> Large language models are impressive statistical text predictors — genuinely useful tools that excel at pattern matching and interpolation.
Slightly OT: It's interesting how many (smart!) people in tech like the author of this article still can't conceptualize the difference between training objective and learned capability. I wonder at this point if it's a sort of willful ignorance adopted as a psychological protection mechanism. I wonder if they're going to experience a moment of severe shock, just gradually forget that they held these opinions, or take on a sort of delusional belief that AI can't do XYZ despite all mounting evidence to the contrary.
replies(2):