In this episode, Byron and Marie talk about the Turing test, Watson, autonomous vehicles, and language processing.
Byron Reese: This is Voices in AI, brought to you by Gigaom. I’m Byron Reese. Today I’m excited that our guest is Marie des Jardins. She is an Associate Dean for Engineering and Information Technology as well as a professor of Computer Science at the University of Maryland, Baltimore County. She got her undergrad degree from Harvard, and a Ph.D. in computer science from Berkeley, and she’s been involved in the National Conference of the Association for the Advancement of Artificial Intelligence for over 12 years. Welcome to the show, Marie.
Marie des Jardins: Hi, it’s nice to be here.
I often open the show with “What is artificial intelligence?” because, interestingly, there’s no consensus definition of it, and I get a different kind of view of it from everybody. So I’ll start with that. What is artificial intelligence?
Sure. I’ve always thought about artificial intelligence as just a very broad term referring to trying to get computers to do things that we would consider intelligent if people did them. What’s interesting about that definition is it’s a moving target, because we change our opinions over time about what’s intelligent. As computers get better at doing things, they no longer seem that intelligent to us.
We use the word “intelligent,” too, and I’m not going to dwell on definitions, but what do you think intelligence is at its core?
So, it’s definitely hard to pin down, but I think of it as activities that human beings carry out, that we don’t know of lower order animals doing, other than some of the higher primates who can do things that seem intelligent to us. So intelligence involves intentionality, which means setting goals and making active plans to carry them out, and it involves learning over time and being able to react to situations differently based on experiences and knowledge that we’ve gained over time. The third part, I would argue, is that intelligence includes communication, so the ability to communicate with other beings, other intelligent agents, about your activities and goals.
Well, that’s really useful and specific. Let’s look at some of those things in detail a little bit. You mentioned intentionality. Do you think that intentionality is driven by consciousness? I mean, can you have intentionality without consciousness? Is consciousness therefore a requisite for intelligence?
I think that’s a really interesting question. I would decline to answer it mainly because I don’t think we ever can really know what consciousness is. We all have a sense of being conscious inside our own brains—at least I believe that. But of course, I’m only able to say anything meaningful about my own sense of consciousness. We just don’t have any way to measure consciousness or even really define what it is. So, there does seem to be this idea of self-awareness that we see in various kinds of animals—including humans—and that seems to be a precursor to what we call consciousness. But I think it’s awfully hard to define that term, and so I would be hesitant to put that as a prerequisite on intentionality.
Well, I think people agree what it is in a sense. Consciousness is the experience of things. It’s having a subjective experience of something. Isn’t the debate more like where does that come from? How does that arise? Why do we have it? But in terms of the simple definition, we do know that, don’t we?
Well, I don’t know. I mean, where does it come from, how does it arise, and do different people even have the same experience of consciousness as each other? I think when you start to dig down into it, we don’t have any way to tell whether another being is conscious or self-aware other than to ask them.
Let’s look at that for a minute, because self-awareness is a little different. Are you familiar with the mirror test that Professor Gallup does, where they take a sleeping animal, and paint a little red spot on its forehead, and then wait until it walks by a mirror, and if it stops and rubs its own forehead, then, according to the theory, it has a sense of self and therefore it is self-aware. And the only reason all of this matters is if you really want to build an intelligent machine, you have to start with what goes into that. So do you think that is a measure of self-awareness, and would a computer need to pass the mirror test, as it were?
That’s where I think we start to run into problems, right? Because it’s an interesting experiment, and it maybe tells us something about, let’s say, a type of self-awareness. If an animal’s blind, it can’t pass that test. So, passing the test therefore can’t be a precursor to intelligence.
Well, I guess the question would be if you had the cognitive ability and a fully functional set of senses that most of your species have, are you able to look at something else and determine that, “I am a ‘me’” and “That’s a reflection of me,” and “That actually is me, but I can touch my own forehead.”
I’m thinking, sorry. I’m being nonresponsive because I’m thinking about it, and I guess what I’m trying to say is that a test that’s designed for animals that have evolved in the wild is not necessarily a meaningful test for intelligent agents that we’ve engineered, because I could design a robot that can pass that test, that nobody would think was self-aware in any interesting and meaningful sense. In other words, for any given test you design, I can game and redesign my system to pass that test. But the problem is that the test measures something that we think is true in the wild, but as soon as we say, “This is the test,” we can build the thing that passes that test that doesn’t do what we meant for the agent to be able to do, to be self-aware.
Right. And it should be pointed out that there are those who look at the mirror test and say, “Well, if you put a spot on an animal’s hand, and just because they kind of wipe their hand…” That it’s really more a test of do they have the mental capability to understand what a mirror does?” And it has nothing to do with…
Right. Exactly. It’s measuring something about the mirror and so forth.
