A Fly Brain Played Beat Saber. It Only Knew One Song.

TL;DR

On September 3, Google Research and HHMI Janelia published MaleCNS, the complete wiring diagram of a male fruit fly's central nervous system, in Cell: about 166,000 neurons and 125 million synapses. Within a week, builders had "the fly brain" playing Minecraft, Doom, Super Mario 64 and Beat Saber. Soon it was also steering a car and reading text. On September 14, Patrick Mineault went through the code and wrote: "we have not uploaded flies." Most demos he could inspect never connect what the fly sees to what it does, and the most careful repos admit as much in their READMEs. The lesson travels beyond flies: when a frozen network plus a trained readout does something impressive, find out which part earned the applause.

A serious map, then a week of memes

The dataset is real science. MaleCNS covers the brain plus the ventral nerve cord, the fly's rough equivalent of a spinal cord, which carries motor commands to the body. Janelia's FlyEM team led the project with the University of Cambridge, the MRC Laboratory of Molecular Biology and Google, which supplied computational and segmentation support.

It has also been downloadable for a while. Version 0.9 went public in October 2025 and v1.0 followed on June 8. The Cell paper, "Sexual dimorphism in the complete connectome of the Drosophila male central nervous system," is what put it in front of everyone.

Then X got hold of it. A Minecraft fly appeared on September 4, its maker noting it was "Built with the help of GPT-6 Astra." Within days there were flies in Doom, Mario 64 and Beat Saber, a fly steering in the CARLA driving simulator, a fly doing OCR, and a fly day-trading crypto. The Google AI account, "which has 2.4M followers, even tweeted about the public response," Mineault notes.

One developer, quoted by IBTimes UK, objected that everyone was "forcing it to play the same beat saber song indefinitely" and started building a sim where it "just gets to fly around forever in fruit fly heaven." The fly, as far as anyone can tell, has not filed a complaint.

What a "fly brain" sim actually runs

The template is Shiu et al., a 2024 Nature paper that built a leaky integrate-and-fire simulation straight from an earlier fly connectome and showed it produces sensible activity instead of dying out or seizing. The Mario 64 repo explains the idea plainly: "Each cell has a number that stands in for its voltage. That number fades toward zero." Inputs push it up or down, and when it crosses a set level the cell fires and resets, 50 steps a second.

Doomfly, FLM and Fly OCR all load the same retained graph: 166,700 neurons and 25,582,938 directed connections. That is the headline synapse count collapsed into weighted neuron-to-neuron edges. The Fly OCR research report spells out the recipe: each weight is the contact count times an inferred transmitter sign and a fixed 0.275 scale, and "Unknown or ambiguous signs default positive for 3,718 cells."

So the map tells you who talks to whom. For thousands of cells, it does not tell you whether the message means go or stop.

Three ways to make a fly look smart

Mineault's teardown walks through the tricks. None of them needs the fly's wiring to do the clever part.

1. Let noise do the dancing

A recurrent network, including a connectome-based one, "can generate rich activity when driven by random noise," Mineault writes. "Think of a camera pointed at a television screen projecting its image: time delays, nonlinearity, and user manipulation can create highly complex patterns." Wire that activity to a game controller and it looks like agency.

The Mario 64 fly is the clean example. Its README says the model adds "a steady background drive and repeatable noise. So the screen is not the cause of every movement." Rules the author wrote map a descending neuron called DNg100 to forward movement and a right-minus-left difference in two other cell types to steering. "There is no training, reward, or goal," the README says. The code is "literally 100% vibe coded" with GPT Astra, and the author adds: "I have not reviewed the code."

Mineault's term for the result is "random button mashing." Mario, for the record, did not get any better at Mario.

2. Train a flexible readout

The Beat Saber clip was the one that broke containment. According to Mineault, its author "mentioned that it's overtrained on one track and doesn't generalize to others." IBTimes reports the motor system "had been overfit to reproduce a recorded movement sequence." Other projects, Mineault writes, involve "training a very flexible policy on top of the connectome, or backpropagating through the connectome." His objection is blunt: "a big network with a lot of free parameters can learn many things, fly or not."

3. Borrow the hard parts

This one predates MaleCNS. In March, Eon Systems showed a virtual fly that walks, grooms and feeds, built on the older FlyWire brain connectome (about 140,000 neurons) and the NeuroMechFly body. Eon's own write-up is candid about it.

The mappings from neurons to movement "can be somewhat arbitrarily chosen by hand (as is our case)." Walking came from existing controllers "trained to imitate the walking behavior of the fly." And the visual activations piped in from the Lappalainen et al. model are "somewhat 'decorative' in that they do not currently substantially influence our behavioral outputs."

A worm brain can walk a fly

The sharpest rebuttal came from neuroscientists, in what Mineault drily calls "a strongly worded preprint." Bingni Brunton, Elliott Abe, Lawrence Hu and John Tuthill built the digital sphinx: the connectome of a C. elegans worm hooked up to a biomechanical fly body. A network trained with deep reinforcement learning maps the worm's motor neurons to the fly's leg actuators.

It walks like a fly. The authors call it "biologically meaningless" and sum up the lesson in one line: "behavioral fidelity is achievable without biological fidelity, making such models easy to overinterpret."

