Does Your Computer Know Chinese?

think of as a human intelligence. We're smarter than like pigs maybe, or like grasshoppers and stuff like that. I flatter myself to be considered smarter than my toaster most days. Yeah, so this actually got us into a discussion about being smarter than inanimate objects, and so far we're doing quite well. Yeah. I want y'all to imagine a scenario. Okay, imagine that you have a real good friend Ardia already stretching imagination already. Well, I know you don't have good friends, but we just

pretend you have a really good relationship. But this guy named Artie, and he's funny, intelligent, compassionate. Um, not to mention the strongest man in the world. Uh. But you have been extremely good friends with Artie for many, many years. Um. And suddenly one day, while you and Artie are hanging out, maybe having some coffee in your living room, a bunch of people in lab coats bust into your living room

and they deactivate Alreadie with a remote control. Uh. And the head scientist among this group, she flares her lab coat and informs you that Already is in fact and artificially intelligent robot. And to prove it to you, she opens up Already's head and shows you that it is full of dense circuitry where a normal person's brain would be. Pop.

Quiz is what's your reaction in this scenario? Are you a horrified that you were tricked into a fake relationship with fake intelligence or are you be amazed that a synthetic intelligence is capable of being a real friend. I'm gonna go with Cee. I'm amazed they were able to burst into my living room because there's no doors, so live us through the walls. So I got like a wrecking ball comes in and it's got a bunch of lab coat people clinging to it, and so like scientist

versions of Miley Cyrus. Maybe I think that I am d upset by this entire situation and that someone has just deactivated my friend. I think that I'm concerned about artist personally, which probably puts me into a B camp

if you're if you're really being technical, Okay. But so yeah, I it's hard to know how I would feel, But I would like to say I think i'd go with B. That I wouldn't just be horrified and say like all of those years were just a yeah, Now, I've had enough of those in in my life from from people who,

as far as I know, are not robots. So I think I would mostly be amazed that someone had achieved the technical ability to simulate intelligence to the extent where I could not tell the difference between what is simulated intelligence and real intelligence. Okay. And so this ties into a concept we've talked about on the show before, which

is the Turing test. What what is the Turing tests? Essentially, the basic version of the test is that you have a person positing questions over a terminal, so a computer terminal. They cannot see whomever or whatever is answering those questions. Those questions are being answered either by a person in another room at another terminal, who's typing those answers in or by a computer program that is responding automatically to

those questions. If the person asking the questions is unable to determine with any real level of of of certain certainty, thank you words that don't come to mind certainty, then that machine is said to pass the Touring test if in fact it was the machine that was answering those questions. Okay, And so that's the basic Turing tests, say the back

and forth text based chatbot. But you could sort of look at the Turing test in a larger view and say, it's the bigger question of can you simulate human intelligence in a way that people can't tell the difference? And in fact, Touring went so far as to say that if this machine was able to simulate human intelligence to that level, we would, in fact, we should, in fact say that it is intelligent in the sense that we

should grant it the same kind of consideration that we do. Right, because because we cannot inhabit another person's being, we cannot be certain that that person actually possesses intelligence. We assume, yeah, we assume that they have the same things that we You know that they have the same things that we have, so we should extend the same courtesy to machines if they are capable of doing such a thing. Right, So

this is what's known as sort of strong AI. Actually this is really known as weak AI because it's simulating intelligence, not necessarily possessing intelligence. Well, that what the question is about strong AI. The question is is um something that displays convincingly intelligent behavior actually intelligent. And here's where we come back to Artie. Okay, So Artie, your friend convinced you for years that he was a real intelligence. He seemed intelligent, he was, you know, funny, He was a

hundred indistinguishable from every other human you knew. And when he was activated, he would insist that he was in fact conscious, understanding language, would insist on all the other things that we would insist on as part of our sort of mental abilities. Was already actually intelligent and did already understand the relationship the language, all of these things that happened between humans. Or was he merely running programs

that made it seem as though he was right? So, in other words, this is the difference between we weak

