Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!rutgers!mcnc!rti!ntpdvp1!kenp From: kenp@ntpdvp1.UUCP (Ken Presting) Newsgroups: comp.ai Subject: Re: No more Chinese rooms, please? Summary: The Chinese Room puzzle cannot be solved by Computer Science alone Message-ID: <605@ntpdvp1.UUCP> Date: 17 Jul 90 21:22:16 GMT References: <25422@cs.yale.edu> <593@ntpdvp1.UUCP> <31329@cup.portal.com> <50741@iuvax.cs.indiana.edu> Organization: SNA Solutions Inc., Contract Programming Group Lines: 142 In article <50741@iuvax.cs.indiana.edu>, dave@cogsci.indiana.edu (David Chalmers) writes: > > Talk of "machines" tends only to confuse the issue, anyway. All we need > is the notion of *program* (a formal object), and *implementation of program* > (a physical system). It's not clear that all implementations will be > describable as running on pre-existing machines. > > In this framework, the strong AI claim becomes: > > EXISTS P (program) such that FORALL S (physical system): > S is an implementation of P => S is intelligent. > This version of the Strong AI claim is anticipated by Searle, on p.29 of the Sci.Am. article: The thesis of Strong AI is that any system whatsoever ... not only might have thoughts and feelings, but _must_ have thoughts and feelings, provided only that it implements the right program, with the right inputs and outputs. Notice two (small) differences: 1) Searle is insistent on the issue of programs "constituting" minds, and physical objects "causing" thought. So the arrow above must not be read as simple material implication - it must be necessary implication, or entailment, or some other counterfactual (eg causal). 2) As I read him, Searle is lumping together all automata which compute the same function, independently of the algorithm they use. If I follow Daryl McCullough's last article, he for one would not accept this version of Strong AI. Probably few would. I myself balk at (2), because I believe a thinking thing must contain a representation of the concept of "truth", which cannot be finitely defined in I/O terms. Searle is perfectly willing to face Strong AI defined in terms of "implemented systems", but there is probably a difference between his concept of implementation and Dave's. The usual software engineering concept of "implemented system" involves: a) an independently pre-existing "machine" which can run most any "program" b) a "machine-readable" copy of a program c) a mechanical process for "loading" the program into the machine d) a user-initiated process of "running" the program. Searle seems to view "implementation" in this fashion, which I'll call GOFR, for "Good Old Fashioned Running" (apologies to John Haugeland). Note that Hilary Putnam claims to have shown that every physical system is an *instantiation* of every finite automaton. Unless Dave's concept of "implemented system" can be distinguished from Putnam's "instantiated automata", Searle could object that Dave's version of Strong AI implies panpsychism. "Go ahead," Searle would say, "write your magic program. Now find a system that *doesn't* implement it, or else explain why all these implementations lying around on the ground still act so stupid." GOFR is not subject to Putnam's argument, because to identify an object as a "machine" requires the concurrent specification of the processes of loading and running the program, and a coding scheme for machine- readable copy. (See _Representation and Reality_ for the argument) This is the first problem with the Systems Reply - the high powered abstract concept of implementation reduces Strong AI to an absurdity, while the traditional GOFR concept does not neatly excise the "pre-existing machine" and its gripes about not understanding its data. > > Anyway, with this in place, Searle needs to show > > FORALL P, EXISTS S such that S is an implementation of P but S does not > produce intelligence, > > which is what the Chinese Room purports to show. Of course it doesn't show > that, but that's another story. Suffice to reiterate the often-made point > that the fact that the pre-existing machine (i.e. the person in the room) > that implements the program fails to understand is quite irrelevant. > Implementing machines aren't what counts: implemented systems are. Let me try to give a formal analogue of this objection, in the forlorn hope of clarifying the issue once and for all. Let U(n,m) be the function computed by a Universal Turing Machine, where 'n' is the Goedel number of an arbitrary TM, and 'm' is the Goedel number of an arbitrary starting configuration of an input tape. Let C(m) be the function computed by a TM that, when implemented, can pass the Turing Test in Chinese. Let 'c' be the Goedel number of some TM that computes C(m). Finally, suppose for a moment that we have made sense of the concept of "implementation", and let S and R be implements. Note first that on any acceptable concept of "implementation": For any S, S implements U(c,m) if and only if S implements C(m). Now, according to Systems Repliers, all that Searle shows is: There are S, R such that S implements U(n,m) and R implements U(c,m) and S does not think. And of course, this is irrelevant. We only care whether R thinks. If I finally got it straight, this is the gist of Bob Kohout's "Viola" (:-) article last spring. NOBODY, NOT EVEN SEARLE, IS THAT STUPID. The Putnam-based objection I gave above is never raised by Searle, because he has a more straightforward counter on p. 30 of the Sci.Am. article: The point of the original argument was that symbol shuffling by itself does not give any access to the meanings of the symbols. But this is as much true of the whole room as it is of the person inside. Searle's point is that there is nothing a Universal TM can do with a program that he cannot do just as well himself. When he says "But I still don't understand Chinese", he is not just reporting a subjective state of ignorance. He is correctly emphasizing that the syntactically specified operations he is performing on the symbols are unrelated to their semantics. Everyone agrees that Searle-without-books not understanding is completely irrelevant. The controversial issues are: Problem 1: How much do we have to add to Searle in order to get an "implemented system", and what "causal powers" will that system have? Problem 2: Once we have an "implemented system", what connection, if any, is there between the operations of the system and the semantics of the symbols? There are (min) two distinct threads in the CR debate. One is not an issue for AI at all - Problem 1. General Computer Science should be able to handle that issue, but not until the semantics of programs is understood and the concept of implementation is cleared up. Problem 2 is specific to AI, but it can only be studied if we make some general assumptions about how symbols get their meanings. IMO, we can get these assumptions from Quine and Davidson, so there's no need to think in a vacuum. Ken Presting ("Anybody else abhor a vacuum?")