>>NEIL DEGRASSE TYSON: Welcome to the universe, yes. Hi, everybody. I’m Neil deGrasse Tyson, the Frederick P. Rose director of the Hayden Planetarium, and welcome. Is this our 18th or 19th annual Isaac Asimov Panel Debate. This has become a very hot ticket in New York City and I almost feel apologetic, because we can’t accommodate everyone who wants to see it. We have to go to a lottery model to get tickets out. And short of going to a bigger venue or charging more, we’re still trying to work this out, but you’re here in the audience now, and that’s good. So did they tell you what tonight’s topic is? It’s a very hot topic on every frontier. We’re talking about artificial intelligence and, are people afraid of it? Do people embrace it? Should we be doing it? Should we not be doing it? And it’s all over the news. Not the least of which, in today’s business section of The New York Times, today.
This is a paper version of the news— the newspaper. It’s got a, sort of an android robot holding the national flag of China. And the title is, “China’s Blitz to Dominate AI.” And I just came back 48 hours ago from the United Arab Emirates, and they have a newly-established minister of artificial intelligence. There are countries around the world that see this and recognize it as a way to sort of leapfrog technologies, and I think this is a, there’s another…
Here it is. “China’s blitz to rule AI meets with silence from the White House.” So I just thought I would just say that I’m just saying. We’re trying to burn clean coal. That’s what our priorities are. >>TYSON: But I’m just saying. >>TYSON: Don’t get me started, because… That’s the topic tonight. We combed the country to find some of the top AI people in the land, and we are delighted for this mix of five panelists we have this evening. Let me first introduce to you, who’s right on the wings, Mike Wellman. Michael Wellman is a professor of computer science and engineering at the University of Michigan, and he leads the strategic reasoning group. Michael Wellman, come on. >>TYSON: And… oh, I was supposed to introduce you in a different order than that. Yeah. Yeah, I will get back to you in a minute. >>MICHAEL WELLMAN: Okay. >>TYSON: Just talk among yourselves there, for the… And next up we have a friend and colleague in the astrophysics community who’s directed his attention to AI.
Max Tegmark. Come on out, Max. >>TYSON: Professor of physics. , Max. >>MAX TEGMARK: Excellent, man. >>TYSON: He’s doing research in AI at MIT, and he’s also president of the Future of Life Institute. So Max, welcome. We also have, get my order straight, here. Here we go. When I was… Nope, that’s not it. Yeah. So, Next we have… You couldn’t do this without representation from industry, and that’s precisely what we obtained for this panel. John Giannandrea, come on out. John. >>JOHN GIANNANDREA: Thank you. >>TYSON: John is, he’s a senior vice president of engineering at Google where he leads the Google search and the Google AI teams.
So we got Google in the house. Google in the house. Next, I’ve got Ruchir Puri. Ruchir, come on out. >>TYSON: Ruchir is the chief architect of IBM Watson, and he’s also an IBM Fellow. So we got him. >>TYSON: And we’ve got Helen Greiner. Helen, come on out. Helen. >>HELEN GREINER: Thank you. >>TYSON: Helen, cofounder of the iRobot Corporation, maker of the Roomba. The Roomba. We all know the vacuuming robot. She’s also founder of the drone company CyPhy Works.
She makes drones, now. Is that good or bad? >>TYSON: I don’t know, we’ll find out. Ladies and gentlemen, thank you for coming. This is our panel, everyone. Yes. >>TYSON: So, Mike, you’re a professor at University of Michigan. So what do you do? >>WELLMAN: Well, I study artificial intelligence from the perspective of economics. You know, economics is a social science that treats its entities, its agents as rational beings— ideally rational— >>TYSON: Really. really. >>WELLMAN: Artificial intelligence is the subfield of computer science that’s trying to make ideally rational beings. So it’s a very natural fit. >>TYSON: Can an irrational being make a rational being? >>WELLMAN: We can do our best. >>TYSON: And so you teach a course on this. I’m just curious, how do you frame a course around something that’s so dynamic and so changing and so emotionally fraught with fear. >>WELLMAN: So what I do, and what one does in teaching an AI course is you bring together the standard frameworks and representations and algorithms, techniques that AI people have developed over the years to address thinking-like problems and reasoning and problem solving, decision making, learning using very standard forms of algorithms.