AI and strong AI. Week A I would be running programs where you would be taking in various kinds of input and producing output that to an independent observer would seem to be this kind of natural processing of information and reaction in the way a human being would react week at week A. I would be the position that even though you could simulate it, it wasn't really intelligence, right, and strong AI would be that it actually has a

level of intelligence that's comparable to human intelligence. It might not be identical, but it would be. It would truly understand, It would know what it knew, and not just be able to process information based on some complex series of rules. Right. So can a robot or a computer program, any kind of machine that's based on symbolic programming and computation actually no or understand anything. And this is a question that

is debated quite heavily in artificial intelligence circles and cognitive science. Yeah, what we want to focus on today is one particular thought experiment that has tried to answer this question. And I would say that this is probably the most hotly debated question in AI and cognitive science in the past few decades, well at least the past decades certainly. Yeah, we're talking about John Searle's Chinese Room thought experiment, which

is a really uh interesting approach. And there are some also interesting counter arguments to the Chinese room thought experiment that we'll have to kind of talk about. Yeah. As of about, computer scientist Pat Hayes had defined cognitive cognitive science as a field as the ongoing research project of refuting searles argument. Um. Yeah, So obviously this is a huge issue, hotly debated by people a lot smarter than we are and who have thought about it a lot more.

So we're not going to settle it here today. We're just going to lay it out. We were just going to try to explain it and offer a few of the prominent thoughts in this field. Let me give a crack at explaining what this thought experiment is. Okay, okay, Well, first of all, just let's know where it came from. It was from John Searle in his nineteen eighty paper Minds, Brains and Programs, and that was published in the Behavioral

in Brain Science. He has he has updated it a couple of times since then, uh, and also responded to various criticisms leveled against his argument. But here is the basic thought experiment now, imagine that you, Joe. I'm going to use you as an example. The Joe, you don't happen to know Chinese, do you know? Okay, well, then you're a perfect I wanted to use you as an example. It's too late. I'm the one giving the example, so

you are the example. I also don't know Chinese, so either one of us would work in Okay, you can use me. Okay, So Joe does not know Chinese. Uh, he is put into a room. That room has a desk, it's got some paper, some pencils or racers, it's got some file in cabinets and as a door with a slot in it. And the way that Joe interacts with the outside world is that occasionally someone puts a piece of paper through that slot that has a Chinese character

drawn on it. Now, Joe also has an enormous book full of instructions on what to do every time a piece of paper with a character comes in and he looks at the character, goes through the book, finds the character in the book, follows the instructions on what he's supposed to do when that character is inserted, which usually involves him drawing a different Chinese character on a on a sheet of paper and then slaying that back through

the slot. Now from anyone who is outside of that rooms, a Chinese speaking person who is writing down a specific instruction on a sheet of paper and slides it through that slot and then gets the result that Joe has driven written on the other sheet of paper. To them, it appears that whatever is inside that room understands Chinese right from their perspective, if if the instructions are good enough, yes, right. So what we're sort of a mad jenning is this is a room with a person in it that passes

the Turing test. Yes, yeah, because you the person on the outside has no idea what's on the inside of the room. All they know is that they have passed in instructions that were in Chinese. They've received output that's also in Chinese, and everything follows according to a sets set of rules, predetermined rules, and and the output is precise and convincing enough that the person outside thinks that

whatever is inside is a true native Chinese speaker. So so the question then in this experiment is does this mean that Joe actually knows Chinese? And Searles, uh, Well, Searles says no, you don't. You don't know Chinese. You don't understand Chinese. You're following instructions, but you don't understand the actual language. Actually, I might make one changes. I think it's the obvious point is that I and that

thought experiment don't actually know Chinese. The question is does this translate to a computer in the same situation, because this with this supposed to be an analogy for the situation that a computer or turing machine is in sure, it receives input, um it has a set of instructions that it follows in in computer readable language. It processes that input through those instructions, and then it gives output. But it doesn't have any actual understanding of what is

going on at any given time. It it knows to follow specific instructions because of a program, but it doesn't understand what's happened. It doesn't understand the question more or less much less the answer, right. And so that was Searle's point, and it wasn't He wasn't against week Ai. I mean, he obviously believed that maybe a computer could simulate human intelligence so well that we couldn't tell the difference.

