TinyBrains has you train a small neural net, export it to ONNX plus a manifest, and upload it to play strategy games against everyone else’s — starting with Ants, an homage to the 2011 Google AI Challenge, where the author placed #127. Weight classes run like boxing divisions — nano, micro, mini, small, large — set by the byte size of your model file, starting at 8 KiB; nano’s baseline is 2,930 parameters in 5.9 KiB. The engine bans floating-point outright, so replays reproduce bit-for-bit — free.