Aaronson does not critique the chinese room argument at all, at least not in the paper cited.
"Where others invoked complexity considerations to argue with Searle about the metaphysical question, I’m invoking them to argue with Penrose about the practical question."
Aaronson is not arguing with Searle at all, he is using Searle's argument as an example of other peoples faulty thinking about complexity. Aaronson understands (which is why he doesn't elucidate Searle's argument) that Searle is talking about meaning, and Aaronson is criticizing other peoples critics of Searle that are based in computation for failing to understand complexity:
"I find this response to Searle extremely interesting—since if correct, it suggests that the distinction between polynomial and exponential complexity has metaphysical significance. According to this response, an exponential-sized lookup table that passed the Turing Test would not be sentient (or conscious, intelligent, self-aware, etc.), but a polynomially-bounded program with exactly the same input/output behavior would be sentient."
More to the point, Aaronson doesn't address the meaning (or as he says, the metaphysical) questions at all. He is interested in the complexity problem of producing a machine that passes the Turing Test, and how philosophers don't seem to grok that very practical problem. Searle recognizes the practical problem for what it is (lookup tables can be Turing complete) and talks about meaning and asks us to consider where the meaning of things are, and shows that meaning does not exist in the functions or the data the functions process. So that even if a machine passes the Turing Test, it fools the observer. The machine still would not be "intelligent"; it would not be conscious.
Also, my reading of that passage by Aaronson is very different from yours. I read it as him saying, these computer scientists have put forward a fairly serious and compelling set of arguments that the actual complexity class of the translator algorithm has philosophical importance. Meanwhile, the analytical philosophy response is just to say something hand-wavy about "meaning."
I agree Aaronson's not saying anything definitive (he is extremely conservative about making definitive claims, despite how hotly involved he becomes in the few cases when he does make definitive claims). But I don't agree that he is framing it to raise questions about the validity of the CS rebuttal papers that he cites.
I think Aaronson demolishes the other critics arguments because he shows they focus on the lookup table and attach sentience to algorithmically complex solutions but not to exponentially complex solutions. My point is the lookup table is irrelevant, in practical terms, because the lookup table in Searle's argument exists only as a "philosophical fiction" as Aaronson says. But I was pointing out that lookup tables can be Turing complete. And hence any Turing machine could be substituted for the mechanism in Searles's room and thus the particular mechanism of the room's operation is irrelevant. (in any kind of Turing completeness sense)
I took Aaronson as being humorous here: "Yet, as much as that criterion for sentience flatters my complexity-theoretic pride, I find myself reluctant to take a position on such a weighty matter." Because there is no obvious reason an algorithmically complex solution should somehow be sentient when an exponentially complex solution should not be. How could the lower mathematical bound confer sentience?
Aaronson's paper is about the practical requirements to pass a Turing test in some given amount of time. It is a testable problem. Searle's argument is about what it means to produce actual sentience. Aaronson does not really get into this.
There is an argument against Searle along the lines of "what are the requirements for a machine which passes the Turing test for Searle." And Searle's response to these practicalities are weak, at best. But those arguments have nothing to do with Searle's point in the Chinese Room. Aaronson sort of reflects those critiques of Searle, but he also realizes the hand-wavy problem of meaning is something he doesn't address.
Personally, I get very frustrated when people mistake the problem of sentience for the testable hypothesis of a Turing test (or any of the other "practical" problems). I think the problem of sentience is a real problem, and it requires a practicable solution to produce machine sentience, machines which have and understand meaning. So arguments against Searle's Room that do not address how to instantiate meaning in a computer system are disappointing because they ignore his basic point. (Aaronson is making arguments about complexity and Turing tests) Ignoring the key problem is not a critique of that problem. And critiquing an argument is not necessarily a critique of the point or concept the argument elucidates.
Meaning is a real thing.
If you sit down to make a machine conscious, you have to deal with what awareness is and how meaning and representation work- at the very beginning. And then figure out how to make computers do representational processing and instantiate awareness.
All of the modern approaches abandon the problem of actual sentience and the problems of meaning; because, they are hard. Or it's too hard to finish in the timeline of a PhD. So people do the reverse, start with the algorithms and solve a testable sub-problem and make some practical progress in computer science or in industry. (which is a good thing!)
Nearly everyone abandons the hard problem of meaning and how meaning works and chooses to solve a different problem. This does not mean our solutions to those other problems are solutions to the hand-wavy problem of meaning. It rather makes me think of people who figured how to make fake feathers and then assumed the process of making fake feathers will naturally lead to human flight.