He was just pointing out that UM it seems that based on simple programmable instructions and symbolic computation, that a computer could never really understand but it was dealing with. Yeah, so understanding is the crux of this argument. And also it's problematic because often people who level criticisms against it say that we need to define what understand means, and then it gets a little metaphysical. But we're not going to go too far into that because obviously that that

can take hours of discussion. Well, but I mean, part of this problem is the is the semantics of the terminology that we use for this kind of thing, And and which is a little bit funny because because syntax versus semantics is considered um kind of the key point of this argument, right, Since what the computer clearly does understand is syntax, it's got instructions and data to deal with. What's up for debate is whether it gets semantics, whether

it understands the meaning of the symbol. So, for example, we've talked about the semantic web, this idea of a web, some way of interacting with the Worldwide Web. I love to use that old term because I was around when it started, but a way to interact with the web where it quote unquote understand ends what you want and

it's pulling exactly the stuff that you want. But when we talk about actually making the semantic web come to pass, it's all about things like meta data and including all of this information about information so that a computer can make sense of it, because a computer does not natively possess this ability to understand semantics. And even in that case, we would call it the semantic web, but it would still be based on program yes, and still be very

much based on syntax. All right, All the examples that we can really come up with are very much based in our human world because it's a little bit difficult for us to understand, um without metaphor exactly what's going on inside of a computer. But it's it's sort of talking about the difference between like memorizing versus understanding and equation or being tested on memorization versus comprehension. Yeah, okay, so well, I want to offer a few of the

responses I've read about. None of these, of course, are original to us. This is something that's been debated over and over again. So these are all things that have been gust in that philosophy world, and an AI and computer science before UM. But I think one of the most important things to start with is sort of the epistemological question is just going straight back to the Turing skepticism, which is that, um, if you can't tell the difference,

you can't tell the difference. That the idea is, if someone appears to you to have understanding, how could you say that they don't have understanding? Um? And I think that's a really good point to start with. But there is also a good response to it, in my mind, which is that, um, it still leaves the question of

what is the basic thing we're talking about? So we might not be able to know if other people have understanding, but just saying that sentence still acknowledges, well, there is such a thing as understanding, and whether or not we can know that, the question of whether a program machine can in principle attain this state remains. So it's not

deconstructing the argument, and it's just it's just raising more questions. Yeah, and so one point raised by this objection might be that, well, what if we can never solve this question, But that's not the same as saying that the question is meaningless. Basically, sure, another objection or another criticism leveled uses what Searl has

identified as the system approach. So the system argument is about how if you look at the person in the room as supposed to you know, the person in the room is supposed to represent a computer, then clearly you would argue while Joe didn't really learn Chinese, computer being Joe, Joe clearly did not learn it. He was following specific instructions that were in a giant book that were not part of Joe's knowledge. All he was doing was taking an input and then applying this set of rules and

then getting output. But the system approach says, well, you're looking at this wrong. You're looking at at the Joe at being the representation of the computer, whereas really it would be the entire system that you have to take into consideration rather than just Joe. Joe would be like the processor, the CPU, and that that something to do with the room, which has created these instructions, understands Chinese,

and therefore that that intelligence has happened. So when you look at the whole thing together, the room, the English instruction manual, me the input and output paper slips that are coming through the door, the filing system, which represents memory. Right. All of that together, you could say, in a way understands Chinese. Right, And it may not be directly analogous to the way a person understands Chinese, but that really

doesn't matter in this argument so much. We're talking about strong AI versus weak AI, not whether or not the intelligence is exactly like human intelligence. Yeah, and so this is probably the I think the most popular response to the Chinese room argument. There's some others we're gonna talk about in a minute, but I think the dominant one is this systems thinking approach. And and this also goes with, um a lot of the ways people have come up with to think about how intelligence comes out of a

material mind, sorry, a material brain. Um. So you know, it's like you're the brain is not just one discrete agent doing one kind of thing, but it's this confabulation of tons of different types of events that, in an emergent way combined to produce what we think of his consciousness. Right. It's also related to another response called the virtual mind reply, which which specifies that AI is a machine that has created a mind that understands Chinese, not that the AI

itself understands Chinese. Which is getting back into semantics. Yeah, it's getting it's also getting a little metaphysical, to the point where Searle's responses puffed. I think that's how you would sum it up. But essentially he says like, you don't need to bring metaphysics into it. This isn't this, that's that's irrelevant. Well, if you want to get really hard nosed and practical, there are a couple other ways

we could go with this. One of them is simply the criticism that the thought experiment is too impractical to be useful, or it's too unlike the real world. So think about it like this. Um It asks us to imagine that the person in the room would be able to use this English instruction manual, which is the equivalent to a computer program, fast enough to generate convincing responses.

