FlyWire (FAFB v783)
The whole adult female brain, optic lobes included. The dots on this page come from here.
FlyWire v783 · every dot is a real cell
A fruit fly brain was cut into thousands of 40-nanometre slices, imaged with an electron microscope and rebuilt synapse by synapse. Then it was run as a program, put into a virtual body — and handed to the internet. Here is everything that came of it.
Lab
Pick what the fly will sense. The signal enters through real sensory neurons and runs along real connections — 15 million of them, not one made up.
This is not a video or a pretty animation. Every flash is a spike in a run of the Shiu et al. model (Nature, 2024) on the full FlyWire v783 connectome: 138,639 neurons and 15,091,983 connections, each connection signed by its predicted transmitter. We re-implemented the model ourselves in numpy instead of Brian2: on their reference sugar experiment our twin activates the same neurons, with a 0.94 correlation of firing rates. Each stimulus was run 10 times for one second; rates in the list are averages, the 3D view replays the first run. This fly has no body, only a brain.
Smells had to be given at low rates: in this model the olfactory system works like a switch. Below threshold the signal never leaves the receptors; above it, within 20–60 ms, the network runs away into the same state of 10,000 neurons, whatever triggered it. Heat, cold and humidity run away already at 20 Hz, so they have no buttons of their own — only “Runaway”.
Story
Eight years from the first image to a fly that walks, grooms and eats in a virtual world on its own.
Bargmann and Marder write the classic paper on why a connectome is not enough: the worm C. elegans has had its wiring map since 1986 and is still hard to understand. The argument goes on today.
Zheng et al. (Cell) publish an electron microscopy volume of an entire adult fly brain. Still a picture, not a map: every neuron has yet to be traced. ~100,000 neurons
Half of the central brain, mapped to the synapse. The first big adult insect connectome you can open in a browser through neuPrint. ~25,000 neurons
Winding et al. (Science) release the complete brain connectome of a fruit fly larva. 3,016 neurons · 548,000 synapses
The fly’s “spinal cord” — the ventral nerve cord that runs the legs and wings — is mapped separately, for a male (MANC) and a female (FANC). 23,665 neurons in MANC
Nine papers in Nature. A consortium of 50+ labs and hundreds of volunteers proofread the automatic segmentation of FAFB. The same day, the Shiu model comes out: the whole brain run as a spiking network on a laptop — and it predicts which neurons make a fly eat. 139,255 neurons · 8,453 types
Lappalainen, Turaga et al. (Nature) build a model of the visual system from the connectome and train it to see motion. Its predictions matched 26 independent experiments.
The Ramdya lab (EPFL) gives the fly a physical body with 87 joints, eyes, smell and sticky feet — in the MuJoCo simulator.
Nern et al. (Nature): the complete wiring of one optic lobe — over a third of all the fly’s neurons live in its eyes. ~53,000 neurons · 732 types
A detailed body that walks and flies, steered by neural networks trained with reinforcement learning. No connectome yet — a platform for a future brain.
Sandia runs the whole of FlyWire on 12 Intel Loihi 2 chips — hardware that itself computes in spikes.
A startup puts the Shiu model into the NeuroMechFly body: the fly looks for food, eats and grooms in a virtual world. Loud headlines, harsh criticism — see “Honestly” below.
Two connectomes in one day. BANC (Nature) is the first to join the brain and nerve cord of one female. Janelia opens MaleCNS — the whole nervous system of a male. ~188,000 neurons in BANC
The paper on the male’s complete nervous system. Google, who did the segmentation, blogs about it — and the internet notices. 166,700 neurons · 11,710 types · ~125M synapses
Within a week, dozens of projects: Doom, Mario 64, Minecraft, Beat Saber, crypto trading, haiku, tic-tac-toe and wildfire detection. “Bad Apple” on a fly brain gets 11.5M views. By 12 September the fly had played Doom 6,385 times and never beaten it.
Maps
A connectome is the full wiring diagram: which neuron connects to which, and with how many synapses. The fruit fly already has several, all open to everyone.
The whole adult female brain, optic lobes included. The dots on this page come from here.
The male’s whole nervous system: brain, eyes and nerve cord. Almost the entire September 2026 wave of games is built on it.
The first connectome where brain and nerve cord come from one animal. Key finding: movement is controlled mostly locally, not from the brain.
One right optic lobe — complete, with every cell type.
The ventral nerve cord of a male and a female — what moves the legs and wings.
Half of the central brain. Almost every tool people use today started with it.
The complete brain of a fruit fly larva — the first whole insect brain.
Models and bodies
A wiring diagram does nothing by itself. For the fly to “come alive” you need a neuron model that runs signals through the wiring, and a body with physics for that brain to steer.
