Path: utzoo!attcan!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: Why the argument against program's relevance is important Message-ID: <596@ntpdvp1.UUCP> Date: 28 Jun 90 03:20:07 GMT References: <3285@usceast.UUCP> <2837@skye.ed.ac.uk> <25422@cs.yale.edu> <965@idunno.Princeton.EDU> Organization: SNA Solutions Inc., Contract Programming Group Lines: 120 In article <965@idunno.Princeton.EDU>, markv@gauss.Princeton.EDU (Mark VandeWettering) writes: > In article <593@ntpdvp1.UUCP> kenp@ntpdvp1.UUCP (Ken Presting) writes: > > > >Searle gets a lot of heat for using loose language, but there are > >important cases where he says just what he means, in no uncertain > >terms. > > Searle's language is CRIMINALLY loose. Concepts such as understanding, > causal powers, the distinction between syntax and semantics are > not ever defined in any paper of his that I have read. In particular, > his recent Scientific American article was not a "proof": he merely > assumed that his conclusion was correct and proceeded. Don't bother with the places where his language is loose. If we are serious about wanting to defeat this argument, we need to look at those few places where he is clear. We are all familiar with the process of "desk debugging" - simulating a program by hand to trace it's operation. Searle is not in a vacuum at UC Berkeley. He has talked to *plenty* of programmers, and knows about hand simulation. The CR does NOT get its staying power from the vagueness. This tar baby is sticky because it's based on everyday common sense. A programmer can runs through the steps of his program just as well (if not as fast) as a chip. And once you've "learned" a foreign language, you can do a lot more than just generate replies to written notes. But we need a stronger solvent than common sense. Let's talk logic: > >For example, he says that the CR example itself is presented only > >to show that semantics is not reducible to syntax. > > ... which has NOT been shown at all ... Searle doesn't really need to argue for this - he just wants a compelling example. Tarski has already proved it, and it is quite beyond debate. Tarksi's theorem states that no predicate P can satisfy the following criterion for al sentences S: P('S') is provable if and only if 'S' is true. Note that this criterion does not involve the provability of S, but only the provability of a sentence *about* S. Truth is a semantic property, while any "P" in the theorem is a syntactic property (because it takes a quoted sentence as its operand), so Tarski shows that truth cannot be reduced to any syntactic property. Pat Hayes has correctly pointed out that programs are more than syntax, which is very important. A single floppy disk with my program on it is physically different from the same disk with your program on it, and running the different programs will produce physically distinguishable output. This issue is important for the question of whether programmed computers have any specific "causal powers," but is independent of the question of understanding. > Searle is trying to prove the following: > > > For any program P whatsoever, and for any machine M whatsoever, > > the following inference is always invalid: > > > Machine M runs Program P, therefore Machine M understands. > > > . . . Searle is *not* trying to show that no program can > >think, or that no machine can think. He is too clever for that. He > >is not attacking the *goal* of Strong AI. > > Interesting distinction, but why would ever use the fact that program A > causes "understanding"? The only really valid test for understanding is > demonstrating it. > Suppose you want to sell your latest "Conversational Chinese" program. Wouldn't you like to claim that running your program will make the computer speak Chinese? It's a question of whether the way to make an AI is by writing programs or by building machines (or some combination). Strong AI says (according to Searle, and I think he is not far off) that given the right program, any machine that is big enough to run it will actually be intelligent, while the program is running. I think this claim is true, but Searle does not. He thinks something else must be said about the machine before we can conclude that it understands. Perhaps you agree with Searle? > >He is attacking the *argument* behind Strong AI. This is much easier > >to do, but almost as devastating. Write any program you want, and > >run it any way you want, on any hardware, parallel, serial, or cerebral. > >But if you want to claim that the system is thinking, you'll need a > >better reason than "It's running my program". > > Searle's Chinese Room was designed to attack the concept of the Turing test > as a valid test for intelligence. It simply fails. To Searle, it makes > no sense to say that the Chinese Room _understands_ Chinese. This is simply handwaving. Look - when I'm in my office, my office could pass the Turing test in English, but it still makes no sense to say that my office understands English. If you want to claim that Searle defies common sense in *his* argument, then you better not defy common sense in your *own*! Go back to the logic of the problem. Take any programmed Turing machine that can pass the Turing test. This machine can do nothing more than implement a syntactic algorithm, since the input tape can never contain anything other than a string of symbols. Therefore, by Tarski's theorem, (and Church's thesis) the machine could represent a syntactic predicate but not a semantic predicate, such as truth. Now, *whatever* it is that constitutes understanding, it must have something to do with knowing what words mean, and that certainly requires semantics. We have just seen that no Turing machine can represent the concept of "truth", so there is at least one word which no Turing machine can understand. This is a different conclusion than Searle wanted, but that's because I have used slightly different premises (getting Searle's original conclusion would require a more general concept of "semantic information"). I hope we can put an end to the vague language of both sides, in this thread. Ken Presting ("Let us calculate")