It looks like the forum ate a couple of posts in this thread, including my own. Oh, well.
I've been mulling over my own interpretation of Hoffman's work, and I don't think the implications are that bad. So, in any theory of perception, the basic problem is interpreting a stream of raw information being received by the mind. These could be neural spikes coming from the sensory organs, as materialism assumes, or it could be an actual stream of raw information coming through the agent-network, as conscious realism assumes (as would absolute idealism, transcendental idealism, information realism, etc.).
Since the 'camera model' has been ruled out by optical illusions, the sophistication of the visual system, etc., the usual assumption is that the human perceptual system is meant to accurately model the environment. When our perceptions don't match with objective (or, really, inter-subjective) reality, it's attributed to our imperfect sense organs (colour vs. EM spectra), our finite cognitive capabilities (change blindness), our lack of complete physical knowledge (quantum mechanics), or the mis-application of heuristics that usually work (optical illusions). Hoffman argues that we've mis-understood what perception is about. It's meant to model the
fitness function of the environment using the
least resources. (Note that he found that the evolutionary advantage of interface strategies did not depend on organism complexity.) This means ignoring some information, blurring over meaningful but irrelevant differences, and otherwise running roughshod over the truth. However, we don't have instinctive knowledge of the (Darwinian) fitness function, or else everyone would be popping out babies to the expense of all other concerns. All you have are some pleasure and pain instincts that are vaguely related. Also, we know that sensory input is required for perception to develop properly, so the interface does not come pre-packaged. It needs to be
learned to some extent.
How do you get perception to match the fitness function, then? This is where I'd diverge from Hoffman's desktop interface analogy. What you want is a set of species-specific
developmental biases that constrain the form of the organism's models of the environment to a subset that correlates well with the fitness function. For an analogy, if you want to match a numerical sequence to some pattern, you can restrict your candidate patterns to (say) polynomials. You need to do something like that, anyway, because if you don't restrict the space of possible sensory interpretations, you can run into the
poverty of the stimulus; also, there's an infinity of useless interpretations that ought to be ruled out (e.g. the whole world is painted on the back of your eyelids). Thus, our perception of a 3+1D space-time continuum containing concrete objects viewed through the five senses is simply part of this instinctive package, which forces us to fit everything into a certain framework (shades of transcendental idealism). It may be a very powerful framework that can account for almost every eventuality, but the map is not the territory, and you can do pretty well without a deep understanding of reality. For this reason, we shouldn't be surprised when (meta)physical theories based on concrete observation turn out to miss the mark under close examination. It's happened before with classical mechanics; will materialism do any better?