Shit, I forgot about this, sorry. Off the top of my head: CF is an umbrella term so I'm mostly attacking the common perception which is of course awful. Someone like Chalmers does a lot better though still falls horribly short IMO. Kanai's work has recently come into my attention and he does a great job of noticing many of the same issues, though so far hasn't quite addressed them (slim odds IMO). Ironically enough I'm very sympathetic to it for **anything other than sentience**. For example, KORs are better understood through some computational role, namely as control rods for processing of novel information, which is why they show up in sensory cortex, what makes no damn sense if you take the "they encode le bad valence". My main issues come well before any empirical question, I'm skeptical of CF at the level of transcendental constitution of its putative object, namely "the computation". Quid iuris? What licenses one to speak of such unconditionally? Does the world offer an objective, canonical factorization between hardware and software, for example? Some of the issues seem to be inherited, others unnecessary, others innate. An example of something inherited is that classically, (A <-> B) V (B <-> C) V (C <-> A) Means that up to equivalence you only get two propositions, so qualities like white get shunted into predicates white(x), which means a semantic domain with some property class that requires independent justification. But then because of Joyal's lemma "a cartesian closed category with a dualizing object is a preorder" your proof theory and hom-set are fucked since you get a trivial semantics for computation or proofs. So on the one hand the naive theory is too fine at the level of identity, and too coarse at the level of individuation of primitives. That is of course what we see in practice. Some posited "functional organization" that no one ever actually directly describes, and no satisfactory account of the qualitative at all, hardly even so much as attempts. This is why the thing is sort of a shitshow methodologically and how you get this sort of empirical-argumentative gacha where arguments and properties are posited à la tête du client. Now for less pointed, but pertinent, scattershot problems. Consider arxiv.org/abs/1608.05972v8 Even in the functionalist's effectively ideal setting of pure mathematics of algorithmic information theory with lossless redescription of a constructed determinate object, redescription is not without issue. The thing is that even though the object may be reconstructed without fault, the surface representation may change and thereby alter measures and interactions, which are representation-dependent. But we're talking brains at least, which are recurrent, non-terminating, self-modifying, realtime, so the object may itself be parametrically coupled with specific representations, representation is not by default some neutral encoding layer. There is no canonical "the same computation". Proof identity is in fact an open problem. Therefore the basic putative object of CF is not quite well-posed. That's before we even get to anything like substrate-independence or how people try to patch CF into something slightly more positive. Factorizing between algorithm and substrate requires justification. Personally I prefer instead to think along opposite lines, of computation-admissible dynamics. You have some physical process that given some standards admits to a computational description. This IMO leads to a much more interesting lens and account of things. Substrate independence itself is an operational regime. There is a vast industrial discipline of informational and instrumental impedance matching before anything resembling a reliable abstract computation shows up, and even then we don't typically guarantee the absence of side channels and hardware vulnerabilities. "The same computation" in addition to specifying no equivalence class, can conceal *inequivalent* demands, much as in the Zenil case cited above. One-tape versus multitape machines, Turing-machine steps versus beta reduction, call-by-name vs call-by-value. There's no "I/O behavior" previous to choice of model of computation and evaluation regime, and those will have different effects. Furthermore, what is the admissible set of transformations, specially of intermediate states? This one is almost never supplied. Duplicate a subcomputation? Share it? Memoize it? Serialize parallel activity? Stretch its timing? Replace a sensor reading with an internally generated value? There are myriad routine shortcuts and modifications in computer science and programming. Scope is left undefined. CF schizzes out between computationalism and functionalism and seems to never really decide between the two. It makes me think of that one meme with Darth Vader and the water pitcher. The dynamics of computation sneaks through the back what the abstraction of functionalism quotients out. Retarded. Rejecting lookup tables and requiring counterfactual organization makes the process matter. Appealing subsequently to “the algorithm,” an I/O relation, or a produced normal form can erase distinctions that the earlier criterion made relevant. Functional roles themselves can be materially constituted. Preflexes is the phenomenon that the visco-elastic properties of muscle are there to pre-empt motor control stabilization because nerve transduction is too slow for it, dissolving part of the problem at the level of format through distributed computation instead of a centralized coordination. If you then go on to admit complexity into the picture due to the demands of relative timing, a generic isomorphism for computation is completely fucked as an option. The sufficiency of structuralism for phenomenology is extremely questionable and has been called out many times under many guises, ironically enough. Newman's objection, Benacerraf, etc. This is unsurprising given the logical difficulty I highlighted previously. Anyway I could keep going on for a long time (going into specific arguments, e.g. the problem with the counterfactual patch, etc.) but this is already pretty rambly. To me the most basic problem, that is, what I would have to be convinced is incorrectly posed to even consider changing camps, is that the qualitative is locative, not alethic. You don't ask if white in itself is true or false, it's either there or not. It's content that's not really amenable to addressable description. It's what gets addressed, the occupant.