gain in connectome generation. Machine learning built the map; now the map runs as a model. AI consumed biology to produce a brain, and the brain is now being consumed as software.

There is also an uncomfortable mirror for the AI field. A 166,700-neuron biological network, running in real time on ordinary hardware, is already generating adaptive behavior in alien environments. Meanwhile the frontier AI models that cost billions to train remain, in a precise sense, unfaithful copies — enormous statistical shadows of brains nobody has mapped. The fly simulation is tiny, but it is exact. Every synapse corresponds to a measured biological connection. It is the only “neural network” in existence whose architecture is ground truth rather than learned approximation, and watching it learn is the closest anyone has come to watching a real brain learn from the inside.

What to watch

The fly connectome era is just beginning. HHMI’s AI@HHMI initiative is building molecularly annotated connectomes that record not just which neurons connect but what chemical signals they exchange, and Janelia is at work on the larval zebrafish — a vertebrate — pairing wiring diagrams with whole-brain activity recordings to build predictive models of brain function. Each new map will arrive with the same property MaleCNS just demonstrated: it will be immediately loadable, immediately runnable, and immediately mobbed by engineers looking for something wild to plug it into.

The honest caveats deserve the last word. The fly is not good at Doom. Reddit skeptics correctly note that a simulated neural net playing Doom is not so different from any other neural net playing Doom — the novelty is the wiring’s provenance, not yet the scoreboard. Whether connectome-based agents can ever match trained artificial networks on performance is an open question; whether they can beat them on sample efficiency — leveraging innate circuitry instead of learning everything from scratch — is the scientifically interesting bet.

But as a milestone, the week stands on its own. A complete brain was published, and within days engineers had it fighting demons. The 1986 worm connectome took a decade to influence computational neuroscience. The 2026 fly connectome took about 96 hours to start shooting imps. Whatever else the connectome era brings, one thing is settled: if you publish a brain, the internet will run Doom on it.