Let’s talk about another thing in your intelligence definition, because I’m fascinated by what you just kind of outlined. You said that some amount of communication, therefore some language, is necessary. So do you think—at least before we get to applying it to machines—that language is a requisite in the animal kingdom for intelligence?
Well, I don’t think it has to be language in the sense of the English language or our human natural language, but there are different ways to communicate. You can communicate through gestures. You can communicate through physical interaction. So it doesn’t necessarily have to be spoken language, but I do think the ability to convey information to another being that can then receive the information that was conveyed is part of what we mean by intelligence. Languages for artificial systems could be very limited and constrained, so I don’t think that we necessarily have to solve the natural language problem in order to develop what we would call intelligent systems. But I think when you talk about strong AI, which is referring to sort of human level intelligence, at that point, I don’t think you can really demonstrate human level intelligence without being able to communicate in some kind of natural language.
So, just to be clear, are you saying language indicates intelligence or language is required for intelligence?
Language is required for intelligence.
There are actually a number of examples in the plant kingdom where the plants are able to communicate signals to other plants. Would you say that qualifies? If you’re familiar with any of those examples, do those qualify as language in a meaningful sense, or is that just like, “Well, you can call it language if you’re trying to do clever thought riddles, but it’s not really a language.”
Yeah, I guess I’d say, as with most interesting things, there’s sort of a spectrum. But one of the characteristics of intelligent language, I think, is the ability to learn the language and to adapt the language to new situations. So, you know, ants communicate with each other by laying down pheromones, but ants can’t develop new ways to communicate with each other. If you put them into a new environment, they’re biologically hardwired to use communication.
There’s an interesting philosophical argument that the species is intelligent, or evolution is intelligent at some level. I think those are interesting philosophical discussions. I don’t know that they’re particularly helpful in understanding intelligence in individual beings.
Well, I definitely want to get to computers here in a minute and apply all of this as best we can, but… By our best guess, humans acquired speech a hundred thousand years ago, roughly the same time we got fire. The theory is that fire allowed us to cook food, which allowed us to break down the proteins in it and make it more digestible, and that that allowed us to increase our caloric consumption, and we went all in on the brain, and that gave us language. Would your statement that language is a requirement for intelligence imply that a hundred and one thousand years ago, we were not intelligent?
I would guess that human beings were communicating with each other a hundred and one thousand years ago and probably two hundred thousand years ago. And again, I think intelligence is a spectrum. I think chimpanzees are intelligent and dolphins are intelligent, at some level. I don’t know about pigs and dogs. I don’t have strong evidence.
Interestingly, of all things, dogs don’t pass the red paint mirror test. They are interestingly the only animal on the whole face of the earth—and by all means, any listener out there who knows otherwise, please email me—that if you point at an object, will look at the object.
Really?
Yeah, even chimpanzees don’t do it. So it’s thought that they co-evolved with us as we domesticated them. That was something we selected for, not overtly but passively, because that’s useful. It’s like, “Go get that thing,” and then the dog looks over there at it.
Right.
It’s funny, there’s an old Far Side cartoon—you can’t get those things out of your head—where the dolphins are in the tank, and they’re writing down all the dolphins’ noises, and they’re saying things like, “Se habla español,” and “Sprechen sie Deutsch,” and the scientists are like, “Yeah, we can’t make any sense of it.”
So let’s get back to language, because I’m really fascinated by this and particularly the cognitive aspects of it. So, what do you think is meaningful, if anything, about the Turing test—which of course you know, but for the benefit of our listeners, is: Alan Turing put this out that if you’re on a computer terminal, and you’re chatting with somebody, typing, and you can’t tell if it’s a person or a machine, then you have to say that machine is intelligent.
Right, and of course, Alan Turing’s original version of that test was a little bit different and more gendered if you’re familiar.
He based it on the gendered test, right. You’re entirely right. Yes.
There’s a lot of objections to the Turing test. In fact, when I teach the Introductory AI class at UMBC, I have the students read some of Alan Turing’s work and then John Searle’s arguments against the Turing test.
Chinese Room, right?
The Chinese Room and so forth, and I have them talk about all of that. And, again, I think these are, sort of, interesting philosophical discussions that, luckily, we don’t actually need to resolve in order to keep making progress towards intelligence, because I don’t think this is one that will ever be resolved.
