Comment by jeremyjh

I wouldn't use the term "token predictor" or "statistical pattern matcher" to refer to a post-trained instruct model. Technically that is still what it is doing at a low level, but the reward function is so different - the updates its making to weights are not about frequency distribution at all.