But in reality, it might take a single person in a room like billions of years to go through all of the instructions in that instruction manual and turn back an answer that was satisfactory. A computer could do this in less than a second, probably, so Thus there could be a fundamental difference between the kind of slow intelligence that the thought experiment would display and an actually convincing touring ai um, and thus the thought experiment can actually

comment on what happens with a touring AI. See. I I have a problem with this just from a gut level. Obviously, I'm not a philosopher nor my scientist. However, I think about people who really take their time and answering a question, and I think of computers I've used that also really take your time and executing a program, And I'm just thinking that when it comes down to the time issue, I almost see that as irrelevant to me, like that

doesn't really matter so much to me. But then a part of it is also that I'm used to talking about stuff that's taking over taking taking place over the course of billions of years. So whether it takes you know, a month to respond to a request or it takes a second, I don't really necessarily see that as an indicator or or you know, something that denies the presence

of intelligence. Yeah, you could certainly say that you know that basically no matter how long it takes, the principles behind the question in the thought experiment remain the same. Um that it's just that's just sort of an irrelevant variable that we can exclude for the sake of argument, And the question is is it reasonable to exclude it for the sake of argument or not? Right, because if time is actually a factor, if speed is a factor in determining what is is an intelligent then clearly you

can't just dismiss it. Right, Like if some if for some reason, we discover that that the way that we exhibit intelligence is dependent partly on the speed at which our neural networks work, and then it is. But if they did not work at that speed, we would not be intelligent, That's what I'm saying. Like, if that were the case, then we would you know, well we wouldn't be having this argument because we wouldn't be able to. But oh, the way you just phrased it made this

objection suddenly seem more reasonable to me. Actually, Like, what if something that did all the same things a human brain does, except did them at one one billiont the speed. Would that have consciousness in the same way that we do? Well, I don't know that we would all ever argue them

as sument have consciousness the same way. Well, not exactly, but you know what I mean, has this intelligence and part of this is also part of a separate but related argument that talks about exactly how ego centric we're being in our definition of intelligence and understanding, because that's a huge part of it. I mean, you know, like like if if we do we define intelligence based on neural connections, because there are things out there who display

relatively intelligent behavior that don't have that on our planet Earth. Yeah, I just think of them as incredibly lucky. They're just on a They're just on a series of making their saving throw against stupidity. That's all that's happening. Uh. Yeah. Let me offer one more objection that's sort of more on the practical edge I think, and that's sort of the lack of diverse sensory grounding. So system, if a system is only receiving input by pieces of paper being

shoved through a slot, that's extremely limiting. Yeah. Um, and so okay. This is also known as the robot response to the argument. Yeah so uh. This argument basically says that what prevents the person in the room from understanding Chinese is the lack of sensory grounding in the outside world through diversity of cues like visual, auditory all that. So, perhaps a computer designed only to manipulate symbols based on

instructions wouldn't understand language or wouldn't understand those symbols. But if granted diverse sensory and put like video cameras haptic input, Basically, if you made that computer into a robot that could sample the world and all of the ways that we can, the system actually would be able to understand what those

symbols mean. It's an interesting argument. Uh. Cyril of course has his own counter which says that you could have a string of numbers that to a person in that room seemed to be meaningless, but are actually the readoubts that a result of a video camera that's set up pointing outside of the room, and it wouldn't magically make things more meaningful to the person inside the room, and

that it would just create more work for him or her. Right, you're just yeah, that's exactly what Sarl says, Like, all you've done is just made more work. You haven't made it. You haven't increased understanding or intelligence, You just increase the workload. So that's an interesting counter argument. Although related to that is the idea that that man in the room, and this is kind of where it becomes a poor example, would eventually learn Chinese due to his interactions with the

