-I _also_ had more to say about philosophy of categories: I thought I could be more rigorous about the difference between "caring about predicting different variables" and "caring about consequences", in a way that Eliezer would _have_ to understand even if Scott didn't. (Scott had claimed that he could use gerrymandered categories and still be just as good at making predictions—but that's just not true if we're talking about the _internal_ use of categories as a [cognitive algorithm](https://www.lesswrong.com/posts/HcCpvYLoSFP4iAqSz/rationality-appreciating-cognitive-algorithms), rather than mere verbal behavior: it's always easy to _say_ "_X_ is a _Y_" for arbitrary _X_ and _Y_ if the stakes demand it, but if you're _actually_ using that concept of _Y_ internally, that does have effects on your world-model.)
+Moreover, you _can't_ just give an orc a clever misinterpretation of an oath and have them believe it. First you have to [cripple their _general_ ability](https://www.lesswrong.com/posts/XTWkjCJScy2GFAgDt/dark-side-epistemology) to correctly interpret oaths, for the same reason that you can't get someone to believe that 2+2=5 without crippling their _general_ ability to do arithmetic. We weren't talking about a little "white lie" that the listener will never get to see falsified (like telling someone their dead dog is in heaven); the orcs _already know_ the text of the oath, and you have to break their ability to _understand_ it. Are you willing to permanently damage an orc's ability to reason, in order to save them pain? For some sufficiently large amount of pain, surely. But this isn't a choice to make lightly—and the choices people make to satisfy their own consciences, don't always line up with the volition of their alleged beneficiaries. We think we can lie to save others from pain, without ourselves _wanting to be lied to_. But behind the veil of ignorance, it's the same choice!
+
+I _also_ had more to say about philosophy of categories: I thought I could be more rigorous about the difference between "caring about predicting different variables" and "caring about consequences", in a way that Eliezer would _have_ to understand even if Scott didn't. (Scott had claimed that he could use gerrymandered categories and still be just as good at making predictions—but that's just not true if we're talking about the _internal_ use of categories as a [cognitive algorithm](https://www.lesswrong.com/posts/HcCpvYLoSFP4iAqSz/rationality-appreciating-cognitive-algorithms), rather than mere verbal behavior: it's always easy to _say_ "_X_ is a _Y_" for arbitrary _X_ and _Y_ if the stakes demand it, but that's not the same thing as using that concept of _Y_ internally as part of your world-model.)