.. But the house always wins ?

Right? I hope that at some point the weights system will become clear enough that one could bias against wrong answers. Just “telling” it that is wrong biases for the current token cache and doesn’t back propagate to the model. Of course back propagating user input would be very exploitable so I don’t see that happening.

Quite possibly. They are subsidized right now, but those things are really expensive.