> In contrast language models are trained over trillions of tokens comprising the entirety of human knowledge. Not even close! At best it's a small subset of t

> In contrast language models are trained over trillions of tokens comprising the entirety of human knowledge. Not even close! At best it's a small subset of the internet + published books. The vast majority of human knowledge isn't even in the training sets yet. I would question the use of a model fed everything, though.