Each neuron is the simplest integrate-and-fire unit, connections come from FlyWire, their sign from the transmitter. Nothing else. And this model predicted which neurons make a fly eat and groom — confirmed in live flies. It’s what runs in the lab above.
The Shiu model plus flyvis vision, wired to the NeuroMechFly v2 body at 15 ms steps. The fly finds food, eats and grooms. The body code isn’t open; motor output goes through a handful of hand-picked neurons.
A fly body in MuJoCo: 87 joints, compound eyes, smell, sticky feet, rough terrain. The most popular body for connectome experiments.
The most detailed fly body: it walks and flies, steered by trained neural networks. No connectome brain inside yet — it’s a test bench for one.
The visual system from the connectome: 64 cell types, unknown parameters fitted by training. The wiring can’t see on its own — it needed tuning to a task.
All of FlyWire on 12 Intel neuromorphic chips: orders of magnitude faster than ordinary simulation, especially when activity is sparse.
The connectome as a ready-made neural network architecture: the brain graph is trained to walk with reinforcement learning, and it learns faster than ordinary networks.
An exam for digital flies: does the simulation do what a real fly does? Sugar → proboscis, shadow → giant fiber, and so on.
Catalog
Once the brain was opened to everyone, everyone got to work: scientists, game makers, artists and traders. The fly plays Doom, drives a car, trades crypto and writes haiku.
The community’s most-starred project: a fly crawls over your macOS desktop, with 23,210 of its neurons in a side window.
The most detailed fly body: walks and flies with RL-trained controllers. A test bench for a future connectome brain.
The Shiu model on six engines — Brian2, CUDA, PyTorch, NEST GPU, GeNN — with benchmarks and parity checks.
The fly trades bitcoin; profit and loss hit its dopamine cells. $100 balance, $10 per trade, 24 trades a day.
Paper trading by default; no profitable learning shown.
MaleCNS access from R, with morphology tools.
The fly body in MuJoCo: 87 joints, compound eyes, smell, sticky feet. The most popular body for connectome experiments.
FlyWire's 138,639 neurons spike in the browser, play three games and write exactly reproducible bug reports, with no LLM.
The author measures steering vs goal at zero correlation and says so up front.
The larval L1 connectome (Winding et al.) modulates a segmental VNC controller driving an 11-segment MuJoCo larva.
The browser demo uses a reduced controller; the full 2,952-node network runs locally.
Tic-tac-toe against a full MaleCNS LIF simulation; a readout on descending and VNC neurons is trained by self-play REINFORCE.
Wiring never modified; only the linear readout is trained. Self-hosted, no public demo.
A pet fly on your phone, offline: FlyWire's 138,584 neurons stepped as LIF cells in Rust, behaviour read from the activity.
The app shows on screen which input and output mappings are the authors' own choices.
FlyWire plus a NeuroMechFly body 'does' 1C:Enterprise labs: a PDF manual becomes configs that load into a real 1C base.
The fly only picks among moves the program prepares via named descending neurons; a small external decoder is trained.
The Shiu et al. brain model on Apple MLX with a Metal kernel: 1 s of brain time in 0.29 s on an M4 Pro vs 62.6 s in Brian2.
FlyWire v630, paper parameters unchanged; nothing is trained.
Honestly
Headlines say “a fly was uploaded to a computer”. What was uploaded is the wiring diagram, and that is not the same thing as a brain.
A connectome has no ion channels, hormones, neuromodulators or memory. It tells who connects to whom, not how those cells actually behave. Even Eon says so.
Walking is done by a ready-made body model; the brain only switches its programs through a few hand-picked neurons. The vision, by the authors’ own admission, is “somewhat decorative”. Critiques: Carboncopies, LessWrong.
The “91%” and “95%” from press releases measure how well the Shiu model predicts experiments, not how similar the whole virtual fly is to a real one. Even Eon’s neuron count jumps between 125 and 140 thousand.
How a Doom frame becomes a signal for the eyes and which neuron means “shoot” is chosen by the author. Doomfly honestly says: “getting better is unproven”. In Beat Saber a memorised sequence is playing.
No project has shown anything like experience in a digital fly. The NeuroCraft author says it plainly: this is not evidence of consciousness or a complete copy of a living fly.
A mouse brain is about 560 times bigger. Over 95% of a connectome’s cost is manual proofreading, and it hasn’t got cheaper yet. Sceptics like Kenneth Hayworth think even an honest fly upload is years or decades away.
The hype brought “fly brain” tokens and repos that clone other people’s projects. There are deliberately no links to them here.