An introduction to worst-case AI safety — LessWrong