Comment by dnautics

> outside of explicit training

what does this mean anymore? i can invent a proof assistant language that was not in the training set and the llm does a fantastic job with it.

(github.com/ityonemo/bpa)

well it learned some sort of programming language and some sort of logic

well can you not see that those sorts of analogies can effectively make nothing in the knowable universe out of distribution? if you did such a categorization for human learning you could likewise say, "humans never can do anything outside of training set" too.