On the future of language models — LessWrong