Path: utzoo!censor!jeff From: jeff@censor.UUCP (Jeff Hunter) Newsgroups: comp.ai Subject: Re: Question on Chinese Room Argument Summary: fair enough, I understand you better Message-ID: <399@censor.UUCP> Date: 10 Mar 89 08:11:29 GMT References: <4298@pt.cs.cmu.edu> Organization: Bell Canada, Business Development, Toronto Lines: 105 ... harnad@elbereth.rutgers.edu (Stevan Harnad) writes: > [me condensed. assume an atom-by-atom model of a human is sufficient rules > to reproduce linguistic behaviour. now "record" atoms of a real human.] > > This is followed by a long list of arbitrary variations and permutations > on Searle's simple Chinese Room -- including real people, atomic > copies, synthetic copies and neural networks, as well as symbolic and > human "simulations" of all the foregoing, in unspecified languages -- > none of which seem to elucidate anything. I'm sorry if the examples seemed arbitrary. I was listing some alternative "copies"of the original human, from flesh-and-blood to a purely mental simulation. I was unsure where you would draw the line and say that understanding ceased. > I am asked which ones can > understand English. The simple answer is that those that can pass the > TTT (including the English LTT) can. Atomic and synthetic robotic > copies could do that in principle (so what?); mere symbol-crunchers > (including symbolic simulations of atomic copies, neural nets, etc.) > cannot, for a number of reasons, among them Searle's Chinese Room > Argument. A clear answer. Thank you. I went back and re-read your un-expired articles here. Let me try my understanding (that word again) of your position. "A purely symbolic simulation is, by definition, not real. A system which has (at least some) real parts is immune to Searl's Chinese Room argument, and thus can potentially understand. A system which can pass the Total Turing Test (simulate a real human) can probably understand." Please correct me if I've mis-stated your opinions. I'm curious as to the bounds you put on the TTT. Does a candidate have to look exactly human even under X-rays, etc... or does it just have to be able to pass the LTT, look vaguely humanoid, and be able to pick up a glass? An interesting property of any of my "symbolic simulations of atomic copies" is that it is possible, in principle, to reconstitute a living human from the symbolic information of the position of the atoms. This new human seems as good a candidate to pass the TTT as any random man on the street, and so can "understand" with the best of them. (You may at this point say "of course, so what?". Please don't get annoyed. I'm just trying to fathom your concept of understanding, which I find strangely counter-intuitive.) Do you agree that a re-embodied simulation can understand? Now let us take a non-Chinese-speaking human, and record her. She goes in a sealed room, with only an optical fibre link to the rest of the world. (Note: purely symbols in, symbols out.) Her simulation is put in a similar (but simulated) room. Both of them are taught Chinese (by correspondence, telephone, video, etc) until the real human understands Chinese, and the simulated one has simulated understanding. To an external observer they should be as alike as twins. There is no way that I can see to distinguish which is real without opening the room (or probing it with cosmic rays). (Remember this is a thought experiment with an arbitrarily good simulation at atomic level.) Now open the real room and let the volunteer out. Reconstitute the simulated room, open it, and let the second copy of the volunteer out. Presumably both can pass the TTT. Each will state that they understood Chinese while they were in the room, so the introspective "I know when I understand something as opposed to just manipulating meaningless symbols" property of understanding seems to be common to both real and simulated humans. I can't see any functional difference, inside or out, between the real and simulated understanding. Do you see any difference aside from the fact that the simulated one is purely non-physical? Next thought experiment: take a human. record the brain, then remove it (and store it, who knows when a spare brain will come in handy :-). add hardware to each nerve to record incoming signals, send outgoing ones, and mask the truncation of the nerve. add more hardware to remove glucose from incoming blood, etc.. add a room containing a miniaturized ultra-quick Searl running an atomic level simulation of the recorded brain. The hardware communicates signals with Searl using some real actuator (a series of laser pulses perhaps). The result would seem to be a human, and should be able to pass the TTT. Do you agree? Oh it wouldn't get concussions due to brain bruises. Add a few acceleration sensors. We're back in business.) Now if the original human spoke Chinese we have: Searle does not understand Chinese. You repeatedly ridicule the notion that "Searle + rules" can understand Chinese. "Searl + rules + laser" and "Searl + rules + laser + interface hardware" might or might not understand (I don't know where you stand on this). "Searl + rules + laser + interface hardware + body" can pass the TTT, and therefore you should believe that this can understand. Do you? I find it hard to believe that adding a few peripherals to the processor (I'm a programmer don't cha know :-) magically adds understanding somehow. Please try to explain again. Thanks. > > Burying Searle's simple, straightforward point in a labyrinth of arbitrary > complications serves neither to understand it nor to refute it. > Well it's so simple that dozens of messages later there still is no clear agreement on what you see as the difference between real and simulated understanding. I'm hoping to accelerate the process a bit. .... keep smiling ... -- ___ __ __ {utzoo,lsuc}!censor!jeff (416-595-2705) / / /) / ) -- my opinions -- -/ _ -/- /- No one born with a mouth and a need is innocent. (__/ (/_/ _/_ Greg Bear