Fruit Fly / Fruit Fly Gaming
A Real Fly Brain Inside Minecraft: How the Fly Brain Minecraft Mod Runs the Male Fruit Fly Connectome, and Why It Isn't AI Hype
A Minecraft fly driven by the complete male fruit fly connectome: 176,422 neuron records, 6.29 million connections and a published neuron model. How the mod turns real wiring into feeding, escape and grooming, what was built by hand, and why there is no AI model inside the brain.
Akmal Alif · 10 October 2026 MYT
On 6 September 2026 the developer Blendi (@BlendiByl) posted a short clip of a fruit fly in Minecraft. Beside the fly, one panel showed a map of its whole nervous system with neurons flashing as they fired. A second panel read out what the fly had decided to do (Blendi, 2026a). The code went up the same day as Fly Brain Minecraft, an open-source mod (Blendi, 2026b).
It looks like a game trick. It is closer to a neuroscience model with a game attached. This post explains what the clip shows, how the fly's brain is simulated, and why the result is not AI hype.
What you are looking at
The clip shows the mod's two heads-up panels (Blendi, 2026b):
The brain view. A live map of the whole nervous system: both optic lobes, the central brain between them and the nerve cord below. Each neuron lights up at its real cell-body position when it fires, then fades over about 300 ms.
The neuroscope. The readout. It shows the decoded motor channels (forward, turn, back, stop, land, flight, feed, groom, song), the firing rates of key neuron populations, a spike raster, and the fly's view of the world through its simulated retina.
In one frame of the clip the decoder is in feeding mode, with about 52,000 spikes in a 50 ms tick and 129,000 neurons active. The real-time factor reads 0.66, meaning the brain simulation was running a little behind the game at that moment.
Two posts, two days
The clip was a reply. Two days earlier, on 4 September, the user @evnsnclr had posted a fly in Minecraft driven by the male fly connectome, saying it was built with the help of GPT-6 Astra and promising code soon. That post drew about 14 million views (evnsnclr, 2026). Blendi's reply: "did something similar". This time the code, data pipeline and validation notes were public on day one.
Both build on the same dataset. The male central nervous system connectome was published in Cell in September 2026 by teams at HHMI Janelia, Google Research and the University of Cambridge. It is a complete wiring diagram of the male fruit fly's brain and nerve cord: 166,700 proofread neurons, mapped synapse by synapse from electron microscopy (Berg et al., 2026). The mod's changelog dates its first version to the day that paper appeared.
How the fruit fly model works
The mod rebuilds the fly's brain from measured wiring and a simple, published neuron model. Step by step:
The wiring diagram. The mod loads every neuron record in the public neuPrint dataset male-cns:v1.0: 176,422 entries, of which the paper's 166,700 are the proofread neurons. It keeps the 6.29 million connections of five or more synapses, which carry 90 million of the dataset's 125 million synapses. Each neuron keeps its neuPrint ID, so anything on screen can be looked up in the public database (Plaza et al., 2022).
One simple neuron model, copied from a peer-reviewed paper. Every neuron is a leaky integrate-and-fire unit with the parameters of Shiu et al. (2024). Its voltage drifts back towards rest at −52 mV. It spikes when it crosses −45 mV, then resets and stays silent for 2.2 ms. Each incoming synapse nudges the voltage by 0.275 mV, scaled by one global gain, after a delay of 1.8 ms.
Excitation or inhibition, from the transmitter. Neurons that release acetylcholine excite their targets. Those that release GABA, glutamate or histamine inhibit them. Neurons with monoamine or unclear transmitters are treated as excitatory.
No spontaneous activity. A silent world gives a silent brain. Every spike can be traced back to a sensory neuron.
The game feeds the fly's real sensors. The world drives the matching sensory neurons as random spike trains: photoreceptors, smell and taste receptors, the hearing and wind organ in the antenna, and touch bristles.
Motor neurons drive the mob. The firing rates of real descending and motor neurons are decoded into what the fly does. A priority ladder runs from escape and landing down to walking, grooming and song.
In real time. Each fly gets its own thread, with 50 ms of brain time per game tick in 0.5 ms steps. Only neurons that are away from rest or have input pending are computed, so the cost scales with activity, not with brain size. On a 32-core machine, one fly needs 25–60 ms of compute per 50 ms tick.
The model is a direct descendant of Shiu et al. (2024). That team built the same kind of simulation from the earlier female fly brain map, which has more than 125,000 neurons (Dorkenwald et al., 2024), and used it to make predictions. In the model, activating sugar-sensing neurons predicted which neurons respond to taste and drive feeding. The team then confirmed predictions like this in real flies by switching neurons on with light (optogenetics) and watching their behaviour (Shiu et al., 2024).
