Along the same lines, it's often useful to think of an LLM's "understanding" a question or a context. Sometimes responses from smaller models show a shallower "understanding" of the question than you'd get from a bigger model.

Obviously, we shouldn't anthropomorphize too much here, and even the most powerful LLMs don't "understand" or "reason" or "think" the way humans do. But whatever they're doing, it's at least analogous to what we do. These concepts are genuinely useful for making better use of these tools.