"Aargh!" he said out loud in real life. David, are you disagreeing with me here or do you honestly not understand what I'm getting at?

The whole idea is that an agent can fully understand, model, predict, manipulate, and derive all relevant facts that could affect which actions lead to how many paperclips, regarding happiness, without having a pleasure-pain architecture. I don't have a paperclipping architecture but this doesn't stop me from modeling and understanding paperclipping architectures.

The paperclipper can model and predict an agent (you) that (a) operates on a pleasure-pain architecture and (b) has a self-model consisting of introspectively opaque elements which actually contain internally coded instructions for your brain to experience or want certain things (e.g. happiness). The paperclipper can fully understand how your workspace is modeling happiness and know exactly how much you would want happiness and why you write papers about the apparent ineffability of happiness, without being happy itself or at all sympathetic toward you. It will experience no future surprise on comprehending these things, because it already knows them. It doesn't have any object-level brain circuits that can carry out the introspectively opaque instructions-to-David's-brain that your own qualia encode, so it has never "experienced" what you "experience". You could somewhat arbitrarily define this as a lack of knowledge, in defiance of the usual correspondence theory of truth, and despite the usual idea that knowledge is being able to narrow down possible states of the universe. In which case, symmetrically under this odd definition, you will never be said to "know" what it feels like to be a sentient paperclip maximizer or you would yourself be compelled to make paperclips above all else, for that is the internal instruction of that quale.

But if you take knowledge in the powerful-intelligence-relevant sense where to accurately represent the universe is to narrow down its possible states under some correspondence theory of truth, and to well model is to be able to efficiently predict, then I am not barred from understanding how the paperclip maximizer works by virtue of not having any internal instructions which tell me to only make paperclips, and it's not barred by its lack of pleasure-pain architecture from fully representing and efficiently reasoning about the exact cognitive architecture which makes you want to be happy and write sentences about the ineffable compellingness of happiness. There is nothing left for it to understand. This is also the only sort of "knowledge" or "understanding" that would inevitably be implied by Bayesian updating. So inventing a more exotic definition of "knowledge" which requires having completely modified your entire cognitive architecture just so that you can natively and non-sandboxed-ly obey the introspectively-opaque brain-instructions aka qualia of another agent with completely different goals, is not the sort of predictive knowledge you get just by running a powerful self-improving agent trying to better manipulate the world. You can't say, "But it will surely discover..."

I know that when you imagine this it feels like the paperclipper doesn't truly know happiness, but that's because, as an act of imagination, you're imagining the paperclipper without that introspectively-opaque brain-instructing model-element that you model as happiness, the modeled memory of which is your model of what "knowing happiness" feels like. And because the actual content and interpretation of these brain-instructions are introspectively opaque to you, you can't imagine anything except the quale itself that you imagine to constitute understanding of the quale, just as you can't imagine any configuration of mere atoms that seem to add up to a quale within your mental workspace. That's why people write papers about the hard problem of consciousness in the first place.

Even if you don't believe my exact account of the details, someone ought to be able to imagine that something like this, as soon as you actually knew how things were made of parts and could fully diagram out exactly what was going on in your own mind when you talked about happiness, would be true - that you would be able to efficiently manipulate models of it and predict anything predictable, without having the same cognitive architecture yourself, because you could break it into pieces and model the pieces. And if you can't fully credit that, you at least shouldn't be confident that it doesn't work that way, when you know you don't know why happiness feels so ineffably compelling!