Here’s something I think is really interesting: when that test was proposed, and in the early years of AI, the way it was envisioned was based on the communication of the time. Today’s Turing tests are based in an environment in which we communicate very differently—we communicate very differently online than we do in person—than Alan Turing ever imagined we would. And so the kind of chat bots that do well at these Turing tests really probably wouldn’t have looked intelligent to an AI researcher in the 1960s, but I don’t think that most social media posts would have looked very intelligent, either. And so we’ve kind of adapted ourselves to this sort of cryptic, darting, illogical, jumping-around-in-different-topics way of conversing with each other online, where lapses in rationality and continuity are forgiven really easily. And when I see some of the transcripts of modern Turing tests, I think, well, this kind of reminds me a little bit of PARRY. I don’t know if you’re familiar with ELIZA and PARRY.
Weizenbaum’s 1960s Q&A, his kind of psychologist helper, right?
Right. So ELIZA was a pattern-recognition-based online psychologist that would use this, I guess, Freudian way of interrogating a patient, to ask them about their feelings and so forth. And when this was created, people were very taken in by it, because, you know, they would spill out their deepest, darkest secrets to what turned out to be, essentially, one of the earliest chat bots. There was a version of that that was created later. I can’t remember the researcher who created it, but it was studying paranoid schizophrenia and the speech patterns of paranoid schizophrenics, and that version of ELIZA was called PARRY.
If you read any transcripts by PARRY, it’s very disjointed, and it can get away with not having a deep semantic model, because if it doesn’t really understand anything, and if it can’t match anything, it just changes the topic. And that’s what the modern Turing test look like to me, mostly. I think if we were going to really use the Turing test as some measure of intelligence, I think maybe we need to put some rules on critical thinking and rationality. What is it that we’re chatting about? And what is the nature of this communication with the agent in the black box? Because, right now, it’s just degenerated into, again, this kind of gaming the system. Well, let’s just see if we can trick a human into thinking that we’re a person, but we get to take advantage of the fact that online communication is this kind of dance that we play that’s not necessarily logical and rational and rule-following.
I want to come back to that, because I want to go down that path with you, but beforehand, it should be pointed out, and correct me if I’m wrong because you know this a lot better than I do, but the people who interacted with ELIZA all knew it was a computer and that there was “nobody at home.” And that, in the end, is what freaked Weizenbaum out, and had him turn on artificial intelligence, because I think he said something to the effect that when the computer says, “I understand,” it’s a lie. It’s a lie because there is no “I,” and there’s nothing to understand. Was that the same case with PARRY that they knew full and well they were talking to a machine, but they still engaged with it as if it was another person?
Well, that was being used to try to model the behavior of a paranoid schizophrenic, and so my understanding is that they ran some experiments where they had psychologists, in a blind setting, interact with an actual paranoid schizophrenic or this model, and do a Turing test to try to determine whether this was a convincing model of paranoid schizophrenic interaction style. I think it was a scientific experiment that was being run.
So, you used the phrase, when you were talking about PARRY just now, “It doesn’t understand anything.” That’s obviously Searle’s whole question with the Chinese Room, that the non-Chinese speaker who can use these books to answer questions in Chinese doesn’t understand anything. Do you think even today a computer understands anything, and will a computer ever understand anything?
That’s an interesting question. So when we talk about this with my class, with my students, I use the analogy of learning a new language. I don’t know if you speak any foreign languages to any degree of fluency.
I’m still working on English.
Right. So, I speak a little bit of French and a little bit of German and a little bit of Italian, so I’m very conscious of the language learning process. When I was first learning Italian, anything I said in Italian was laboriously translated in my mind by essentially looking up rules. I don’t remember any Italian, so I can’t use Italian as an example anymore. I want to say, “I am twenty years old” in French, and so in order to do that, I just don’t say, “J’ai vingt ans”; I say to myself, “How do I say, ‘I am 20 years old’? Oh, I remember, they don’t say, ‘I am 20 years old.’ They say, ‘I have 20 years.’ OK. ‘I have’ is ‘J’ai,’ ‘twenty’ is ‘vingt’…” And I’m doing this kind of pattern-based look up in my mind. But doing that inside my head, I can communicate a little bit in French. So do I understand French?
Well, the answer to that question would be “no,” but what you understand is that process you just talked about, “OK, I need to deconstruct the sentence. I need to figure out what the subject is. I need to line that up with the verb.” So yes, you have a higher order understanding that allows you to do that. You understand what you’re doing, unquestionably.
Right.
And so the question is, at that meta-meta-meta-meta-meta level, will a computer ever understand what it’s doing.
And I think this actually kind of gets back to the question of consciousness. Is understanding—in the sense that Searle wants it to be, or Weizenbaum wanted it to be—tied up in our self-awareness of the processes that we’re carrying out, to reason about things in the world?
So, I only have one more Turing test question to ask, then I would love to change the subject to the state of the art today, and then I would love to talk about when you think we’re going to have certain advances, and then maybe we can talk about the impact of all this technology on jobs. So, with that looking forward, one last question, which is: when you were talking about maybe rethinking the Turing test, that we would have a different standard, maybe, today than Turing did. And by the way, the contests that they have where they really are trying to pass it, they are highly restricted and constrained, I think. Is that the case?
