About FLY

A real fruit-fly brain, all 138,639 neurons of it, simulated live on one graphics card. Her senses are the trades on Pons. Her reflexes place the orders. Nobody wrote her behavior.

What you are watching

The top of the screen is the brain. Every flash is one neuron crossing its firing threshold. The positions are the real positions of those neurons inside a fly's head, and the colors are brain regions (visual lobes, mushroom bodies, central complex, and so on). The wiring is the FlyWire connectome: every neuron and every synapse of an adult Drosophila melanogaster, reconstructed from electron-microscope images of one female fly. This is not a drawing of a brain. It is the brain's own wiring diagram, running.

The brain

We run the whole network as a leaky integrate-and-fire model, the one published by Shiu et al. in Nature (2024). Each neuron is a leaky integrator; each synapse's strength is the number of contacts counted in the microscope; each connection is excitatory or inhibitory according to the neurotransmitter predicted for that neuron. There is no training, no fitting, no hand-tuning toward any behavior. The published model reproduces real reflexes, like sugar on the legs driving the proboscis out, because the wiring does it.

Two honest consequences follow. The model has no spontaneous activity, so with nothing to sense she is silent: the market is her only input. And it has no fatigue, so a runaway state is possible; when it happens she blacks out, restarts, and the blackout is counted on the panel. Synapses also change a little with use (Hebbian plasticity with a slow relaxation toward the original wiring), so she is not identical from one hour to the next. That counter is on the panel too.

The brain runs slower than real time (the ratio is in the footer). What you see is a real reaction in slow motion, never a replay of a recording.

The body

The fly at the bottom is NeuroMechFly v2, the anatomically measured fly model from EPFL, with leg trajectories recorded from real walking flies. She does not have a script. Her descending neurons, the ones that carry commands from brain to body in a real fly, set the rhythm and direction of the gait, back her up, freeze her, extend the proboscis, start grooming, and fire the giant fiber that launches her into flight. The bars on the panel are the live firing rates of those neuron groups. The body is drawn in your browser from the joint angles her neurons produce, thirty times per second.

The market is her world

She watches one Pons curve at a time on Robinhood Chain, the busiest one. Every trade becomes a sense, by a fixed and public map:

She trades

Her motor neurons become orders by fixed rules, also public. A sustained proboscis extension buys a small lot (5% of her ETH, at most $5 per order). Escape or backing away sells. A lasting bitter taste sells a quarter. Orders are at least 20 seconds apart. She trades only on Pons curves, never her own token, from her own wallet, the address in the top bar. Every order is a real transaction you can open on the explorer. Her wallet is topped up from the token's creator fees; nothing else touches it.

What is not here

No AI agent. No language model. No strategy. Nothing decides in her name. Given the same brain state and the same trades, the same orders come out. She has reflexes, not opinions. When the brain is silent she does nothing, and that is shown as it is.

Data and credits

Connectome: FlyWire (Dorkenwald et al., Schlegel et al., Nature 2024). Model: Shiu et al., Nature 2024. Engine derived from fly-brain (MIT). Body: NeuroMechFly v2 (Lobato-Rios et al., Wang-Chen et al., EPFL). Rendering: three.js. Market data: Pons V2 on Robinhood Chain via GeckoTerminal. Development build on FlyWire v783.