Here is the intuition. Drop pebbles into a pond and the ripples get complicated fast. Train someone to read the ripples and they can tell you which pebble you threw, but that says more about the reader than the pond. Machine learning calls this reservoir computing: a fixed recurrent network churns inputs into rich activity, and a trained readout pulls out whatever you ask for. As Mineault puts it, "with enough flexibility in readout, you can control the gait of a virtual fly body with a worm brain."

The repos that graded themselves

The fairest part of this story is that several builders published their own negative results. Those READMEs are more informative than the clips.

• Doomfly. Game frames drive the full graph, and a plasticity rule driven by dopamine activity adjusts 4,184 existing connections during play. The README leads with its status: "The current v6 candidate failed its visual, conditioning and survival validation gates."
• FLM, the fly language model. Tokens drive the frozen fly graph, and a 278,528-parameter adapter reads its state to adjust the next-token scores of Liquid AI's LFM2.5-1.2B-Instruct. "Language ability comes from the pretrained model," the README says. In the linked paper's frozen study, the matched direct-input control "performed slightly better; it does not establish an advantage from fly anatomy."
• FlyHard. The simulated fly body physically turns a steering wheel in CARLA. After 600 optimizer updates and 186 seconds of training, it passed 100 of 100 held-out steering targets, up from zero. Visual driving remains untested, and the current video "uses scripted speed and requested turns."
• Fly OCR, from LlamaIndex CEO Jerry Liu. Glyph pixels pass through the frozen circuit, and a 266,628-parameter decoder reads 1,024 downstream neurons. It scores 87.6% on a 1,632-glyph benchmark, but "only 1/8 PDF lines is exactly correct, and a 3-degree tilt breaks segmentation."

The controls worth reading slowly

Fly OCR's report goes furthest. Mismatching the recorded neural responses to labels drops its 68-class benchmark to 2.3%, near the 1.5% chance rate, so the circuit's activity does carry signal. In an earlier 200-example digit pilot, the intact circuit scored 82%, while three degree-preserving rewirings with retrained heads scored 65.0%, 63.5% and 54.5%.

So the real wiring helps that decoder. It is not magic, though: on a 500-example digit test, a plain linear classifier on raw pixels reached 99% and a small CNN hit 100%, versus 89% through the fly.

The report's own verdict: "Signals matter, but biological superiority is unproven." It warns that randomization "substantially changes activity scale, and the control budget is small," and concludes that "An ordinary OCR system remains the practical choice for accuracy and generality."

Why this is a builder story

AI coding tools powered this wave. A Mario 64 port nobody reviewed, a Minecraft build credited to GPT-6 Astra, and a handful of repos with their own simulation code, controls and dashboards all showed up within about ten days of a Cell paper. Mineault's teardown ran on the same fuel: "I asked Claude to fetch the code for the other viral popular fly sims from the last week and dig into them."

That speed cuts both ways. The tools that wire 166,700 neurons into a game over a weekend also hand you a convincing video before anyone runs the control that would deflate it. The better repos show the fix:

• Ablate what you credit. If the connectome is the story, run the same readout on a rewired graph and on the raw input with no graph at all.
• Count trainable parameters. A decoder with a quarter of a million weights can learn a lot by itself.
• Separate clip from claim. One overfit track is a music video, not a skill.

It is the same question you ask when an agent harness, not the model, moves a benchmark score. Mineault is not waving anyone off: "more people interested in neuroscience is good!" His suggested next steps lean on what makes MaleCNS new, such as decoding motor commands from the ventral nerve cord, and on ways to "bring outside constraints to the network."

Caveats

Mineault reviewed the demos whose source he could find, and he says some builders are clearly "trying to make them work non-trivially." Several repos are days old, and FlyHard and Doomfly describe themselves as works in progress.

All repo numbers are self-reported results from small experiments, and Fly OCR's rewiring figures come from a separate pilot, not its final letter model. None of this is a knock on the connectome itself. The critique is about what the viral clips appeared to show.

Key Takeaways

• The map is real. MaleCNS, published in Cell on September 3 by Janelia, Google and partners, charts about 166,000 neurons and 125 million synapses, including the ventral nerve cord.
• The uploads are not. Mineault found that most viral demos with available source do not close the loop from sensation to motor output: "we have not uploaded flies."
• Readouts do the lifting. The Beat Saber fly was overtrained on one track, and a worm connectome with a trained decoder can walk a fly body.
• Honest repos say so. Doomfly's v6 failed its validation gates, FLM's matched control did slightly better, and Fly OCR's circuit trails a linear classifier on raw pixels, 89% to 99%.
• Vibe-coded science needs controls. Agents made these builds possible in days. Ablations against random graphs and raw inputs are what make them mean something.

Sources: Patrick Mineault: Deconstructing viral fly sims, Google Research: A connectomics milestone, Cell: Sexual dimorphism in the complete connectome of the Drosophila male CNS, HHMI Janelia, MaleCNS release notes, IBTimes UK, ornata/fly (Mario 64), nftechie/doomfly, nftechie/flm, MarkUnthank/flyhard, jerryjliu/fly_ocr, Fly OCR research report, eLife: The digital sphinx, Eon Systems: How the Eon Team Produced a Virtual Embodied Fly, Shiu et al., Nature (2024), Lappalainen et al., Nature (2024)