Some background for reasoning about dual-use alignment research — LessWrong