2013-06-30 ABC The Philosopher's Zone AI - Think Again
Duration: 00:25:14
Transcript
Fran Kelly
Hi, Fran Kelly here from RN Breakfast. Before you listen to this podcast, just a reminder that all our stories are online too. So if you want to download breakfast morning, noon or night, head to the RN homepage.
Joe Gelonesi
On RN, the Philosopher’s Zone, Joe Gelonesi with you. Today, the puzzle of artificial intelligence.
David Deutsch
Because of an unfortunate series of events in the history of philosophy, most philosophers and most scientists have completely the wrong picture of what the creation of new ideas is like as a physical process. And therein lies the problem.
Joe Gelonesi
David Deutsch, pioneer of quantum computing, calling for the philosophers to help clear the roadblock. We’ll hear from the renowned physicist shortly on why a new philosophy of knowledge is desperately needed. Of course, philosophers and scientists haven’t exactly seen eye to eye on the way forward. Hubert Dreyfus was there in the early days, the 1960s. Fiercely critical of the philosophical foundations of AI. His book, What Computers Can’t Do, is a classic. AI research itself has hit a bit of a speed bump. What do you think it would take now? What sort of philosophical insight do you think it would take to move it along?
Hubert Dreyfus
I was just talking with a student about that yesterday, so it’s sort of on my mind. Because good old fashioned AI, as my former student John Haugeland calls it, was symbolic AI, which had a view that the mind had in it representations of the world and concluded from those what to do and so forth. That just failed. I think everybody realizes that that’s good old fashioned AI. So then what is there instead? Well, there’s a whole bunch of interesting possibilities, but they haven’t gotten anywhere yet. The good old fashioned AI people accepted what was called the Turing test. Alan Turing, the famous guy who’s the main thinker of digital computers, said that the test to see whether you actually have captured intelligence if you’re trying to do artificial intelligence, is to see whether the computer could carry on a conversation with a human being in such a way that the human being couldn’t recognize that it was a computer. Another way to put that is computers don’t have common sense knowledge. That’s the way AI people put it when they realize there’s a big deal problem. Common sense shouldn’t be called knowledge, but we do have something like common sense familiarity with the everyday way that people do things. But that’s a big complicated total something or other. Philosophers call it the background sometime. Heidegger is the main guy. But the question is, is the kind of computational software, let’s say, that leads to this seeming start toward intelligence going to get you anywhere? So there’s really interesting things going on. There’s the Google car that goes into traffic and on real freeways and gets around with no driver, doing it with artificial intelligence. Not with good old fashioned AI, not with symbolic representations, but mostly with neural networks, which are a way of using computers to simulate in a very, very rough way what goes on at the level of neurons. And the question is whether the kind of intelligence that you get with Google car driving could ever get you to passing the Turing test, that is, showing human common sense behavior. And the answer, I think, well, it doesn’t look promising, though that’s not a criticism of the kind of intelligence that you can get with neural nets. Then there comes another one which is really kind of stunning because I can’t really understand it. And that is the highest kind of thing would be the Watson program that IBM has that actually played Jeopardy and actually beat the two best Jeopardy players in the history of Jeopardy. Of course, that was in what seemed like ordinary English, the questions in Jeopardy. And the computer seemed to understand the questions, and not only understand them, but relate them to everyday life and the real world and answered the questions. And whatever it is they do requires millions and millions of computations in fractions of a second in which all sorts of data that is possibly relevant is related to all sorts of other data which is possibly relevant until some item comes up which seems to score high on relevance. And then the computer answers with that. And lo and behold, that turns out to be the right Jeopardy answer a lot of the time.
