He asserts that we can prove that it is better for so-and-so to perceive according to "fitness" rather than according to "truth", and maintains this is true with mathematical rigor; fine.
This is observably true (allowing me the possibility of observation for the space of this conversation) - predators, for example, usually have forward-facing eyes and vertically-slitted pupils, which allows them to focus and react to motion of prey in front of them; prey often have widely-spaced eyes allowing a greater field of view. Fitness literally makes you see the world differently.
However, there is an important dimension that I think he missed (or several), which is that the fitness function governing a trait is a vast, multi-dimensional space.
For example, using color we may distinguish between unripe, hard-to-eat fruit (green) and ripe, healthy fruit (reddish-orange). According to his argument, we should therefore only need to see these two colors. Of course this is incorrect, because if our vision were so limited we'd just be running into trees all the time. Now our fitness function is spread across two, completely unrelated goals - suddenly, it becomes much more attractive (and simpler) to come close to apprehending reality accurately, thereby killing two birds with one stone. It only gets worse as the number of goals we add to our fitness function for perception increases.
Furthermore: the space of "goals" is itself a function of the genetic diversity in your population. A species with a lot of spare genetic diversity lying around will be able to refine traits with selection much more efficiently. This means we can develop all sorts of wonderful toys (like a face-matcher) that can augment our perception of reality.
Of course, in the end, we're just modeling reality, we can't Know it in some Buddhist sense; but there is good reason to believe that perception should lean heavily towards a simple objective representation of reality, and that fitness is not such a close rein on it all the time.