@darius that's a lot of words and jargon to say "imitation learners shouldn't extrapolate, but they do"

…all while missing Emily (&c)'s point that the models don't actually "know" *anything* , they just produce statistically probably sequences