AI helpers often don’t know new users’ tastes and can’t keep up when those tastes change.
Large language models usually learn by guessing the next word, then get a tiny bit of instruction-following practice; this paper flips that by turning massive web documents into instruction-and-answer pairs at huge scale.