Chase

Collect the orbs. Don’t get caught. The red hunter chasing you has no scripted AI at all: every half-second it imagines nine possible futures — rolling candidate moves through the same two tiny neural nets that run this page’s physics — scores them by whether they end with you caught, and takes the best one. That’s model-predictive control, live in your browser, and the glowing threads are its actual imagination: every future it considered, the brightest being the one it chose.

You’re faster than it is. It can only win by prediction — it chases where the model says you’re going, not where you are. Cut corners, feint, use your speed. And you can borrow the same crystal ball: hold Shift (or the Premonition button) to see the model’s dream of your own next second.

As on /arena/, there is no physics engine here — the ball you drive, the hunter, every wall bounce is the output of two ~2,000-parameter MLPs trained offline by hand-rolled backprop, with learned motion gates holding still things still (chase_train.js, no ML libraries). The hunter’s cunning is not in the nets: it emerges from searching them.