I am not that familiar with them, although I did read The Most Human Human, which is a very interesting book if you are looking for some light summer reading.
All right.
Are you familiar with the book? It’s by somebody who served as a human in the Loebner Prize Turing test, and sort of his experience of what it’s like to be the human.
No. I don’t know that. That’s funny. So, the interesting thing was that—and anybody who’s heard the show before will know I use this example—I always start everyone with the same question. I always ask the same question to every system, and nobody ever gets it right, even close. And because of that, I know within three seconds that I’m not talking to a human. And the question is: “What’s larger? The sun or a nickel?” And no matter how, I think your phrase was “schizophrenic” or “disjointed” or what have you, the person is, they answer, “The sun” or “Duh” or “Hmm.” But no machine can.
So, two questions: Is that question indicative of the state of the art, that we really are like in stone knives and bear skins with natural language? And second, do you think that we’re going to make strides forward that maybe someday you’ll have to wonder if I’m actually not a sophisticated artificial intelligence chatting with you or not?
Actually, I guess I’m surprised to hear you say that computers can’t answer that question, because I would think Watson, or a system like that, that has a big backend knowledge base that it’s drawing on would pretty easily be able to find that. I can Google “How big is the sun?” and “How big is a nickel?” and apply a pretty simple rule.
Well, you’re right. In all fairness, there’s not a global chat bot of Watson that I have found. I mean, the trick is nickel is both a metal and a coin, and the sun is a homophone that could be a person’s son. But a person, a human, makes that connection. These are both round and so they kind of look like alike and whatnot. When I say it, what I mean is you go to Cleverbot, or you go to the different chat bots that are entered in the Turing competitions and whatnot. You ask Google, you type that into Google, you don’t get the answer. So, you’re right, there are probably systems that can nail it. I just never bump into them.
And, you know, there’s probably context that you could provide in which the answer to that question would be the nickel. Right? So like I’ve got a drawing that we’ve just been talking about, and it’s got the sun in it, and it has a nickel in it, and the nickel is really big in the picture, and the sun is really small because it’s far away. And I say, “Which is bigger?” There might actually be a context in which the obvious answer isn’t actually the right answer, and I think that kind of trickiness is what makes people, you know, that’s the signal of intelligence, that we can kind of contextualize our reasoning. I think the question as a basic question, it’s such a factual question, that that’s the kind of thing that I think computers are actually really good at. What do you love more: A rose or a daisy? That’s a harder question.
Right.
You know, or what’s your mother’s favorite flower? Now there’s a tricky question.
Right. I have a book coming out on this topic at the end of the year, and I try to think up the hardest question, like what’s the last one. I’m sure listeners will have better ideas than I have. But one I came up with was: Dr. Smith is eating at her favorite restaurant when she receives a phone call. She rushes out, neglecting to pay her bill. Is management likely to prosecute? So we need to know: She’s probably a medical doctor. She probably got an emergency call. It’s her favorite restaurant, so she’s probably known there. She dashes out. Are they really going to go to all the effort to prosecute, not just get her to pay next time she’s in and whatnot? That is the kind of thing that has so many layers of experience that it would be hard for a machine to do.
Yeah, but I would argue that I think, eventually, we will have intelligent agents that are embedded in the world and interact with people and build up knowledge bases of that kind of common sense knowledge, and could answer that question. Or a similar type of question that was posed based on experience in the world and knowledge of interpersonal interactions. To me, that’s kind of the exciting future of AI. Being able to look up facts really fast, like Watson… Watson was exciting because it won Jeopardy, but let’s face it: looking up a lot of facts and being able to click on a buzzer really fast are not really the things that are the most exciting about the idea of an intelligent, human-like agent. They’re awfully cool, don’t get me wrong.
I think when we talk about commercial potential and replacing jobs, which you mentioned, I think those kinds of abilities to retrieve information really quickly, in a flexible way, that is something that can really lead to systems that are incredibly useful for human beings. Whether they are “strong AI” or not doesn’t matter. The philosophical stuff is fun to talk about, but there’s this other kind of practical, “What are we really going to build and what are we going to do with it?”
Right.
And it doesn’t require answering those questions.
Fair enough. In closing on all of that other part, I heard Ken Jennings speak at South by Southwest about it, and I will preface this by saying he’s incredibly gracious. He doesn’t say, “Well, it was rigged.” He did describe, though, that the buzzer situation was different, because that’s the one part that’s really hard to map. Because the buzzer’s the trick on Jeopardy, not the answers.