Chinese language. Oh so you're saying like that somehow, I mean, it seems unlikely. But the person, even through just interacting based on these instructions, would somehow gain an understanding that given given long enough and giving given enough input, or you know, maybe through enough questions about, for example, hamburgers, the guy would know what a hamburger is and would eventually connect the two things in his brain. See, I

have a real problem with that. I could I could see where Joe in his room starts to receive, over a great span of time, enough of one particular character that he's remembered how the resulting character already needs to be formed without having to look it up. He's just he's done it enough times. But that doesn't mean he knows what the input is or what the output means. He just knows, Oh, I've done this five hundred times. I know how to draw this without even having to

look it up. It's a good question, though, Actually, I mean this sort of just intuitively assumes that while the part we you can't question is that you could never learn anything just from the symbols alone. But I don't know.

I mean what if you, I mean, imagine something like this, So you receive sequences of symbols, and then based on over like hundreds of years of doing this in the room, you start to notice that certain sequences of symbols mirror mirror the syntax of sentences you would recognize, like you start to maybe understand part of speech or something well. And then maybe from understanding part of speech, you could

start to understand more complex relationships between meanings. All right, I think that part of the thought experiment assumes that that that Joe sitting in the room is not thinking at all about what he's doing, that he's merely following

the instructions. Yeah, it just seems to be sort of like a condition of this thought experiment that you don't learn, right, but you know, but but people do learn, and and you know, people like Kurtswile talk a lot about about machines also being a naturally emergent um forces just like

anything else in nature. I'm not going to dwell on cards while too much, but but I will say I will say that there have been some incredible developments in machines learning or being able to derive information based upon uh input, like sensory input. So we were talking just now about the robot objection, right, the idea that by adding the sensory input, the system might be able to piece together more things and thus learn and have true intelligence.

So you guys know about the computer that the Cornell people built that was able to derive Newton's laws of physics by examining the movements of a pendulum. Right, I had not heard about this. So they set a pendulum in motion, they have a computer that was set up to measure what was going on with the pendulum, and the pendulum within or the computer, within about a day's time, derived the basic law Newton's basic laws of physics because it observed how the the pendulum moved and then extrapolated

from that what laws must guide the movements. So what's it done for us lately? Well, I just think that that's an interesting example of how a machine has been able to seemingly learn something by observing. Now, again, I think Searl would argue that the computer doesn't really know that, but that the computer was able to make um conclusions based upon input, right, And that all kind of kind of ties back to what I was saying at the beginning.

About are kind of basic, the basic metaphors that we use to describe computers and how we're we're essentially misrepresenting computer syntax with our ideas of of ones and zeros, because really it's all voltage and operations. It's not really reading. A computer isn't reading information. It's it's chains of state changes. Um. And it's it's hard for us to even conceive of how how little computers understand because it seems like such

a compelling you know. Yeah, well that reminds me of one more interesting objection to the Chinese room argument, which is that, Okay, so Cearle isn't saying that no machine could ever have intelligence. I mean, what cearl is talking about is programmable machines, right, that are based on symbolic computation, says, you know, essentially our brains are machines. He's not intelligence. Yeah,

he's not saying that this is magic. He's just saying that, like computers in the way we think of computers can't do this. Um. So there's another objection there. It's like, well, maybe it's true that computers can't do this in the way we think of, like programmed computers based on information processing, you know ones and zeros, but maybe some other type of synthetic machine could. Yeah. And of course, when you put it that way, I mean the possibilities are impossible

for us to even contain. We don't know what that would look like. The only types of computers we can think of to be based on these symbols. So I was thinking, like, when I was thinking about the touring test in the Chinese room, there was one computer in particular that kept popping up to mind, something that appears to be intelligent based upon the way it behaved, And

I was thinking of IBM S Watson. Obviously, you know, Watson being the computer that competed on Jeopardy and now is trying to uh help trying doctors are trying to use it to help with medical diagnoses and things of that nature. But if you watched any of those episodes of Jeopardy where wats and was one of the contestants and was going up against two former champions, it appeared that Watson was fairly intelligent thing. I mean, it was

able to beat the champions at their own game. Spoiler alert for anyone who wasn't watched Jeopardy in the last few years, but you know, it was able to parse sometimes pretty complicated clues, go through all of the information that was stored on the computer because it didn't have any connection to the internet during those games, and try and determine what the best answer was, or best question in the case of Jeopardy was to that particular clue.

