I wanted to see if decisions models could keep up with a FPS. It turns out they can and even small models were more capable then than I thought they'd be at a task that is likely out of training distribution.

This is libre quake, kev and laya compiled as wasm