Thanks to @Syderas for bringing this point up and doing a pretty good job at summarising it. I think Negarestani's point as it regards AI is intriguing and thought-provoking, but, more importantly, it stems from a broader and far more profound principle.
I feel like the concept of a 'toy model' is probably pretty familiar for the academically-inclined, but maybe a little elaboration is in order. A toy model is a deliberately-simplified or abstracted model that is used for the purposes of making first-order estimations or calibrating intuitions. The physicist's joke about spherical cows is a humorous exaggeration of this practice (and its shortcomings): tl;dr, if you want to know how aerodynamic a cow is, it's easier to just think of it as approximately resembling a sphere, because then the math becomes a lot simpler. Most physics classrooms ignore air resistance for similar reasons, assuming that everything exists in a vacuum.
In my experience though, toy models are far more prevalent in economics. It's variously accepted (at least, imo, by good schools), that the vaunted 'perfectly competitive market' doesn't exist — it's an ideal type, a set of mathematical relations that are basically never fully manifest in reality, in the same way things that we identify as circles basically never have circumferences of exactly 2πr. All other economic models are similar constructions: contingent on assumptions and hacks and spherical cows, which 'bake-in' limitations and deviations from reality.
As it is said, all models are wrong (but some are useful). Or, in Korzybski's gloss, the map is not the territory, and a 1:1 map is not one of those useful models. There seems to be a fundamental tradeoff between total/absolute specificity and comprehensibility.
(Consequently, all of our physical models can be 'wrong', too. It's easy to see the failings of Newtonian physics, because it breaks down at the level of quantum physics — but this doesn't mean that quantum physics itself has no blind spots.)
I think the received interpretation of these dicta is humbling, and well-understood by the rationalist community over at LessWrong, incidentally, but I also think it somewhat boringly transforms empirical study into a practice of model-refinement and calibration. This can feel pretty grim, clinical, and depressing. I think the profound thing that Negarestani is highlighting here is that even our wrong models can be useful; combined with the right truth-seeking kind of attitude, they give us contrapositive indication of what a 'right' model looks like. We can get to a fuller picture of what 'intelligence' is by proposing (inherently limited) models of human intelligence, and then seeing how artificial intelligences stack up against it. We need to experiment, and 'get it wrong' sometimes, in order to know how to get it right.
Where is the link to memes and meme studies? Well, I sometimes find myself getting frustrated by the oft-abstruse and overcomplicated systematic and formal descriptions of memetic behaviour and practice (sorry, but the Jreg thread makes me a tiny bit irate — it feels like the String Theory of irony). But, my gripes with these kinds of models, if I can articulate them, can actually be important and useful stepping-stones to a more-integrated, easier-to-comprehend schema. As Syderas glosses, our 'views' (models) need to be "open to change and redefinition", and continual refinement. We shouldn't be afraid of putting forward new concepts and models in experimental spirit, but we also need to be ready to 'kill our darlings', or otherwise abandon them, if something else (including a future version) is more fit-for-purpose.
Are memes themselves toy models? Maybe, but I would say only insofar as a meme encodes or represents a specific "model of the world".