Only a locally trained neural model steers this fly. Watch it change lanes, leap low barriers and approaching carts, and flatten beneath gates. The deliberately spacious course makes learning visible.
A random 14-input, 16-state GRU predicts left, right, jump, duck, or stay every 0.1 simulated seconds. Generated teacher demonstrations and current observations train weighted categorical cross-entropy with analytic backpropagation, 24-step truncated sequences, Adam, and global gradient clipping at 1.5. This is supervised sequence learning, not full-episode reinforcement learning. Crashes restart the world and recurrent state while keeping learned weights. Reset starts learning from scratch.
The warm anatomical brain is illustrative: 12 compartments, 1,536 somata, and 3,106 connected synapses. Its functional rings show actual inputs, reset gates, update gates, candidate state, hidden state, and five action probabilities. Packet brightness follows neural contributions; signals travel on drawn paths at 3.5× the original packet speed. These are signal packets, not literal electrons or a biological connectome. The easy course does not require memory, though the trained recurrent paths affect decisions.
Speed changes world time. Drag the brain to orbit it. There are no steering controls, sound, camera, microphone, uploads, external calls, or downloaded model weights. Training stays in this tab; hidden tabs pause. Collider volumes follow the fly's chitin body and stance; wings, antennae and fine legs are cosmetic.