Searle (John Searle (1997), The Mystery of Consciousness, p9) says “A computer is by definition a device that manipulates formal symbols”. What does he mean? How do you define a piece of electronic hardware (rather than, say, describe it)? Or is he defining the term "computer" to mean a formal-symbol manipulating device? Is he making an empirical statement (that computers by their fundamental nature necessarily manipulate formal symbols and only formal symbols) - which might be true or false - or is he merely saying that from now on when he uses the word "computer" he is referring to formal-symbol manipulating devices? Or does he mean something else?
Yes, he (correctly) means "a formal-symbol manipulating device". But "manipulation" needs a little clarification. The usual textbook formalism is given by the lambda calculus, e.g., https://en.wikipedia.org/wiki/Lambda_calculus#Formal_definition (google "lambda calculus" for lots more).
But electronic devices natively implement only a much simpler (architecture-dependent) machine language, something typically more-or-less like Knuth's MIX, e.g., https://en.wikipedia.org/wiki/MIX (google "Knuth MIX").
However, a lambda-calculus-interpreter can always be written in any such machine language. Indeed, "Church's Thesis" and "Turing Completeness" (more terms for you to google) guarantee that just about any sensible formal idea of "manipulation" is ultimately equivalent to all others.
That is, there's a class of so-called "computable functions" (google that), meaning that a sequence of symbols representing input (the function's argument) can be manipulated into another sequence representing output (the function's value). If no such manipulations exist, the function's not computable (google "halting function" for a non-computable example). And all computer languages, i.e., all formal ideas of "manipulation", can ultimately calculate exactly the same class of computable functions.
So Searle's ultimately and equivalently saying that consciousness isn't (can't be simulated by) computable functions. But your comment, "...defining a computer as a symbol-manipulating device just seems to add fog to the AI landscape", seems to conflate AI with consciousness. AI, aka "expert systems", are certainly computable, but certainly not consciousness. And nobody ever claims any such thing. I think maybe your question arises from a misunderstanding about that.
Edit... Hmm, now taking a look at your profile, I wouldn't imagine you'd have any confusion/misunderstanding about that. But then I don't see how you'd have any question about Searle's remark whatsoever. So what's your question, more exactly?
Edit (reply to Roddus' comment below)... Firstly, for concreteness, let's please do away with this unnecessarily vague "voltage level" terminology, which you've used here and in preceding comments above. See, e.g., https://en.wikipedia.org/wiki/Bit#Physical_representation for the correspondence between bits and voltage levels. We're talking about Searle's "symbols" (sequences of bits), regardless of their physical representation, which simply happens to be voltage levels in electronic digital computers.
Regarding "meaning", with respect to computers, I'd guess what you might want to study could be "denotational semantics" (e.g., https://en.wikipedia.org/wiki/Denotational_semantics and many other google hits) and "domain theory" (e.g., https://en.wikipedia.org/wiki/Domain_theory ). Then it's the so-called "semantic function", which maps syntax (the denotation represented by strings of symbols) to semantics (the syntax's meaning represented in a so-called Scott domain), that captures the mathematical idea of "meaning" as it pertains to computability. The short paper https://www.cs.colorado.edu/~bec/courses/csci5535/reading/densem.pdf seems like a pretty good intro to me (I'm not googling any wikipedia-type stuff that comes anywhere near adequate). A longer-but-more-comprehensive intro is (seems to me) http://homepage.divms.uiowa.edu/~slonnegr/plf/Book/Chapter9.pdf And/or try googling "semantic function" denotational (put "semantic function" in quotes as shown, followed by denotational) for additional tutorial papers.
As for Searle and his Chinese Room conclusions, you'd need to compare and contrast "meaning" with respect to computers, versus "meaning" with respect to consciousness. But only the former is well-enough-defined for any rigorous comparison. And if you're really interested, I think you might need to further study domains, maybe particularly the idea of approximation as represented by their poset ordering, whereby "meaning" can start out vague and become better-and-better defined with more-and-more syntax. But that would involve a pretty significant effort, far afield from any direct relation to "consciousness". So I wouldn't recommend it unless you're really, really interested.
