I think this is assuming that the model doesn't get better through additional training or feedback, but it seems to me that reward/punishment is actually what d

I think this is assuming that the model doesn't get better through additional training or feedback, but it seems to me that reward/punishment is actually what drives better prediction over the medium/long run. Right now we are definitely in a place where the models are over confident but I think with time the self correction will become very good.