I'm getting obsessed by the idea of a neural network that just learns from the "mistakes".
Great, but how would it relate to online dice?
Let's put a supercomputer with a neural link to work and give him a dice site to beat, where all it can do is click "roll". Oh, I lost, roll again or not? Let's roll, Oh, I won, roll or not?
All I can foresee is that computer getting addicted and stuck playing dice forever