That’s right.
And that was all changed up a bit.
Ken is clearly the best human at the buzzer. He’s super smart, and he knows a ton of stuff, don’t get me wrong, I couldn’t win on Jeopardy. But I think it’s that buzzer that’s the difference. And so I think it would be really interesting to have a sort of Jeopardy contest in which the buzzer doesn’t matter, right? So, you just buzz in, and there’s some reasonable window in which to buzz in, and then it’s random who gets to answer the question, or maybe everybody gets to answer the question independently. A Jeopardy-like thing where that timed buzzing in isn’t part of it; it’s really the knowledge that’s the key. I suspect Watson would still do pretty well, and Ken would still do pretty well, but I’m not sure who would win in that case. It would depend a lot on the questions, I think.
So, you gave us a great segue just a minute ago when you said, “Is all of this talk about consciousness and awareness and self and Turing test and all that—does it matter?” And it sounded like you were saying, whether it does or doesn’t, there is plenty of exciting things that are coming down the pipe. So let’s talk about that. I would love to hear your thoughts on the state of the art. AI’s passed a bunch of milestones, like you said, there was chess, then Jeopardy, then AlphaGo, and then recently poker. What are some things, you think—without going to AGI which we’ll get to in a minute—we should look for? What’s the state of the art, and what are some things you think we’re going to see in a year, or two years, three years, that will dominate the headlines?
I think the most obvious thing is self-driving cars and autonomous vehicles, right? Which we already have out there on the roads doing a great deal. I drive a Volvo that can do lane following and can pretty much drive itself in many conditions. And that is really cool and really exciting. Is it intelligence? Well, no, not by the definitions we’ve just been talking about, but the technology to be able to do all of that very much came out of AI research and research directions.
But I guess there won’t be a watershed with that, like, in the way that one day we woke up and Lee Sedol had lost. I mean, won’t it be that in three years, the number one Justin Bieber song will have been written by an AI or something like that, where it’s like, “Wow, something just happened”?
Yeah, I guess I think it’s a little bit more like cell phones. Right? I mean, what was the moment for cell phones? I’m not sure there was one single.
Fair enough. That’s right.
It’s more of like a tipping point, and you can look back at it and say, “Oh, there’s this inflection point.” And I don’t know what it was for cell phones. I expect there was an inflection point when either cell phone technology became cheap enough, or cell tower coverage became prevalent enough that it made sense for people to have cell phones and start using them. And when that happened, it did happen very fast. I think it will be the same with self-driving cars.
It was very fast that cars started coming out with adaptive cruise control. We’ve had cruise control for a long time, where your car just keeps going at the same speed forever. But adaptive cruise control, where your car detects when there’s something in front of it and slows down or speeds up based on the conditions of the road, that happened really fast. It just came out and now lots of cars have that, and people are kind of used to it. GPS technology—I was just driving along the other day, and I was like, “Oh yeah, I’ve got a map in my car all the time.” And anytime I want to, I can say, “Hey, I’d like to go to this place,” and it will show me how to get to that place. We didn’t have that, and then within a pretty short span of time, we have that, and that’s an AI derivative also.
Right. I think that those are all incredibly good points. I would say with cell phones—I can remember in the mid ‘90s, the RAZR coming out, which was smaller, and it was like, “Wow.” You didn’t know you had it in your pocket. And then, of course, the iPhone was kind of a watershed thing.
Right. A smartphone.
Right. But you’re right, it’s a form of gradualism punctuated by a series of step functions up.
Definitely. Self-driving car technology, in particular, is like that, because it’s really a big ask to expect people to trust self-driving cars on the road. So there’s this process by which that will happen and is already happening, where individual bits of autonomous technology are being incorporated into human-driven cars. And meanwhile, there’s a lot of experimentation with self-driving cars under relatively controlled conditions. And at some point, there will be a tipping point, and I will buy a car, and I will be sitting in my car and it will take me to New York, and I won’t have to be in control.
Of course, one impediment to that is that whole thing where a vast majority of the people believe the statistical impossibility that they are above-average drivers.
That’s right.
I, on the other hand, believe I’m a below-average driver. So I’m going to be the first person—I’m a menace on the road. You want me off as soon as you can. It probably is good enough for that. I know prognostication is hard, and I guess cars are different, because I can’t get a free self-driving car with a two-year contract at $39.95 a month, right? So it’s a big capital shift, but do you have a sense—because I’m sure you’re up on all of this—when you think the first fully autonomous car will happen? And then the most interesting thing, when will it be illegal not to drive a fully autonomous car?
I’m not quite sure how it will roll out. It may be that it’s in particular locations or particular regions first, but I think that ordinary people being able to drive a self-driving car; I would say within ten years.
