flæ

Every one of the fly's 165,122 neurons is simulated as a leaky firing-rate unit, connected by 6.1 million real synaptic connections whose wiring is never changed. The board drives sensory neurons, and descending neurons (the brain's outputs to the body) are read out as a running preference over moves, plus a signal for when to stop thinking.

The model has firing rates rather than individual spikes, so “active” means a neuron is firing above 10% of its maximum rate. “Recruited” counts every neuron that was active at any point while the fly thought about the current move.

The brain never resets during a game and keeps running while you think, too. Its guess at your reply comes from the same readout.

This fly plays by smell. Its brain is the real wiring without the optic lobes and nerve cord (39,109 neurons), simulated spike by spike after Shiu et al. (2024). Each candidate move becomes a smell of chess facts (material, hanging pieces, threats, what your greediest reply would win) on its antennae, and the smell travels through the antennal lobe to the Kenyon cells of the mushroom body.

It learned which smells are good the way flies do: sugar or a shock right after a smell, delivered by dopamine neurons, weakened particular Kenyon-cell synapses onto the mushroom body's output neurons. How attractive a move smells is the share of that output going to the “approach” side.

It sniffs the legal moves one at a time, in a random order, for half a second of brain time each, and plays the first one that smells clearly better than usual, or the best it has smelled when it runs out of moves or time. “Active” means a neuron spiked in the last 10 ms of brain time. Nothing is learned during play.