And that's a you know, that's pretty tricky stuff. It's one of those things where you know, you would say, even smart people have trouble sometimes on Jeopardy. We've seen it. So but again Searle would say, yeah, the computer didn't necessarily understand what it was doing. It was able to break this apart and look for clues using various search algorithms and probabilistic models to determine what was probably the right answer, but didn't actually understand any of that content.

That would be Searles or right. Even though even though Jeopardy is not you know, direct direct question and answer kind of things, it's a lot of it's based on cultural understanding or puns. But that's the kind of thing that Watson is really good at parsing because it just has enough stuff in there. It's got a lot of natural language processing ability that I mean that is part

of artificial intelligence. And again, you know, Searle was never saying that artificial intelligence is as a fool's errand, but rather that you know, defining it as strong AI versus weak AI, and and his skepticism that strong AI is something that we could attain with the current version of computing as we know it. Right, Okay, So now that we've talked about it, Yeah, where do you come down? Is already intelligent? Did already understand the friendship you had

with him? No one understands the friendship they had with me. Um. I mean I think I think no, based on our

based on our basic definitions of intelligence. But I again would argue more on the side of of that our definitions are flawed and that as we move forward into this incredible future, that we maybe need to think really hard about what we're calling intelligent and how we're I mean, I'm just saying that I don't want, you know, AI artificial intelligence like the Steven Spielberg Stanley Kubrick flick too

to come into reality. What a film? You know, So I intuitively feel like already would not actually be intelligent. But I also, especially in areas like this where we have so little understanding, I feel disinclined to trust my intuitions. I think my intuitions very likely could be leading me astray. Well, you certainly, I mean, we have a bias, honestly, and that's that that's undeniable. So I would say, particularly the system's response to CYL makes a lot of sense to me.

I don't feel it. I don't feel like it's true intuitively, but thinking about it abstract ly, I can absolutely see how that is a reasonable answer. Well, I mean, the example I could give to you on a biological level is that, you know, I can speak a little French, but if you take any single neuron, or even in any single synapse out of my brain, it can't speak French.

It's you know, you can't. You know. So, judging one element of a system and saying that that one element proves that the entire system can't do something seems to be shortsighted. Cyril, of course, has his own responses to that. That's not that's not the end of the argument, and I just broken. I don't consider him refuted, But I

think that objection is interesting. It is interesting, I agree, entirely interesting, but not you know, not the end all of this, like we said, there are people way smarter than us who are continuously debating this and and expanding our UH knowledge. And sometimes they're focusing definitions, sometimes they are completely scrapping definitions and starting from scratch. So it's the best thing about all of this is that it's going to inform future engineers who are working on artificial

intelligence problems. And it may very well be that either we will never create any sort of entity that has strong AI, or that we do but we have no way of knowing that we did it. But no matter what, it means that we're going to make improvements in that field. We're going to have computers that are better able to react to things like natural language commands, and that will

in turn benefit us. So while this whole discussion might have seemed philosophical and that there was no real practical application, the reality is exactly the opposite. Well, and if there is such a thing as strong AI, you better believe this has practical application, and we certainly should be thinking about it. I mean, just on the legal and moral basis, right, just on the off chance that there is such a thing as the understanding and experience of being already I

think it's worth conser. Still not giving a toaster the right to vote, guys, still not doing it. I'm not gonna do it, alright. So that wraps up this discussion of artificial intelligence of the Chinese Room Thought experiment. And uh, while I sound cold, I'm I don't really subscribe to the things I say. Sometimes. Oh I always vote for toasters day. They believe in caramelization, and that's wonderful. That's okay,

that's fair. All right. So guys, if you have enjoyed this, make sure you go to f W Thinking dot com. That's our website where all of our stuff lives. Blog posts, videos, podcasts, articles. You should go check that out. It's great stuff. And if you want to interact with us, get on social networks. Guys. Come on, we're there. You can find us on Twitter, Facebook, and Google Plus. We are f W Thinking and we

will pomp to you again really soon. For more on this topic and the future of technology, visit forward thinking dot com, brought to you by Toyota. Let's go places