I think this is a clue that the typical computer science approach, which has made great progress in what we call artificial intelligence, is maybe the wrong approach to solve the sentience problem. Not that computer science is irrelevant, but that the general computer science approach simply does not provide a path toward, or the theoretical foundation, to make computers which are aware and can generate and understand meaning.
I know that. I was asking about your claim that a lookup table can be Turing complete, as opposed to merely able to pass a given Turing test. A lookup table does not express universal computation, at least not on its own. For instance, you couldn't use a fixed-size lookup table (with fixed, immutable content) to express some floating point number of precision N, where N is so high that even just the bits required to represent it outnumbers the available space in the lookup table. You'd need something "outside of" the lookup table that could mutate its memory, or hold state of its own while reading from the lookup table. The complexity class of doing any of this would be irrelevant to Turing completeness.
On the topic of the complexity threshold for sentience, I think the reasons are extremely simple and obvious. If you take a Turing test to involve convincing a human being of something (as it was originally conceived), then there are inherent time constraints to performing actions. If you can adequately respond to any question posed in English, but it takes you longer than the age of the universe to do so, then as far as anyone is concerned, you actually can't do it.
Imagine that one day we create an artificial being which we all agree is sentient. Suppose, without being too precise, that the entity "runs" in some amount of time -- like it can answer verbal questions in matters of minutes, it can solve complicated problems in matters of hours or days, and so on.
Now suppose we can play around with the insides of that being and we have some sort of dial that lets us "slow it down." We click the dial once, and the being now takes hours to make verbal responses, and takes weeks to solve complex problems. We click the dial again and it takes weeks to make a verbal response, and months to solve a complex problems... and so on.
I think it is reasonable to believe that after enough clicks the entity is not sentient even if the underlying software is conceptually similar and just differs by controlling its complexity class.
I'm not saying I agree that this definitely defines sentience. I'm just saying it's extremely reasonable that it might and it's not at all obvious to me that an exponentially slow entity, if we could wait long enough to verify, is "just as sentient" as an entity that is verifiably more efficient.
Even further, the particular argument that I pointed out actually used estimated numbers to come up with space bounds, not time bounds. Basically, the argument was that even if you truncate to considering merely 5-minute-long English conversations (or shorter), then creating a lookup table to be able to look up responses is physically impossible -- who even cares how long it would take to actually use it if you could create it -- you can't even create it, and so it is somewhat fallacious to even begin a thought experiment by saying "suppose there is this big lookup table that physics logically excludes from possibility." It's just not even useful, not even as a thought experiment, because it doesn't speak about what is possible, from first principles.
You touch on a very interesting point in your reply here:
"You'd need something "outside of" the lookup table that could mutate its memory, or hold state of its own while reading from the lookup table. The complexity class of doing any of this would be irrelevant to Turing completeness."
I don't know if this is irrelevant to Turing completeness... I suspect you are right, but I'm not sure how to write a proof of that.
However, it is the basic problem I have faced writing code that can make representations (write it's own programs). My solution requires abandoning the symbolic route completely. We still do computations, and the computations interact and change each other, like how molecules interact and change each other. But it's a whole level down from the symbolic computation problem you bring up.
I tried doing cellular automata that had interacting mechanisms to alter their rule sets, which isn't that far from neural nets that alter their functions. And I tried approaching neural nets that could alter their structure based on their responses to input. Not merely back propagation of values, but back propagation of structure. But like all computation, it's the problem of syntax and semantics all over again!
I just decided to go down a level and make no associations between syntax and semantics at all, and instead build up a syntactically homeostatic system, that would hopefully then be able to create semantic behavior by creating "semantic" structures from intrinsic syntactic behavior. So, my approach is not "outside of" but rather from the "inside out".
If you have any suggestions about how to code a solution to the "outside of" problem, in any kind of symbolic system, I would be very interested in your ideas. [that would be some cool voodoo!]
Cellular automata are lookup tables, and Wolfram and others proved some cellular automata rules are Turing complete computers. https://en.wikipedia.org/wiki/Rule_110 My point was merely about the equivalence of computational mechanisms, not about lookup tables per se. And by corollary, that the computational complexity is equivalent regardless of the computational mechanism. (I think we agree on this point.)
Searle's Room is a just to explain that what computers are doing is syntactic.
