A testing simulation of a fruit fly with its real brain. All 138,639 neurons of the FlyWire connectome, firing live in your browser.
Fly Simulator runs on your kem108 account, the same one you use on kem108.dev and Le Horse.
Checking your kem108 sign-in…
Getting the brain ready
The wiring is real. It's the FlyWire connectome of an adult female fruit fly (version 783): 138,639 neurons and 15.1 million connections, every one traced from electron microscopy. Strong connections (5+ synapses) load first; on fast devices the weaker ones stream in after.
The neurons follow the published model from Shiu et al. (Nature, 2024): leaky integrate-and-fire, with their exact constants. It runs live in a background thread, only computing neurons that could be near firing. Checked against the paper: tasting sugar drives MN9, the proboscis motor neuron, at about 97 Hz and lights up about 400 neurons.
Two changes. Transmitters come from FlyWire's literature-confirmed annotations where they exist (Kenyon cells release acetylcholine, not dopamine as the classifier guesses). Dopamine, serotonin and octopamine act as slow learning signals instead of fast excitation, which in the published version sends smells and heat into a seizure-like state.
It learns for real. Dopamine neurons change the actual synapses from Kenyon cells (the smell code) to the mushroom body's output neurons. That's how it learns that things are scary or safe. Repeated harmless scares also make it habituate to them.
What isn't brain. The nerve cord (the fly's spinal cord: walking rhythm, leg coordination) isn't in the dataset, so the simulation plays that part. The same goes for the optic lobe's motion maths, which is fed into the real looming and motion neurons (LC4, LPLC2, LC16, LPLC1, LC10, LC11, HS, VS), and for pain and sugar reward, which arrive through the brain's own dopamine neurons. Decisions such as escape jumps, eating, grooming, backing up and steering toward moving things come from the descending and motor neurons that fire.
Data: FlyWire Consortium (Dorkenwald et al. 2024; Schlegel et al. 2024; Eckstein et al. 2024), CC BY 4.0. Model: Shiu et al. 2024 (MIT). 3D: three.js (MIT), Kenney asset packs (CC0). Sounds: NASA black hole sonifications and mission audio, Freesound (CC0). The black hole is the kem108.dev one.