The impressive part is that it is low latency enough to serve high quality answers at game speed through the model instead of a pre generated ad-hoc machine.