The theoretical computer science aspects might (again judging from your profile) be adequately interesting, but maybe not the related-to-consciousness aspects... Are computers (i.e., can computers simulate) "conscious"? Or can they exhibit behaviour indistinguishable from consciousness? Searle apparently says "no". Whether or not his argument's conclusive is maybe debatable, but his definition of "computer" is entirely adequate. Any further argument would have to discuss the ultimate capabilities of "symbol manipulation" -- just how far can that take you? And that's indeed somewhat of an open question. It's closed with respect to computable function theory, but meaning/semantics/etc not equally well closed.
I believe that Searle, like almost everyone in computing agrees fully with Turing's definition of mechanical computation. By "computer" Searle is speaking to any mechanical computational device (think Enigma machine and other "computers" that would destroy that dumb apple commercial if they saw it) that has a very simple function. While "manipulates formal symbols" isn't the best wording it is definitely not wrong; as a software engineer who took CIS at university I can't disagree with the simplification; in fact on a subtle level the recognition of base level symbols is quite astute on Searle's part (eg. low level vs high level programming languages, lex vs parse, etc)
I believe that in general Searle idea of "a computer" is looser and would include all things that would be considered Turing machines. From the school of names perspective using term "computer" as the inclusive term would be much easier but to would require the explanatory statement in the late 90s
To obtain an answer, one must first make a distinction between a "computer" and its underlying "implementation." Let me use two examples to make this clear.
IBM 370 ------------ transistors
Human brain------- neurons
Both, IBM 370 and a Human brain are capable of manipulating symbols, so it's clear that a "computer" is capable of manipulating symbols, whereas, an "implementation" is not. Therefore, it is sensible, to define a computer as a symbol-manipulating device. As to what was Searle's purpose for doing so, I would guess that it lets the reader know that when he uses the term "computer," he means its capability, not its implementation.
Thanks John. When you say:
1. SYMBOLS. "...for concreteness, let's please do away with this unnecessarily vague "voltage level" terminology, which you've used here and in preceding comments above. See, e.g., https://en.wikipedia.org/wiki/Bit#Physical_representation for the correspondence between bits and voltage levels. We're talking about Searle's "symbols" (sequences of bits), regardless of their physical representation, which simply happens to be voltage levels in electronic digital computers."
I disagree about bits. To talk of a sequence of bits is to use a computational abstraction. I see the physical representation itself as the key to understanding the semantics of the machine (in Searle's sense of semantics – i.e., intentionality). You could take this further and say that computation presupposes an extrinsic semantics (and that the concept of computation needs to be abandoned before trying to understand intrinsic semantics).
It seems important to take what john Searle does as the CPU in the Chinese room, and then say, OK what is the exact equivalent in an actual electronic digital computer? Not in an abstract machine, not in a Turing machine, but in the actual piece of hardware on a desk.
In the Chinese room, cards inscribed with Chinese ideograms drop through the slot in the door. Unknown to Searle these are sensible Chinese questions (he knows no Chinese). He perceives the shapes inscribed on the cards, and manipulates the cards on the basis of the shapes (presumably the rule book - the program - contains examples of the shapes, but Searle also says the rules describe the shapes). Searle reacts to the shape. But unknown to Searle, a meaning has been assigned to the shape by people outside the room. Searle has no access to the meaning, but only to the shape.
The way external people assign a meaning to a shape is to first perceive the shape then go through a mental process of assigning meaning (learning the meaning, and there's more than one way to do this). This learning can be regarded as creating instances of a 2-term relation. One term is the shape, the other the meaning.
So, what's the exact equivalent in the electronic digital computer, if the Chinese room accurately reflects the essentials of the computer? The CPU receives clocked voltage levels. Unknown to the CPU, the clocked voltage levels have been assigned meanings by people outside the computer.
