flymate-chess

Browser bundle of a chess-playing fruit-fly connectome — the complete FlyWire FAFB v783 brain (138,639 neurons, 15,091,983 synapses, wiring and excitatory/inhibitory signs frozen) with a learned encoder, one positive gain per synapse, per-neuron homeostasis, and a decoder, trained on Stockfish-annotated Lichess positions. Recovers Stockfish's top move 29.8% of the time on held-out positions (ChessFly, the original recipe by Maxime Labonne: 30.4%). Trained 3 epochs on a hash-uniform 4.4M-position sample with soft multi-PV policy targets (gofly run 3).

Play it: https://flymate.naklitechie.com · code: https://github.com/NakliTechie/flymate

Files are raw little-endian arrays consumed by the page's WebGPU kernels; meta.json carries shapes, the 1968-move vocabulary and torch reference outputs for four positions. FlyWire's non-commercial terms apply to anything derived from the graph.

Torch checkpoints (checkpoints/)

state_dicts for the gofly FlyNet module (per-edge log-gain θ, per-step folded norm + gain/bias, encoder, heads), with each run's eval log:

file	training	held-out Stockfish top-1 · value MAE
run3-3ep-29.8.pt	Lichess evaluations, hash-uniform 4.4M, multi-PV soft targets (τ=0.05), 3 epochs	29.8% · 0.088 — source of the browser bundle above
run7-4ep-29.9.pt	run 3 resumed for a 4th epoch	29.9% · 0.087 — plateau; not shipped to the app

ChessFly (Labonne / Liquid AI) reports 30.4% · 0.081 on the same task. The plateau at ~30% is consistent with the ablation in naklitechie/flymate-go: a degree-preserving random rewiring of the connectome trains to the same numbers.