Hubert Dreyfus
I mean, all of Wikipedia is in there. All of Google is in there. And you search all that and you get an answer within a few seconds. Is that going to enable you eventually to pass the Turing test just by virtue of the huge amount of meaningless calculations which add up to some relations being more probable than others among all these calculations. And I have no reason to think that that’s going to give us artificial intelligence, that that’s going to be able to search in such a way as to get common sense knowledge. One thing is sure though, and that’s really important, it’s whatever the computer is doing in the Jeopardy program, it’s not doing anything like what we do when we think and when we have common sense knowledge because the neurons in the brain work very slowly and though there are a lot of them, several million, it’s nothing like searching the immense amount of data that Watson program searches. So if AI is using computers to do something that human beings do, the way human beings do it, I think it’s hopeless that our relatively few, several million neurons at the slow speed that they operate could ever come up with anything like winning at Jeopardy. That’s out of the question, I think. The computer just does it faster and makes more correlations than we could ever make. So that goal of artificial intelligence, using computers that in some way capture what human beings do, I think that’s impossible. But what the future of AI is, I really don’t know.
Joe Gelonesi
It seems size doesn’t matter. Heap on the computing power, but the general intelligence lights still won’t come on. Hubert Dreyfus, the original philosopher gadfly. But the intellectual battle has moved on and now a renowned physicist has called for help in the trenches. David Deutsch is a pioneer of quantum computing. We talked on Skype. Firstly, if I could just ask you about some terminology and the term AGI, what does the G refer to in AGI and why is that important?
David Deutsch
The field used to be called AI, in other words, artificial intelligence. And it originally meant an artificial recreation of the distinctive ability called intelligence in human beings, the one that allows us to do all the distinctively human things that human minds can do. However, as computers were used for all sorts of other purposes, such as playing chess, that constellation of abilities came to be called AI as well. And so one needed a new term to refer to what AI originally referred to, namely the artificial implementation of distinctively human abilities. And that, to me, is primarily the creation of new explanations of the world or of abstractions or of right and wrong or whatever. And so that’s what’s called artificial general intelligence. The general meaning now, it’s got to be everything. You know, we’re not going to call playing chess AGI because it’s not general. The only thing that is going to count as AGI, as success in the field of AGI, is something that is capable of all the things that human minds are capable of.
Joe Gelonesi
So if that is the case, what is the state of AGI now? I know you’re very critical about the development of AGI, that in fact it’s gone nowhere for a very long time.
David Deutsch
Yes, that is my view. And I always have to say at this point, because I think most of the people who think that AGI has gone nowhere think that it can’t go anywhere. And I have exactly the opposite view. I think there are fundamental reasons why AGI must be possible. And it must be possible with existing kind of computers. But what is needed is a way of programming computers to make them into AGIs. And that’s the thing that hasn’t succeeded. That hope, that prediction has not been fulfilled. And what we have today is an enormous range of specific abilities of computers, such as playing chess, such as being search engines, such as machine translation, word processing, and of course the whole Internet, all of which are special purpose applications. But in terms of general purpose application, which can emulate a human’s ability even slightly across the board, not only is there no such program today, but no one knows how to write one, and no program even begins to approach that kind of ability. So we don’t know how at present.
Joe Gelonesi
And yet, David, you believe that it can be done, even though everywhere you look we haven’t been able to capture what that general intelligence is in an artificial sense. Why do you believe it can be done?
David Deutsch
The reason I’m sure that it can be done stems from my core research. The reason that I’m interested in the foundations of computer science at all is that I have been interested in the role of computation in fundamental physics, and it has turned out that, first of all, computation is a fundamental physical process, which was a big surprise because computation looks as though it’s just a piece of technology like bridge building, and not anything fundamental like quantum mechanics or thermodynamics. But in fact, it turns out that the theory of computation is a fundamental branch of physics. And in its capacity as a fundamental branch of physics, we know quite a lot about it. Perhaps the most important thing we know about it is that absolutely any physical process can be simulated with arbitrary accuracy by a computer program. And therefore, it follows from fundamental physics that a computer can have all the functionality of a conscious mind, of a creative mind, in other words, all the general abilities of a human mind. This, of course, has philosophers at ten paces and scientists and computer scientists as well.
Joe Gelonesi
If we could just take this in step, what you’re saying is it can be done physically, and there is no reason given the theories of universal computation and universality that it can’t be done. But the difficulty is, and you say this quite explicitly, that there needs to be a breakthrough in philosophy before AGI can go any further. What do you mean by that, given that in some ways you’ve sidelined the philosophical content in this discussion to some degree?
