Path: utzoo!utgpu!utstat!jarvis.csri.toronto.edu!mailrus!tut.cis.ohio-state.edu!rutgers!elbereth.rutgers.edu!harnad From: harnad@elbereth.rutgers.edu (Stevan Harnad) Newsgroups: comp.ai Subject: Re: Question on Chinese Room Argument Summary: On Tool-Making vs. Mind-Modeling Message-ID: Date: 21 Feb 89 01:43:22 GMT References: <4298@pt.cs.cmu.edu> <4296@cs.Buffalo.EDU> Organization: Rutgers Univ., New Brunswick, N.J. Lines: 61 sher@sunybcs.uucp (David Sher) of SUNY/Buffalo Computer Science, in a very revealing posting, asks: " what is the advantage of a machine with "understanding"? Assume that " HAL doesn't understand anything. He merely manipulates symbols so that " he creates an illusion of understanding in his correspondents. In what " way does that inhibit HAL as a useful tool? What could an " "understanding" machine [do] that a merely intelligent (the symbol " manipulator that merely gets the right answer) machine could not? " Unless someone can show me an advantage to it I'm not going to waste " much time designing it into my programs. There is no advantage to worrying about understanding if all you are interested in doing is making "useful tools" -- which is no doubt all that most of AI is interested in. One wonders, though, why a discipline with that motivation tries to push so hard on the repeatedly discredited "Systems Reply" to Searle, insisting that "The System" DOES understand, when the real goal is as superficial as this. Perhaps there is a confusion here between tool-making and mind-modeling. Cognitive psychologists, on the other hand, are interested in modeling the mind, including understanding, so we have no choice but to face the questions Searle (and the mind/body problem and the other-minds problem) raise. Searle's Argument simply shows that purely symbolic models are the wrong ones for our purposes. [Paradoxically, my own work suggests that even cognitive psychologists should not worry too much about capturing understanding: I have given reasons -- empirical, methodological and logical -- for adopting "methodological epiphenomenalism" and the "Total Turing Test (robotic version)" as constraints on cognitive modeling. However, these same reasons also go strongly against symbolic modeling in favor of hybrid modeling, grounding symbolic representations bottom-up in nonsymbolic (analog and categorical) representations.] Two other points: (1) You've got the assumption on the wrong foot: The default assumption is that HAL doesn't understand, not the other way round. You don't have to say "Assume Hal doesn't understand" any more than you have to say "Assume there are no fairies." The default "assumption" is no, unless compelling reasons are given for rejecting it. No compelling (or even coherent) reasons are coming from symbolic AI, and certainly not from proponents of the "Systems Reply." (2) Unless you are willing to think deeply on these questions you certainly ARE wasting your time "designing it [?]" into your programs! One of the reasons I think it's important to get these matters straight is because if you don't, you spend more time over-interpreting what your models are doing than in actually strengthening their performance capacity. This is a deep and subtle point. The Total Turing Test is the methodological goal. Hermeneutics and hyperbole about the "mental powers" of toy models is not the way to get there; it's just a way of covering up how pathetically far away from the goal we really are. -- Stevan Harnad INTERNET: harnad@confidence.princeton.edu harnad@princeton.edu srh@flash.bellcore.com harnad@elbereth.rutgers.edu harnad@princeton.uucp BITNET: harnad@pucc.bitnet CSNET: harnad%princeton.edu@relay.cs.net (609)-921-7771