What is real
The wiring diagram: every one of the 139,255 neurons and 54.5 million synapses of an adult fruit-fly brain, reconstructed from electron microscopy by the FlyWire consortium. Who connects to whom, how strongly, and whether each synapse excites or inhibits all come from that data. Every dot on the brain map is a real neuron drawn at its real position.
What is simulated
The activity. It is not recorded from a living fly. A leaky integrate-and-fire model (the same equations as Shiu et al., Nature 2024) runs on that real wiring, right here in your browser: "if a brain is wired like this and smells this, these neurons fire." The model reproduces known biology, such as sugar triggering the feeding motor program while bitter does not. No server is involved; your own device computes every spike, so a slower phone takes longer to decide.
How the fly picks a move (Innate Fly)
A shallow chess engine lists candidate moves. Each one becomes an odor: how good the move is sets the odor's strength, and the moving piece, captures and checks add their own smells. The odor travels through the antennal lobe into the mushroom body, the fly's learning center. Its output neurons give each odor a value (valence), and the fly plays the most attractive move. It reliably spots clearly better moves, but it cannot tell apart moves of nearly equal value, so it picks among them by its own taste.
Learning Fly
The engine only suggests 8 reasonable moves. All of them smell equally strong, so the engine's opinion of them never reaches the fly. What the fly senses about each move is plain fact: which piece moves, whether it captures or gives check, whether it threatens an enemy piece, and whether it escapes from an attack. Which of these is good, it has to find out. After each move it sees the consequence (material won or lost by the next reply). Dopamine reports the surprise: how much better or worse that was than its moves usually turn out. Better than expected releases reward dopamine (PAM neurons), worse releases punishment dopamine (PPL1). Half of the signal also reaches the previous move, so a threat earns credit when the capture comes one move later.
Dopamine weakens the synapses from the Kenyon cells that were active, exactly where the connectome says those dopamine neurons connect. The Kenyon-cell activity comes from the spiking simulation; the synaptic change is computed in a rate-based layer built on the real Kenyon cell → output neuron synapse counts, because in the spiking model the dopamine-targeted output neurons stay silent. What your fly learns is saved in this browser only.
Pure Fly
In Innate mode the engine shortlists the 8 best moves and turns their quality into odor strength, so most of the chess skill comes from the engine; the Learning Fly still gets the engine's shortlist. Pure Fly removes the engine entirely. Code only reports what the fly would sense: which piece moves, what it captures (its kind, not its value), whether it gives check, and whether the moved piece lands under attack. Every legal move smells equally strong; the mushroom body alone decides. It starts out playing almost at random and can only improve by learning from captures and losses.
Pondering
While you think, the fly simulates its replies to your most likely moves. If you play one of them, its answer is already computed and identical to what it would compute anyway, so it replies instantly. "LIVE" on the brain map means the activity is being computed at that moment; "REPLAY" means recorded activity from earlier.
Dorkenwald et al. 2024 · Schlegel et al. 2024 · Shiu et al. 2024 · Aso et al. 2014 · Hige et al. 2015