Well, of course this is ridiculous. No such meanings have been assigned, nor could they ever be. External people can perceive shapes and assign meanings to them (thus creating instances of the 2-term relation), but humans are biologically incapable of perceiving clocked voltage levels, so cannot assign meanings to them. Cannot create instances of the 2-term relation, the first term of which would be the clocked voltage level.
Hence, the Chinese room does not - semantically speaking – accurately reflect what happens with computers. Semantics is the whole point of the CRA. Searle has failed to properly understand computer processing from the semantic perspective.
This, above, is just a starting point for a detailed comparative examination of the Chinese room verses electronic computers. ( I think various other things are wrong with the Chinese room, too.)
2. MEANING. You say “As for Searle and his Chinese Room conclusions, you'd need to compare and contrast "meaning" with respect to computers, versus "meaning" with respect to consciousness. But only the former is well-enough-defined for any rigorous comparison.“
The history of AI has also been a history of re-defining mental terms to make it seem as though computers have mental properties (when they don't, at least not when running the proffered programs). Minsky was one of the masters of the fine art of academic re-definition in order to get students and funding. In his incredibly influential early book, Semantic Information Processing (he and his graduate students were contributors) the programs he presented have zero semantic content. He even (with wonderful spin) indicates this: “...one cannot help being astonished at how far they [the programs in the book] did get with their feeble semantic endowment.” They actually had zero semantic endowment.
I agree that a severe problem is that the mind is not understood, the higher-level functions of the brain are not understood. Maybe re-defining mental terms using Computer Science terminology (so the re-defined concepts are capable of being realized in a computer) seems the only option. But to rebut the CRA (which is the goal of really carefully examining the Chinese room) the attempted rebuttal needs to use the concepts Searle uses in his arguments and in his descriptions of the room. Appeals to a definition of meaning that Searle does not use is appropriate. The least needed is a convincing translation of his meaning into a Computer Science one (which would be a reduction of meaning in Searle's sense to meaning in the Computer Science sense).
3. DEFINITION OF COMPUTER. You say, “Whether or not his argument's conclusive is maybe debatable, but his definition of "computer" is entirely adequate.”
Searle defines the computer as a universal Turing machine. However, a Turing machine processes things that have extrinsic meanings (0,1, the various shapes reacted to by the universal machine as described in Turing's 1936 paper), but electronic computers don't. The things electronic computers process don't have any semantics. That's the key point. Turing machines are said to be purely computational entities. Computation presupposes an extrinsic semantics. Electronic computers, in the sense of processing things that lack an extrinsic semantics, are hence not computational. If this (really radical) view is adopted, then it probably has relevance to the validity or soundness of the Chinese room argument.
4. SYMBOL MANIPULATION. You say: “Any further argument would have to discuss the ultimate capabilities of "symbol manipulation" -- just how far can that take you? And that's indeed somewhat of an open question.”
I claim that electronic computers don't process symbols in Searle's sense of “symbol”. If this is right, then the idea of symbol manipulation (in Searle's sense of symbol manipulation, which is the Turing machine sense) is inadequate for fully understanding what computers do and could do.
From the start of the Computer Age, electronic computers have been understood using the concepts of computation (hence the name of the device). But what if these concepts are not adequate when it comes to trying to understand how an electronic computer (so named) could think? What if thinking is fundamentally non-computational? And what if, when the electronic computer is understood with different concepts, it becomes clear how a computer could perform the needed non-computational operations of intelligence?
In asking the question "Does it make sense to define a computer as a symbol-manipulating device?", I was asking whether Searle was trying to force down our throats the (false) idea that electronic computers processes objects that have an extrinsic semantics. These objects being symbols is Searle's sense of "symbol". That computers manipulate symbols and only symbols (in Searle's sense of "symbol") is a premiss of various versions of the CRA. By defining computers as symbol-manipulating devices, he seems to be trying to prevent any discussion of the question: Well, do computers process symbols (in Searle's sense of "symbol")? If they don't, then a CRA premiss is false and the argument is unsound.