David Deutsch
Yes. So the philosophical content that I have sidelined is the debate about whether something which had the right functionality would be, quote, truly an AGI. So we can sideline that and just focus on the issue of whether the functionality is obtainable at all. And then there are two issues in philosophy that remain after that, and they are really the things that are holding things up. One of them is that because of an unfortunate series of events in the history of philosophy, most philosophers and most scientists have completely the wrong picture of what the creation of new ideas is like as a physical process, deemed to be a process of transforming inputs into outputs and so on. And that’s completely barking up the wrong tree. So that’s one problem, that they have the wrong philosophy of knowledge, the right philosophy of knowledge being the one proposed by the philosopher Karl Popper in the mid-20th century, which hasn’t really caught on as such. Even people who say they agree with him, in fact, don’t get the point. So that’s one half. We have the wrong philosophy. The other thing is that even if they had the right philosophy, there’s still a breakthrough to be made. We still don’t know what to do. If you can’t program it, you haven’t understood it.
Joe Gelonesi
What is it that we don’t understand about Popper that makes it difficult to get to this next step? And I assume you’re talking about Popper’s understanding of induction and how that works in the reasoning process and the scientific process.
David Deutsch
That’s right. So traditionally, for thousands of years, it was believed that the way we get our knowledge of the physical world is by observing it and then somehow transforming those observations into general ideas, explanations, theories, understanding. And the role of observation is that we generalize it. So we see the sun rising every morning and we form a theory that the sun will rise every morning. And the question is how to do that transformation in the case of AGI, how to make a program such that if you pointed the computer’s camera at the eastern sky every morning and it saw the sun, it would eventually say, hey, maybe the sun rises every morning. And that is impossible. It’s simply not true that theories can even consist of saying that a certain observation will be repeated every day. The real theories in science do not take that form. They are of the form actually the sun is a million kilometer sphere of glowing gas powered by nuclear fusion. And it doesn’t rise at all. It stays where it is. And we rotate. Those are explanations. And you’ll see that those explanations do not take the same form as our observations do. They don’t even mention rising and setting. And if you take even more general theories than those about the sun, they don’t even take the form of any kind of observation. Very often they take the form of saying why what we see really isn’t true. So not even what we see is true. For example, we look at the stars. Our theory says those too are white hot gigantic objects. Our theory of the stars begins by saying they’re nothing like what they look like. They look like they’re cold. They look like they’re tiny. They look like they’re twinkling. But actually they’re doing none of those things. And a scientific theory typically says that our observations are misleading in a number of ways. And the reason they are as they are is because of a sequence of misleading things that happens to the evidence on its way to us. So heading in the direction of generalizing observations or obtaining theories from observations, which is the way we do all other computing, all other computing is about getting an output from an input. But AGI programming cannot be anything like that.
Joe Gelonesi
What you’re saying here about the philosophy, the breakthrough in philosophy that needs to be made is that we need to somehow get to grips with creative leaps and how to actually use the understanding of error within making those creative leaps.
David Deutsch
That’s exactly right. It hits the nail on the head when you mention error. Because one thing we know about the way that we do this is that we make a lot of errors on the way. And even when we do get a sliver of truth, it’s always shot through with a lot of error as well. For example, Newton’s laws were an enormous advance in not just in predicting, but in understanding what the solar system is. And yet we now know that some of its fundamental claims, such as that there’s a force of gravity pulling the Earth towards the sun, were actually misconceptions. And the real reason that the Earth goes around the sun is that space-time is curved, a concept that couldn’t even have been formulated at the time of Newton.
Joe Gelonesi
David Deutsch, quantum computing pioneer, on what it will take to make computers think. And you’re listening to him on the Philosopher’s Zone, here on RN. The thing that’s divided people to a large degree about whether there can be such a thing as artificial intelligence is that the question of consciousness and what it is has to be properly understood before we can take a further step. Is this part of the philosophical breakthrough that might need to be done?
