"Conventional" here I believe are those approaches before large foundation models.

A hidden concept here is Occam's razor. In many cases, we somehow come up with parsimonious theories of the generative rule behind the data. This will (I think) be discussed in the following chapters.

This claim needs context. Consider Figure 1.1: it illustrates what machine learning typically calls few‑shot learning—inferring a task from a small number of labeled examples and then generalizing to new instances of the same task. The crucial question is whether such generalization truly happens without relying on previous experience.

hello

heell