As you say, the difference between extrinsic (observer-based) and intrinsic (subjective) symbol-manipulation semantics is key to the Chinese Room argument (CRA) (cf. the symbol grounding problem). However, I don't see that the abstract-vs-physical distinction between computation theory and physical computing devices addresses this issue directly.
Computation theory is just another mathematical model, like calculus or geometry. Finding logical "0, 1" symbols in a physical computer is no different than finding straight lines and 90-degree angles in buildings. If we believe that the abstract, mathematically based theories of engineering truly keep skyscrapers from falling and planes in the air, then the dancing voltages in my bank's computer also truly represent the logical "balance" of my logical "checking account". Physical computers "manipulate symbols" to the same degree that buildings "obey statics" and planes, "aerodynamics".
Fortunately, the distinction between idealized math and the real world is not needed to fully understand the CRA. Because it is grounded by Turing machine (TM) theory, the CRA's analysis can be much less esoteric and much more precise and objective.
The CRA expertly focuses everyone's attention on the wrong symbols: the Chinese inputs and outputs. Like a magician, Searle makes you ignore the elephant in the Chinese Room: the program. It, too, is a symbolic input to the Searle-computer. (He also hopes you'll miss the fact that it's in English!)
The program (aka, the rule book) is the key symbolic input because it alone dictates how the Chinese symbols are processed. The only reason the Searle-computer can (and must) process the Chinese symbols purely formally is because he also has the program-rule symbols, which he can (and must) interpret non-formally. As a universal TM, the Searle-computer's main responsibility centers on the program itself. Searle directs everyone's attention solely to himself and the Chinese symbols, yet he is wholly superfluous to their computation!
The Searle-computer and its program input could--and should--both be completely removed from consideration. Replace it by a direct, non-programmable implementation of the program and the room would function identically. Hence, Searle's claim that the room processes the Chinese symbols purely formally (as opposed to just its Searle-CPU) is totally unfounded because he fails to account for the program's own computation (whose existence is irrefutably explained by Turing machine theory).
The degree to which the program's computation interprets the Chinese semantically (or not) remains undetermined. Hence, the CRA establishes absolutely nothing about the semantics in the room regarding the Chinese symbols.
OK for Phil-132, I'm not sure I understand all the points, but this is what I think's going on in the Chinese Room. There are two separate symbol-processing systems: (1) the system that processes the Chinese symbols that enter the room from outside, and (2) the system that processes the symbols that compose the program. The room also have a set of basic operations (like the Turing machine's Scan, Print, Erase, Left and Right). In an actual computer, these are built into the hardware. In the Chinese room, they come from Searle understanding what the program symbols mean.
Program symbols define the sequence of these simple operations once the program begins to execute. The simple operations then manipulate the Chinese symbols (and do other things). The program can be replaced by wiring. In this case the sequence of simple operations are not defined by program symbols, but by the wiring. But the "program-as-wiring" still has to be able to treat different Chinese input symbols in different ways. So how is it going to do this? Somehow the shapes of all possible Chinese symbols have to be built into the wiring. But the only way to do this is for the wiring to contain examples of Chinese symbols (which examples can then be matched by the wiring against the ones that come into the room from outside).
So the wiring has to contain symbols. Even though the simple operations can be executed by wiring instead of program symbols, there must still be literals, there must still be examples of all possible Chinese symbols, that the simple operations can use to match against the actual Chinese ones that arrive as input.
This matching process conducted by the wiring is purely formal because all it does is compare the shapes of the pre-existing example Chinese symbols to the shapes of the ones that enter the room as input. "Getting rid of the program" only gets rid of the program symbols that trigger simple operations. It doesn't get rid of the literals contained in the program. So for the program line: IF INPUT = "A" THEN GO DO , getting rid of the program gets rid of: IF INPUT = "" THEN GO DO , but it doesn't get rid of A.