Path: utzoo!utgpu!jarvis.csri.toronto.edu!clyde.concordia.ca!uunet!mcsun!ukc!edcastle!aiai!jeff From: jeff@aiai.ed.ac.uk (Jeff Dalton) Newsgroups: comp.ai Subject: Re: Recursive Searles, or what? Keywords: understanding Message-ID: <1518@skye.ed.ac.uk> Date: 9 Jan 90 16:22:45 GMT References: <12679@phoenix.Princeton.EDU> <12702@phoenix.Princeton.EDU> <7661@sdcsvax.UCSD.Edu> Reply-To: jeff@aiai.UUCP (Jeff Dalton) Organization: AIAI, University of Edinburgh, Scotland Lines: 125 In article <7661@sdcsvax.UCSD.Edu> pluto@beowulf.UCSD.EDU (Mark E. P. Plutowski) writes: >I'm sorry to belabor the point, but I still am not convinced after >months of patiently following this argument. What follows is not >a counter, but rather is intended to be an honest desire to clarify >my 'understanding' of the concept to a level enabling me to not >only say I understand, but prove it to some objective logician. I think you're asking for too much. There are few, if any, things outside mathematics that could be proved to an objective logician. >We believe Searle-in-the-box when he claims to not-understand. We don't have to believe him. But I think you know what it would normally mean to say you understand Chinese. (I know that I don't understand Chinese, for example, and I don't need to be able to articulate a definition of "understand" to know this. I suspect you are in a similar position.) Now suppose you start doing the things one does in the Chinese Room. You might well think that the room (ie, the system) understood Chinese, and that you had some part in it. But would you really be in doubt about whether your prior claim not to understand Chinese was still correct? I'm pretty sure that running the Chinese Room wouldn't cause _me_ to understand Chinese, at least not right away. The whole point of the C.R., I think, is to take advantage of our admittedly informal and imprecise notion of "understand Chinese". That way, the whole question of what "understanding" in general really means doesn't have to come up right away. The system reply (ie, "the system understands") aims to make the question of whether the person in the room understands Chinese irrelevant, thus forcing us back to the more general, difficult questions about understanding. Searle tries to get back to whether the person understands by his argument about internalizing the system. In short, Searle is trying throughout to avoid getting tangled up in the definition game and other philosophical problems. Well, I don't know if this was really an aim on his part, but I do think it's a good way to look at it. Dennett calls Searle's C.R. argument an "intuition pump", and I think that's a good term for it. It isn't meant to be a water-tight logical argument. Turing took a similar approach in his famous "test": he wanted to get at the question "can a machine think?" in a way that didn't require answers to all kinds of difficult or controversial issues. >Why? Don't we have an objective definition of "understanding?" Probably not. >Why don't we have an objective definition of understanding? It's too hard. Suppose Searle had given a definition. [I'm supposing he didn't give one, but maybe he did. I haven't looked at his _Intentionality_ or the like for a while.] Think of all the ways it could go wrong. For example, he has to make sure that the definition doesn't beg the question by automatically making one or another answer true. Other people might accuse him of that anyway. After all, no one has to accept _Searle's_ definition. It's always open to argue the "understanding" ought to be defined some other way and that defining it Searle's way makes it too hard for machines. A definition might be useful for getting a better idea of what Searle's talking about, but I don't think a dispute about how "understanding" should be defined (a definition game) will tell us what we want to know about machines. There's a danger that it would end up being a debate about whether we should call what computers do "understanding" rather than about whether computers are interestingly different from humans. >Why do we have to ask Searle whether or not he understands, in order >to ourselves decide whether or not he understands? We don't. Imagine yourself in his place. You don't have to agree with his answer, although (as I argued above) I think you would. Maybe you agree with him that far (ie, that the person in the Room doesn't understand Chinese), but still think he hasn't properly answered the systems reply. Or maybe for you the Chinese Room just doesn't convince at all. >How can a proof be based upon a definition which is subjective to >an element of the proof itself? It doesn't aim to prove but to convince. It isn't a proof. >If we ask the Chinese Room itself whether it understands, >(in Chinese, of course, so Searle-in-the-box will not understand >the query) and by some fluke in the rules it somehow answers "Yes" >... Now what? Why not believe the box's reply? That's "the system understands" again. We don't have to believe the box any more than we have to believe Searle. For example, we might have good reason to decide the box was "faking it". I'm not sure what such a good reason would be, but here's something that might do it. Suppose we start up the box for the very first time and ask "what were you doing yesterday?" and it starts telling us all kinds of things that we know it didn't do at all. >I am not biased against Searle's basic point: (IMHO) >that the implementation matters. I just don't see the logic >behind the counter to the systems reply, and would like to >hear it expressed in objective terms. I don't find Searle's argument that he could internalize the rules, etc., so that the system was inside him, and still not understand all that convincing. He's still looking at the system from the CPU's point of view (so to speak), which doesn't seem like the right place to see any understanding that might be taking place. In any case, what I think Searle's Chinese Room example indicates is not so much that no understanding is taking place as that it might not be taking place. Behavior isn't enough evidence for understanding on its own. Whether executing the right program is all there is to understanding (the "strong ai" position Searle is trying to refute) is, in my view, still an open question. >Prove to me the Chinese Room does not understand, without >resorting to asking Searle-inside whether (or not) he does. Note that Searle does have another argument, namely the one about syntax not being sufficient for semantics. Jeff Dalton, JANET: J.Dalton@uk.ac.ed AI Applications Institute, ARPA: J.Dalton%uk.ac.ed@nsfnet-relay.ac.uk Edinburgh University. UUCP: ...!ukc!ed.ac.uk!J.Dalton