Arena

Push the crates into the glowing ring. Simple — except there is no physics engine on this page. The ball you drive, the crates it shoves, every bounce and every collision: all of it is predicted, step by step, by two tiny neural networks (a ball net, and one crate net shared by every crate — entities conditioned on each other’s relative state, the same decomposition idea as multi-subject world models). three.js only draws what the nets decide.

Drive with WASD / arrow keys or the buttons. The nets were trained on a couple of minutes of simulated pushing — so the “physics” you feel is whatever they managed to learn from watching it. Each net also predicts its own motion gate: a learned “am I actually moving?” bit that holds resting objects still. (Without it, a residual bias far below the training-loss floor compounds step by step until every crate slowly conveyor-belts itself into a corner — regression can’t learn exactly zero, but a tiny net is great at learning a binary switch.)

Training script is here — same hand-rolled backprop as the /play/ demo, no ML libraries.