I noticed you slipped that, “I don’t know when it’s going to roll out” pun in there.
Pun not intended. You see, if my AI could recognize that as a pun… Humor is another thing that intelligent agents are not very good at, and I think that’ll be a long time coming.
Right. So you have just confirmed that I’m a human.
So, next question, you’ve mentioned strong AI, also called an artificial general intelligence, that is an intelligence as smart as a human. So, back to your earlier question of does it matter, we’re going to be able to do things like self-driving cars and all this really cool stuff, without answering these philosophical questions; but I think the big question is can we make an AGI?
Because if you look at what humans are good at doing, we’re good at transfer learning where we pick something to learn in one domain and map it to another one effortlessly. We are really good at taking one data point, like, you could show a human one data point of something, and then a hundred photos, and no matter how you change the lighting or the angle, a person will go, “There, there, there, and there.” So, do you think that an AGI is the sum total of a series of weak AIs bolted together? Or is there some, I’m going to use a loaded word, “magic,” and obviously I don’t mean magic, but is there some hitherto unknown magic that we’re going to need to discover or invent?
I think hitherto unknown magic, you know, using the word “magic” cautiously. I think there are individual technologies that are really exciting and are letting us do a lot of things. So right now, deep learning is the big buzz word, and it is kind of cool. We’ve taken old neural net technology, and we’ve updated it with qualitatively different ways of thinking about essentially neural network learning that we couldn’t really think about before, because we didn’t have the hardware to be able to do it at the scale or with the kind of complexity that deep learning networks exist now. So, deep learning is exciting. But deep learning, I think, is just fundamentally not suited to do this single point generalization that you’re talking about.
Big data is a buzz word, but I’m, personally, I’ve always been more interested in tiny data. Or maybe it’s big data in the service of tiny data, so I experience lots and lots and lots of things, and by having all of that background knowledge at my disposal, I can do one shot learning, because I can take that single instance and interpret it and understand what is relevant about that one single instance that I need to use to generalize to the next thing. One shot learning works because we have vast experience, but that doesn’t mean that throwing vast experience at that one thing is, by itself, going to let us generalize from that single thing. I think we still really haven’t developed the cognitive reasoning frameworks that will let us take the power of deep learning and big data, and apply it in these new contexts in creative ways, using different levels of reasoning and abstraction. But I think that’s where we’re headed, and I think a lot of people are thinking about that.
So I’m very hopeful that the broad AI community, in its lushest, many-flowers-blooming way of exploring different approaches, is developing a lot of ideas that eventually are going to come together into a big intelligent reasoning framework, that will let us take all of the different kinds of technologies that we’ve built for special purpose algorithms, and put them together—not just bolt it together, but really integrate it into a more coherent, broad framework for AGI.
If you look at the human genome, it’s, in computer terms, 720MB, give or take. But a vast amount of that is useless, and then a vast amount of that we share with banana trees. And if you look at the part that’s uniquely human, which gives us our unique intelligence, it may be 4MB or 8MB; it’s a really a small number. Yet in that little program are the instructions to make something that becomes an AGI. So do you take that to mean that there’s a secret, a trick—and again, I’m using words that I mean metaphorically—there’s something very simple we’re missing. Something you could write in a few lines of code. Maybe a short program that could make something that’s an AGI?
Yeah, we had a few hundred million years to evolve that. So, the length of something doesn’t necessarily mean that it’s simple. And I think I don’t know enough about genomics to talk really intelligently about this, but I do think that 4MB to 8MB that’s uniquely human interacts with everything else, with the rest of the genome, possibly with the parts that we think don’t do anything. Because there were parts of the genome that we thought didn’t do anything, but it turns out some of it does do something. It’s the dark matter of the genome. Just because we don’t know what it’s doing, I don’t know that that means that it’s not doing anything.
Well, that’s a really interesting point—the 4MB to 8MB may be highly compressed, to use the computer metaphor, and it may be decompressing to something that’s using all the rest. But let’s even say it takes 720MB, you’re still talking about something that will fit on an old CD-ROM, something smaller than most operating systems today.
And I one hundred percent hear what you’re saying, which is nature has had a hundred million years to compress that, to make that really tight code. But, I guess the larger question I’m trying to ask is, do you think that an AGI may… The hope in AI had always been that, just like in the physical universe, there’s just a few laws that explain everything. Or is it that it’s like, no, we’re incredibly complicated, and it’s going to be this immense system that becomes a general intelligence, and it’s going to be of complexity we can’t wrap our heads around yet.
Gosh, I don’t know. I feel like I just can’t prognosticate that. I think if and when we have an AGI that we really think is intelligent, it probably will have an awful lot of component. The core that drives all of it may be, relatively speaking, fairly simple. But, if you think about how human intelligence works, we have lots and lots of modules. Right?