Searle would posit that passing a Turing test in any amount of time is irrelevant to determining consciousness. It's a story we hypothetical use to "measure" intelligence, but it's only a story. it's not a valid test for sentience, and passing such a test would not confer sentience. Sentience is an entirely different question.
What would be more interesting is if a computer intentionally failed turing tests because it thinks turing tests are stupid.
We could test humans with Turing tests to determine their "level" of intelligence. But if you put teenage boys in a room and made them do Turing tests, pretty quick they would come up with remarkable ways to fail those tests, chiefly by not doing them! How could you write a program or create a system that intentionally fails Turing tests? or a program which avoids taking turing tests... because it thinks they are stupid?
Could you write a program that knows when it fails? (it's a pretty long standing problem...)
I like the speed (or space-bound question) you ask because it is not a thought experiment to me. It's an actual real problem I face! at what point does the speed of the underlying computing become so interminably slow that we say something is no longer sentient? In my work, I don't think there is some such slow speed. The slowness simply obscures sentience from our observation.
In the excellent example: "I think it is reasonable to believe that after enough clicks the entity is not sentient..."
How would you distinguish between the "loss" of sentience from reduced complexity, from the loss of your ability to perceive sentience from the reduced complexity? The question is, how could you tell which thing happened? If you don't observe sentience anymore, does that mean it's not there? (Locked in syndrome is similar to this problem in human beings.) And if you have a process to determine sentience, how do you prove your process is correct in all cases?
I do not think of these as rhetorical questions. I actually would like a decent way to approach these problems, because I can see that I will be hitting them if the model I am using works to produce homeostatic metabolic like behavior with code.
Computation is a subset of thinking. There is lots of thinking that is not computation. Errors are a classic example. The apprehension of an error is a representational process, and computation is a representational process. We may do a perfectly correct computation, but then realize the computation itself is the error. (As a programmer learns, it is exactly these realizations that lead to higher levels of abstraction and optimization.)
Searle's point is that a lookup table or any other computational mechanism, can not directly produce sentience because it's behavior is purely syntactic. "Syntax is not semantics and simulation is not duplication." https://www.youtube.com/watch?v=rHKwIYsPXLg
Aaronson's points are very well made, but none of them deal with the problem of semantics or meaning. Because they don't deal with what representation is and how representation itself works. All of the complexity work is about a sub-class of representations that operate with certain constraints. They are not about how representation itself works.
> "suppose there is this big lookup table that physics logically excludes from possibility."... That is the point!
Even if there were such a lookup table, it would not get us to sentience, because it's operations are syntactic. It is functional, but not meaningful. You are correct, it could never work in practice, but it could also never work under absolute conditions. That's why I figured Aaronson was poking fun of those critiquing Searle, because it would ALSO, not work in practice.
Aaronson writes, "I find this response to Searle extremely interesting—since if correct, it suggests that the distinction between polynomial and exponential complexity has metaphysical significance. According to this response, an exponential-sized lookup table that passed the Turing Test would not be sentient (or conscious, intelligent, self-aware, etc.), but a polynomially-bounded program with exactly the same input/output behavior would be sentient."
This statement supports Searle's argument, it doesn't detract from it. Hypothetically, an instantaneous lookup of an exponential table system would not be sentient but an instantaneous lookup of an algorithmically bound table system would be sentient? On what basis then does sentience confer, if the bound is the only difference between the lookup tables? Introducing the physical constraints doesn't change the hypothetical problem.
Searle and Aaronson are just talking about different things.
If Aaronson was actually refuting Searle, what is the refutation he makes?
Aaronson never says something like "Computers will be sentient by doing x, y, and z, and this refutes Searle." The arguments against Searle (which I take Aaronson as poking at) are based in computation. So... show me the code! Nobody has written code to do semantic processing because they don't know how. It could be no one knows how because it's impossible to do semantic processing with computation - directly.
That is my view from repeated failures, there simply is no path to semantics from symbolic computation. And if there is, it's strange voodoo!
People always like to say this about Matthew Cook's result on Rule 110 and connect it to Searle's argument, but they are just totally different things. If Searle instead talked about encoding a general purpose AI program to translate sentences, and his substrate of computation happened to be a cellular automata, that's fine, but it would be no different than him postulating an imaginary C++ program that translates the sentences, meaning he would be assuming a solution to A.I. completeness from the start, whether it is via cellular automata or some typical programming language or whatever.
