What is it like doing AI safety work? — LessWrong