Comment by YeGoblynQueenne

>> This ignores of recent work with large language models that do generalize, zero-shot, to novel tasks.

Which work is that?