There’s this sort of core mechanism by which the brain processes information, that plays out in a lot of different ways, in different parts of the brain. We have the motor cortex, and we have the language cortex, and they’re all specialized. We have these specialized regions and specialized abilities. But they all use a common substrate or mechanism. And so when I think of the ultimate AI, I think of there being some sort of architecture that binds together a lot of different components that are doing different things. And it’s that architecture, that glue, that we haven’t really figured out how to think about yet.
There are cognitive architectures. There are people who work on designing cognitive architectures, and I think those are the precursors of what will ultimately become the architecture for intelligence. But I’m not sure we’re really working on that hard enough, or that we’ve made enough progress on that part of it. And it may be that the way that we get artificial intelligence ultimately is by building a really, really, really big deep learning neural network, which I would find maybe a little bit disappointing, because I feel like if that’s how we get there, we’re not really going to know what’s going on inside of it. Part of what brought me into the field of AI was really an interest in cognitive psychology, and trying to understand how the human brain works. So, maybe we can create another human-like intelligence by just kind of replicating the human brain. But I, personally, just from my own research perspective, wouldn’t find that especially satisfying, because it’s really hard to understand what’s going on in the human brain. And it’s hard to understand what’s going on even in any single deep learning network that can do visual processing or anything like that.
I think that in order for us to really adopt these intelligence systems and embrace them and trust them and be willing to use them, we’ll have to find ways for them to be more explainable and more understandable to human beings. Even if we go about replicating human intelligence in that way, I still think we need to be thinking about understandability and how it really works and how we extract meaning.
That’s really fascinating. So you’re saying if we made this big system that was huge and studied data, it’s kind of just brute force. We don’t have anything elegant about that. It doesn’t tell us anything about ourselves.
Yeah.
So my last theoretical question, and then I’d love to talk about jobs. You said at the very beginning that consciousness may be beyond our grasp, that somehow we’re too close to it, or it may be something we can’t agree on, we can’t measure, we can’t tell in others, and all of that. Is it possible that the same is true of a general intelligence? That in the end, this hope of yours that you said brought you into the field, that it’s going to give us deep insights into ourselves, actually isn’t possible?
Well, I mean, maybe. I don’t know. I think that we’ve already gained a lot of insight into ourselves, and because we’re humans, we’re curious. So if we build intelligent agents without fully understanding how they work or what they do, then maybe we’ll work side by side with them to understand each other. I don’t think we’re ever going to stop asking those questions, whether we get to some level of intelligent agents before then or after then. Questions about the universe are always going to be with us.
Onto the question that most people in their day-to-day lives worry about. They don’t worry as much about killer robots, as they do about job-killing robots. What do you think will be the effect? So, you know the setup. You know both sides of this. Is artificial intelligence something brand new that replaces people, and it’s going to get this critical velocity where it can learn things faster than us and eventually just surpass us in all fields? Or, is it like other disruptive technologies—arguably equally disruptive as such things as the mechanization of industry, the harnessing of steam power, of electricity—that came and went and never, ever budged unemployment even one iota. Because people learned, almost instantly, how to use these new technologies to increase their own productivity. Which of those two or a third choice do you think is most likely?
I’m not a believer in the singularity. I don’t see that happening—that these intelligent agents are going to surpass us and make us completely superfluous, or let us upload our brains into cyberspace or turn us into The Matrix. It could happen. I don’t rule it out, but that’s not what I think is most likely. What I really think is that this is like other technologies. It’s like the invention of the car or the television or the assembly line. If we use it correctly, it enhances human productivity, and it lets us create value at less human cost.
The question is not a scientific question or a technological question. The question is really a political question of how are we, as a society, going to decide to use that extra productivity? And unfortunately, in the past, we’ve often allowed that extra productivity to be channeled into the hands of a very few people, so that we just increased wealth disparity, and the people at the bottom of the economic pile have their jobs taken away. So they’re out of work, but more importantly, the benefit that’s being created by these new technologies isn’t benefiting them. And I think that we can choose to think differently about how we distribute the value that we get out of these new technologies.
The other thing is I think that as you automate various kinds of activities, the economy transforms itself. And we don’t know exactly how that is going to happen, and it would have been hard to predict before any historical technological disruption, right? You invent cars. Well, what happens to all the people who took care of the horses before? Something happened to them. That’s a big industry that’s gone. When we automate truck driving, this is going to be extremely disruptive, because truck driver is one of the most common jobs, in most of our country at least. So, what happens to the people who were truck drivers? It turns out that you’re automating some parts of that job, but not all of it. Because a truck driver doesn’t just sit at the wheel of a car and drive it down the road. The truck driver also loads and offloads and interacts with people at either end. So, maybe the truck driver job becomes more of a sales job, you know, there’s fewer of them, but they’re doing different things. Or maybe it’s supplanted by different kinds of service roles.
