This paper (https://arxiv.org/abs/2206.07682) also touches on a pretty fascinating phenomenon - that when scaling up large language models they seem to "naturally" obtain new emergent abilities that do not exist on smaller models.