David Deutsch
This is slightly a separate issue, whether we need to understand consciousness. The reason is this. I happen to think, but this is a separate matter from my views about whether AGI is imminent or that kind of thing. I happen to think that all the properties of minds that are currently deemed to be mysterious, and that includes consciousness and it includes creativity and it includes free will, for example, that they are all actually the same. If so, then we need only solve the easiest of those, which I’m guessing is creativity, because creativity, the creation of knowledge, is the one that we currently know the most about. We know Popper’s theory of knowledge, which is a huge starting point on the way of understanding how creativity can possibly work. But if I’m wrong that those three things are all really the same thing, then that doesn’t affect the point I’m making about AGI, because if we did solve the problem of creativity, but not the problem of consciousness, we would still have a computer program that had the functionality of a conscious being. In other words, it could do what conscious beings do. It could create new ideas. It could solve problems in science. It could pass Alan Turing’s test for being an AGI, namely that it could converse on a human level with a human and the human couldn’t tell that it wasn’t a human. So it could do all those external things and possibly without having consciousness or free will, it would just look as though it had them. Philosophers sometimes call this position the zombie theory, because such an AGI would have the external appearance of a human. Well, actually, they obviously haven’t seen zombie movies, but this would be the kind of zombie that looks exactly like a human, but hasn’t got any personhood in the sense of having human feelings, human sensations that would just look as though it had them.
Joe Gelonesi
Yeah. I was thinking about the zombie thought experiment when you were putting that forward, this idea of the machine that can do it all, but doesn’t really know it’s actually doing it all or experiencing it from the inside.
David Deutsch
Yes, I think it’s implausible that there could be such a machine, but if there could, and if the people who think that creativity is separate from consciousness and all that stuff are right, it doesn’t at all affect my point about AGI, because my point about AGI is focused on the functionality of AGI.
Joe Gelonesi
And this functionality, you say, can be captured if an algorithm could be tapped into, which tapped into creativity and how creativity deals with context and justifying knowledge within the arena of error.
David Deutsch
Yes, creating knowledge within the arena of error. Yes, that’s right. Exactly. If I take an analogy, for example, in the early 19th century when we didn’t know what life was, you know, unless somebody had made the philosophical breakthrough that Darwin made to realize that the shapes of animals are not what the process that made them is about. The process that made them is about these traits, as he called them, or what we would today call genes, which create the shapes of animals as a side effect. So if somebody hadn’t understood that and they had thought that the problem of the origin of species and so on, was why is an elephant the shape it is? Why does it have a trunk and so on? Then they would concentrate their efforts on trying to sculpt elephants on a finer and finer scale in the hope that if they did it on a fine enough scale, the elephant would suddenly start walking away and being exactly like a normal elephant. And they’d be completely wrong because the point of the correct theory of why elephants are the shape that they are is not about the shape. It is about natural selection and replication and variation and differential survival and accurate copying of genes, not of the animal’s body and so on. All those things are hidden when you look at an elephant. But it’s only if you understood those things that you would have the slightest chance of knowing why the elephant has a trunk, let alone making an artificial elephant. My intuition is that it will be very easy to program. It’s the understanding it that’s hard.
Joe Gelonesi
Ah, intuition. Let’s not go there. Well, not today anyway. Pioneer of quantum computation, David Deutsch, calling for a new philosophical understanding to break the impasse on AGI, artificial general intelligence, and he’ll take it with or without consciousness. Next week, a fitting finale for season one, renowned philosopher Simon Blackburn on human nature and science. His subject for the inaugural Alan Saunders Memorial Lecture. The Philosopher’s Zone is produced by Diane Dean. Technical production today by Martin Peralta. I’m Joe Gelonesi. See you next week.
Health Report promo
Will doctors of the future be anti-fat? Looking then at the unconscious results, we found that about 40% of students had a significant anti-obesity bias. Implicit bias in medical students and ways to overcome it. Also, kidney disease, fetal immune programming and a silver bullet for antibiotic resistance. It’s a jam-packed health report with me, Joel Werner. You can download the program from the RN website.
Markdown