What the fly does, and what is real
The repository says plainly which behaviours come from the wiring and which were built by hand (Blendi, 2026b):
Behaviour | Pathway in the connectome | Where it comes from |
|---|---|---|
Feeds on sugar | Sugar taste neurons, through G2N-1 and Fudog, to the MN9 proboscis motor neuron | Emerges from the wiring |
Refuses bitter food | Bitter taste neurons to Scapula, which silences MN9 even with sugar present | Emerges from the wiring |
Jumps and escapes | Looming detectors LC4 and LPLC2 to the giant fibre neuron, then the jump muscle | The decision emerges; the looming signal is computed by hand |
Grooms | Antennal and bristle sensors to the head-grooming neurons | Emerges from the wiring |
Walks, turns, halts | Walking, turning, backward and braking descending neurons | The readout emerges; the gains are set by hand |
Walks towards food | The smell circuit saturates in this model, so a reflex layer steers | Built by hand |
Flies and lands | Takeoff and landing neurons | A hand-built state machine |
Courtship song (males) | pC1 and P1 to the pIP10 song neuron, one wing extended | Wired, not yet demonstrated |
Why this is not AI hype
No AI model is inside the brain. Nothing here was trained on data. The "weights" are synapse counts measured under an electron microscope, plus one global gain. There is no language model in the loop. The fly's choices are spikes flowing through a measured wiring diagram, decoded at real motor neurons.
It has been checked against biology. The mod includes headless benchmarks, run with fixed random seeds (Blendi, 2026b):
With no input, the brain produces 0 spikes.
Driving sugar taste neurons makes MN9, the proboscis motor neuron, fire at 30–90 Hz. Kenyon cells, the learning centre, stay silent.
Adding bitter input silences MN9 completely, while upstream sugar neurons keep firing. The suppression is downstream of them, as Shiu et al. reported.
Driving the looming detectors makes the giant fibre fire at 330–380 Hz, and the decoder enters escape.
Only one number was tuned, and the choice is documented. The published synaptic weight was fitted to the female brain map, which has fewer synapses per neuron. At that weight the male network over-excites. The global gain was set to 0.65 using the paper's own recipe: the smallest value at which sugar drives MN9 while the grooming and escape pathways stay stable. The repository also documents a rejected alternative and an independent replication that argues for a lower gain.
The limits are spelled out. The repository lists them itself:
Every neuron has the same parameters.
There are no neuromodulators, neuropeptides, gap junctions or learning.
The smell circuit saturates, and the motion pathway of the visual system stays silent, so looming is painted onto the looming detectors directly.
Female flies run the male brain.
Absolute firing rates should not be trusted, only which populations respond.
Where AI does come in. @evnsnclr credits GPT-6 Astra with helping to build their version. Coding assistants can plausibly speed up work like this: writing a fast simulator, a game mod and a data pipeline. That is AI helping to write software. The behaviour itself comes from the connectome.
Why put a fly brain in Minecraft?
Because it turns a dataset into something you can poke at. In the mod, you can drive any named population of neurons and watch the fly respond. You can add a population to the neuroscope, or build the whole connectome out of blocks and walk through it while the linked fly's neurons flash. That makes it a teaching tool and a sandbox for hypotheses: a fast way to ask "what happens if…" of a real wiring diagram.
It is not a conscious fly, a learning fly or a complete one. It is a faithful wiring diagram running a deliberately simple neuron model, and the project says so.
We have covered other uses of the same connectome: Eris's fruit-fly next-token experiment and Neural Canvas, a connectome simulation in the browser.
References
Berg, S., Beckett, I. R., Costa, M., Schlegel, P., Januszewski, M., Marin, E. C., Nern, A., Preibisch, S., Qiu, W., Takemura, S.-Y., Fragniere, A. M. C., Champion, A. S., Adjavon, D.-Y., Cook, M., Gkantia, M., Hayworth, K. J., Huang, G. B., Katz, W. T., Kämpf, F., . . . Jefferis, G. S. X. E. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell, 189(18), 5504–5526. https://doi.org/10.1016/j.cell.2026.08.015
Blendi [@BlendiByl]. (2026a, September 6). did something similar, with a map of the brain showing which neurons are firing in realtime [Post]. X. https://x.com/BlendiByl/status/2096436970113454230
Blendi. (2026b). Fly Brain Minecraft [Computer software]. GitHub. https://github.com/blendi-remade/fly-brain-minecraft
Dorkenwald, S., Matsliah, A., Sterling, A. R., Schlegel, P., Yu, S.-C., McKellar, C. E., Lin, A., Costa, M., Eichler, K., Yin, Y., Silversmith, W., Schneider-Mizell, C., Jordan, C. S., Brittain, D., Halageri, A., Kuehner, K., Ogedengbe, O., Morey, R., Gager, J., . . . Zandawala, M. (2024). Neuronal wiring diagram of an adult brain. Nature, 634(8032), 124–138. https://doi.org/10.1038/s41586-024-07558-y
evnsnclr [@evnsnclr]. (2026, September 4). I've successfully run the full retained MaleCNS v1.0 fruit fly connectome, all 166,700 neurons, inside Minecraft [Post]. X. https://x.com/evnsnclr/status/2095975490708291948
Plaza, S. M., Clements, J., Dolafi, T., Umayam, L., Neubarth, N. N., Scheffer, L. K., & Berg, S. (2022). neuPrint: An open access tool for EM connectomics. Frontiers in Neuroinformatics, 16, Article 896292. https://doi.org/10.3389/fninf.2022.896292
Shiu, P. K., Sterne, G. R., Spiller, N., Franconville, R., Sandoval, A., Zhou, J., Simha, N., Kang, C. H., Yu, S., Kim, J. S., Dorkenwald, S., Matsliah, A., Schlegel, P., Yu, S.-C., McKellar, C. E., Sterling, A., Costa, M., Eichler, K., Bates, A. S., . . . Scott, K. (2024). A Drosophila computational brain model reveals sensorimotor processing. Nature, 634(8032), 210–219. https://doi.org/10.1038/s41586-024-07763-9