alpha-fly Loading the brain

A fruit fly's brain would like a game.

Every neuron and synapse of the male Drosophila connectome, wired exactly as nature left it, taught chess by playing itself 322,000 times. Nothing in it was designed for chess. It runs right here, in your browser.

the fly · Black
you · White
Downloading the fly's brain

How a fly ends up playing chess

166,700
neurons, every one in the MaleCNS connectome
10.5M
synaptic connections, signs fixed by their neurotransmitter
0
human games or engine hints in the training data
322,016
games it played against itself
23 h
on one consumer GPU
1320
Elo parity with Stockfish's weakest calibrated level

The wiring is the fly's. In 2026 the MaleCNS project published the complete wiring diagram of a male fruit fly's central nervous system: which neuron connects to which, how many synapses, and whether each one excites or inhibits. This network is that diagram and nothing else. The board is fed to the fly's 17,937 sensory neurons, activity ripples through the wiring for six steps, and the move and the evaluation are read off its 2,333 motor and descending neurons.

What was trained is only how loud each synapse is: one gain per connection, a resting level and a time constant per neuron, and the linear maps in and out. No connection was added or removed, and no sign was flipped. Everything nonlinear the network does, it does inside the connectome.

It taught itself. AlphaZero-style: play games against yourself with a small tree search, learn to predict what the search preferred and who won, repeat. Stockfish only ever measured; it never taught. After 322,000 games the fly held its own against Stockfish at its lowest calibrated strength, and without any search at all it loses about 74 centipawns a move, roughly a 1200-rated human.

Is the fly's wiring special? Probably not for chess. A control with the same neurons and the same number of synapses but random partners learns at least as fast. What seems to matter is the scale and sparsity of the substrate, not who evolution wired to whom. Which is, in its own way, the more interesting finding.

In your browser the whole network is downloaded once and evaluated on your graphics card through WebGPU, or across your CPU cores where that is not available. "Fast" plays at roughly 1150 Elo and "Deep" at roughly 1450, measured in fixed-seed matches against Stockfish at its calibrated levels. Code, training logs and the trained weights are on GitHub.