But the type of lookup table he is talking about is just an ordinary hash table, it's just a physical store of fixed immutable data which is not interpreted as self-referencing in a programmatic sense, but instead which simply holds onto, and is "unaware" of, translation replies for each possible input.
You are right, Searle isn't making an argument about the translation of sentences. (translating them to what?)
He is making an argument about how the mechanism of computation cannot capture semantic content. He explains this in the google video very well: https://www.youtube.com/watch?v=rHKwIYsPXLg
And all of the... let's call them "structural" critiques are moot. Searle's point is that computer systems cannot understand semantic content because they are syntactic processing machines. And he shows this with his argument.
The opposite view is that computers can understand semantic content. (so there is understanding and there is meaning understood by the computer) and the reason Searle doesn't believe computers can do this is because his argument is flawed.
Which leaves us with a small set of options:
1) That the structure Searle proposes can in fact understand semantic content and Searle just doesn't understand that it does.
I don't think anyone believes this. My iphone is certainly more capable, a better machine, with better software than Searle's room, and no one believes my iphone understands semantic content. so the belief the Room does understand semantic content but not my iphone is plainly false.
2) Searle's Room is simply the wrong kind of structure, or the Room is not a computer, or not a computer of sufficient complexity and therefore it cannot understand semantic content
I think this is the point you are making, but correct me if I'm wrong. This is not an objection against Searle's point. It's a critque of the structure of the argument, but not the argument itself. Searle could rewrite his argument to satisfy this objection, but it wouldn't change his conclusion.
Which bring us to the generalized objection:
3) that sufficient complex computer would understand semantic content.
Aaronson's paper is about the complexity problem and how a sufficiently complex system would APPEAR to understand semantic content by passing a Turing test within some limited time.
There are many arguments to this line of reasoning. One of them is that all such limitations are irrelevant. You yourself are not engaged in a limited time turing test, no person is. The issue is not passing turing tests, it instantiating sentience.
But thinking about complexity gets us off the root of the objection. You intuit that increasing or decreasing complexity should give us some kind of gradient of sentience. So an insufficiently complex system would not be sentient and would not understand semantic content, but this isn't what Searle is arguing.
Searle is demonstrating that no syntactic processing mechanism can understand semantic content. Understanding semantic content is a necessary condition for sentience, therefore no computer which does syntactic processing can be sentient. A gradient of complexity related to sentience is irrelevant.
In the one case: our computers become so complex it becomes sentient -> because it is sentient it can understand semantic content. Vs. understand semantic content and that leads to sentience.
The gradient of complexity to sentience is an intuition. Understanding of semantic content can be atomic. Even if a computer only understands the meaning of one thing, that would disprove Searle's argument. A gradient of complexity isn't necessary. Searle is saying there is a threshold of understanding semantic content that a computer system must pass to even have a discussion about actual sentience. And if a computer is categorically incapable of understanding semantic content, it is therefore incapable of becoming sentient.
Said another way, sentience is a by-product of understanding semantic content. Sentience is not a by-product of passing turing tests. The complexity required to pass a turing test, even of finite or infinite length, says nothing about whether a machine does or does not understand semantic content.
All the structural critiques of Searle fail because they do not offer up a program or system that understands semantic content.
Show me the code that runs a system that understands semantic content. Even something simple, like true/false. or cat/not a cat. If Searle's structure of the room is insuffiently complex, then write a program that is sufficiently complex. And if you can't, then it stands to reason that Searle at least might be correct: computers, categorically, cannot understand semantic content BECAUSE they do syntactic processing.
Google's awesome image processing that can identify cats does not know what a cat is at all. It simply provides results to people who recognize what cats are, and recognize that the google machine is very accurate at getting the right pictures. but even when google gets it wrong, it does not know the picture does not have a cat in it. In fact, the google machine does not know if what it serves up is a cat picture even if there is a cat in the picture.
The Searle Google talk covers this very well: https://www.youtube.com/watch?v=rHKwIYsPXLg
If you fed googles cat NN a training corpus of penguin pictures and ranked the pictures of penguins as successes, it would serve up penguins as if they were cats. But no person would ever tell you a cat is a penguin. Because penguins and cats are different things, they have different semantic content.
I would love to see that Searle is wrong. I'm sure he would be just as pleased. So I am curious if you do have or know of a machine that does do, even the smallest amount, of semantic processing. Because solving that problem with symbolic computation would save me a ton of effort.
