Notes on "the hot mess theory of AI misalignment" — LessWrong