One might still insist on the point that Harkness raises, namely that, even if FEP is not particularly sketchy when stripped of empirical content, it is really only an account of what the system should do, rather than what it actually does. There is of course some truth to this, since the mathematical formulation of FEP is an idealization of a system engaged in variational Bayes.
However, perhaps FEP is in a peculiar functionalist category. Its starting point, as I mentioned earlier, is the trivial truth that organisms exist, from which it follows that they must be acting to maintain themselves in a limited set of states, from which it in turn follows that they must be reducing uncertainty about their model. Thus the function described by FEP is not about what the system should or ought to be doing but about what it must be doing, given the contingent fact that it exists.
This starting point differs from commonsense functionalism because it is not based on conceptual analysis but is instead based on a basic observation, plus statistical notions. It also differs from empirical functionalisms (cf. psychofunctionalism) because it does not specify functional roles in terms of proximal input–output profiles for particular creatures. Neither are the functional roles it sets out defined in terms of teleologically-defined proper functions (cf. teleosemantics), except in so far as it could be said that the proper function of an organism is to exist.
This category of functionalism, which I dubbed “biofunctionalism”, seems intriguingly different from other kinds of functionalism. It provides a foundational functional role, which must be realized in living organisms, and from which more specific processes can be derived (for perception, action, attention etc.). This differs from austere functionalisms, which only say how things ought to be working, and it differs from fully mechanistic functionalisms, which specify how particular types of things actually work.