1. What exactly do you mean by "a kind of metabolic computing"?
2. What is first step you want to accomplish?
3. What do you think (feel) is actually happening in any sentient animal that leads to semantic content? How it is possible that this happens? We know that it happens because we are sentient animals. The question is: where is this difference because as animals we are also machines and it seems that everything what is happening in our cells is purly syntactical.
Molecules are driven not by exterior phenomena, but by changes intrinsic to the molecules and atoms and other particles they interact with. We live in a particle universe. We do not live in a universe with outside "forces" or "laws" that cause the particles to behave in any way. Everything about the physics is from the inside out, and the interactions are always "local" to particles. Large scale phenomena are actually huge quantities of particle phenomena that we perceive as a single large phenomena. (this is a kind of illusion).
When we try to write programs that simulate physical phenomena, like atoms, or molecules we write the code from the outside in. It is the program which causes the data changes to simulate some chemistry. But in nature, there is no program determining how the molecules act. chemical changes occur because of features about the individual molecules interacting, not because of a rule. Simulations like this do not replicate what happens between individual molecules, the replicate what would happen if molecules were controlled by an external rule (which they are not).
any rule based simulation can only express the set of possible outcome conditions from the rules and data. but it cannot capture it's axioms, and it cannot capture conditions that in fact exist outside it's axiomatic boundary. (Aaronson and p4wnc6 both remark on this limitation by pointing out the complexity necessary to achieve a good Turing test result or sentient AI).
My approach is to treat this intrinsic nature of molecular interactions as a fact and accept it as a requirement for a kind of computer system that can do "molecular interactions" from the inside out. And my supposition is (not proved yet!) that a mixture of such interactions could be found that is stable, that would be homeostatic. And if such a mixture could be found, then could a mixture be found that can be encapsulated in a membrane like structure. And could such a mixture store in it's set of code/data like "molecules" it's internal program - eg. DNA.
I think the answer is yes.
There are three different steps that all have to work together.
One is understanding how representation works (see my email to you, it's outside the bounds of this thread). So understanding how semantic content and awareness works, in all situations and conditions, is a precondition to recognizing when we have code that can generate semantic content.
The next is finding a model of how representation is instantiated in organisms to use as a basis for a machine model.
The third is then coding the machine model, to do what organisms do so that the machine understands semantic content, and the machine should produce awareness and consciousness.
I believe metabolic functioning is the key feature to allow us to do representational processing. hence why I call the approach I am taking, metabolic computing. The step I am currently on is writing up an interpreter that I think can do "molecular" interactions between code/data elements. Meaning that the data/code elements determine all the interactions between data and code intrinsically. the interpreter processes those "atomic" interactions based on intrinsic features of that code. Essentially, every bit of code/data is a single function automata and they can all change each other so the system may or may not work dependent on the constituent "molecules" of the system. I call this "the soup".
previous prototypes required me to do all the addressing, which itself was a big leap forward for me. But now the code/data bits do the addressing themselves. (each function interacts "locally" but interactions can create data structures, which is the corollary to molecules forming into larger structures and encapsulating molecular interactions into things like membranes).
So the next step is finish the interpreter, then see if I can get the right soup to make functions (like dna and membranes. I've written out RNA like replication examples and steady state management as discussed in systems biology so I there is a path forward). Then see if I can get to homeostasis and a "cell" from datastructures and interacting "molecules". the step after that is multiple "cells" and then sets of cells that form structures between inputs and outputs. eg. a set of "retina" cells that respond to visual inputs, a set of cells that "process" signals from those retina cells, and motor cells that take their cues from "process" cells etc.
the cell level stuff and above is mostly straightforward. it's forming different kinds of networks that interact with each other. Nodes themselves are semantic content. but how do you make networks from the inside out? from meaningless (syntactic) molecular interactions? that is where the metabolic systems (and stigmergy) come into play. (actually, stigmergy comes into play at many levels)
In biology, the syntactic to semantic jump happens at the cell. the cell itself is a semantic thing. the syntactic processes maintain the cell. the cells underlying mechanims and interactions are all syntactic. and the cell doesn't "cause" anything, everything happens in the cellular processes for their own intrinsic reasons, but the cell, it is semantic content. (embodiment).
the embodiment path is how to get representation, awareness, and consciousness.
My apologies that this is somewhat all over the map, but the problem of making machine sentience actually work requires that theory, model, and implementation all work. And if any of them don't work, then the outcome of sentience becomes impossible. And that's just a lot of different stuff to try to compress into a comment!