How Powerful AIs Get Cheap — LessWrong