It's predicting the next token. The best answers, online, mostly come from polite discourse. It's not a big leap to think manufacturing politeness will yield better answers from a machine.