The first AI probably won't be very smart — LessWrong