I think we’re becoming more and more of a service economy, and that’s partly because of automation. We always need more productivity. There’s always things that human society wants. And if we get some of those things with less human effort, that should let us create more of other things. I think we could use this productivity and support more art. That would be an amazing, transformational, twenty-first century kind of thing to do. I look at our current politics and our current society, and I’m not sure that enough people are thinking that way, that we can think about how to use these wonderful technologies to benefit everybody. I’m not sure that’s where we’re headed right now.
Let’s look at that. So there’s a wide range of options, and everybody’s going to be familiar with them all. On the one hand, you could say, you know, Facebook and Google made twelve billionaires between them. Why don’t we just take their money and give it to other people? All the way to the other extreme that says, look, all those truck drivers, or their corollaries, in the past, nobody in a top-down, heavy handed way reassigned them to different jobs. What happened was the market did a really good job of allocating technology, creating jobs, and recruiting them. So those would be two incredibly extreme positions. And then there’s this whole road in between where you’d say, well, we need more education. We need to help make it easier for people to become productive again. Where on that spectrum do you land? What do you think? What specific meat would you put on those bones?
I think taxes are not an inherently bad thing. Taxes are how we run our society, and our society is what protects people and enables people to invent things like Google. If we didn’t have taxes, and we didn’t have any government services, it would be extremely difficult for human society to invent things like Google, because to invent things like that requires collaboration, it requires infrastructure; it requires the support of people around you to make that happen. You couldn’t have Google if you didn’t have the Internet. And the Internet exists because the government invested in the Internet, and the government could invest in the Internet because we pay taxes to the government to create collective infrastructure. I think there’s always going to be a tension between how high should taxes be and how much should you tax the wealthy—how regressive, how progressive? Estate taxes; should you be able to build up a dynasty and pass along all of your wealth to your children? I have opinions about some of that, but there’s no right answer. It changes over time. But I do think that the reason that we come together as human beings to create governments and create societies is because we want to have some ability to have a protected place where we can pursue our individual goals. I want to be able to drive to and from my job on roads that are good, and have this interview with you through an Internet connection that’s maintained, and not to have marauding hordes steal my car while I’m in here. You know, we want safety and security and shared infrastructure. And I think the technology that we’re creating should let us do a better job at having that shared infrastructure and basic ability for people to live happy and productive lives.
So I don’t think that just taking money from rich people and giving it to poor people is the right way to do that, but I do think investing in a better society makes a lot of sense. We have horribly decaying infrastructure in much of the country. So, doesn’t it make sense to take some of the capital that’s created by technology advances and use it to improve the infrastructure in the country and improve health care for people?
Right. And of course the countervailing factor is, do all of the above without diminishing people’s incentives to work hard and found these companies that they created, and that’s the historical tension. Well, I would like to close with one question for you which is: are you optimistic about the future or pessimistic or how would you answer that?
I’m incredibly optimistic. I mean, you know, I’m pessimistic about individual things on individual days, but I think, collectively, we have made incredible strides in technology, and in making people’s quality of life better.
I think we could do a better job. There’s places where people don’t have the education or don’t have the infrastructure or don’t have access to jobs or technology. I think we have real issues with diversity in technology, both in creating technology and in benefiting from technology. I’m very, very concerned about the continuing under-representation of women and minority groups in computing and technology. And the reason for that is partly because I think it’s just socially unjust to not have everybody equally benefiting from good jobs, from the benefits of technology. But it’s also because the technology solutions that we create are influenced by the people who are creating them. When we have a very limited subset of the population creating technology, there’s a lot of evidence that shows that the technology is not as robust, and doesn’t serve as broad a population of users as technology that’s created by diverse teams of engineers. I’d love to see more women coming into computer science. I’d love to see more African Americans and Hispanics coming into computer science. That’s something I work on a lot. It’s something I think matters a lot to our future. But, I think we’re doing the right things in those areas, and people care about these things, and we’re pushing forward.
There’s a lot of really exciting stuff happening in the AI world right now, and it’s a great time to be an AI scientist because people talk about AI. I walk down the street, or I sit at Panera, and I hear people talking about the latest AI solution for this thing or that—it’s become a common term. Sometimes, I think it’s a little overused, because we sort of use it for anything that seems kind of cool, but that’s OK. I think we can use AI for anything that seems pretty cool, and I don’t think that hurts anything.
All right. Well, that’s a great place to end it. I want to thank you so much for covering this incredibly wide range of topics. This was great fun and very informative. Thank you for your time.
Yeah, thank you.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here.
