Any real toeholds for making practical decisions regarding AI safety? — LessWrong