Another paper trying to rein in the chaos of large language models, because what could possibly go wrong with letting them just learn whatever they want, right?