Now, some people are coming to it from the emotional perspective. I sometimes have gotten comments on my teaching evaluations that said, “I signed up for an AI course, and all I got was computer science.” That’s what it is. It’s an engineering discipline, and that’s the most efficient way to make progress. >>TYSON: Excellent. So, Helen, what are you about? >>GREINER: I’m all about the robots. >>TYSON: You’re all about the robots, yeah. >>GREINER: My brother was a huge Star Wars fan when we were young, and, well, for me, it was all about R2-D2, and I’ve wanted to build robots since I saw Star Wars on the big screen. He had everything: character, strategy, loyalty. >>TYSON: You’re telling me Star Wars had like a positive net effect on this world? >>GREINER: I think it had a positive net effect on children.
>>TYSON: Uh-huh. At least one here, yes. >>GREINER: Many, many. So we’ve been trying to build robots like this, and we’ve had great accomplishments, you know? We’ve had robots that have been credited with saving the lives of hundreds of soldiers, thousands of civilians. We’ve got the Roomba, which, best-selling vacuum —not robot vacuum— but best-selling vacuum last year by retail revenue numbers, and I think a little bit of a cultural icon, too. And so I think we’ve come some of the way, but we’re not at R2-D2 yet. So I think some of the debate is about where it needs to go. >>TYSON: So you cofounded the company, iRobot, which I think was the name of an Isaac Asimov novel. Yeah. And that company invested the Roomba. Great word, by the way. Room-ba. Yeah, that’s just good. That was very good. I like words that— >>GREINER: So we asked our engineers what we should call it first, and they said the Mark Master 2000: the Cyber Suck. That’s probably the best marketing dollars that we’ll ever spend to get that name. >>TYSON: Gotcha.
So that cost you money to get that name, okay. >>GREINER: Yeah. >>TYSON: Okay. Does your Roomba count as a kind of AI, would you say? >>GREINER: I believe so. People are starting to use AI to be synonymous with deep learning techniques. But for roboticists, there’s a lot of tools in the tool bag. Roomba runs something called behavior control, which was invented by one of my business partners at iRobot where we have a lot of behaviors that all run in parallel.
The first generation, it wouldn’t fall down the stairs, it did obstacle avoidance, it followed the walls. The latest generation, I think it was something like 13 years later, it does full navigation using a camera system. So visual SLAM techniques. >>TYSON: Okay. Has your Roomba ever killed everyone? >>GREINER: You know, we actually… >>TYSON: Wait, wait. There’s only a yes/no. No! That’s… The… I… >>GREINER: The answer’s certainly— >>TYSON: No sentence. Yes or no? >>GREINER: Certainly no. >>TYSON: Okay, thank you. >>GREINER: But we actually, you know. You know, it’s product design. You have to look at what are the ramifications it could have? Like the worst thing we came up with, maybe it goes into someone’s fire, pulls out the embers, and sets the place on fire. Has never happened, and by the way, they usually have hearths that keep a Roomba out. And screens. But there’s a lot of, you know… >>TYSON: But in the future, you’re… >>GREINER: No Roomba’s killed anyone. >>TYSON: Okay. >>TYSON: Yeah. Ruchir. Ruchir Puri. Did I say that correctly? >>RUCHIR PURI: Yep. >>TYSON: Yeah, thank you.
You’ve been at IBM for more than two decades. >>PURI: Yep. >>TYSON: And so I’m just curious. Before we get to Watson, which you have something to say about, our earliest memories of IBM getting in this game I think was Deep Blue where it was a chess program that beat the world’s best chess player. What made it so good in its day? >>PURI: Well, I think from the point of view, also, I’ve dealt with optimization algorithms for pretty much quarter of the century. And where— >>TYSON: You said optimization algorithms.
>>PURI: Optimization algorithms. >>TYSON: Yeah, so this would just… So that it can calculate as quickly and efficiently as possible towards a goal. >>PURI: It really… There were three things that came together, actually. So search algorithms, really smart evaluation criteria, and the third one is really sort of massively parallel computing application as well. So those three things came together to really give rise to something that, you know, that wow’d people. It’s an application of technology. Again, algorithms coming together from three points of view to give rise to an application that’s really .
>>TYSON: So, but could Deep Blue do anything other than play chess? >>PURI: Interestingly, the Deep Blue was… We have, at IBM, we have something called grand challenges. And we pose these problems to really move the field forward. Deep Blue was a grand challenge posed to the scientists at IBM. Similar to that, actually, Jeopardy was also a grand challenge. >>TYSON: But Jeopardy wasn’t Deep Blue. >>PURI: No. Jeopardy certainly wasn’t Deep Blue, but— >>TYSON: Yeah, but that was Watson, correct? >>PURI: Yes, that’s— >>TYSON: We’ll get to Watson in a minute.
>>PURI: Okay. >>TYSON: I just want to work my way up to that. And I think I have some firsthand knowledge of your grand challenges. I was once invited to address a retreat among IBM engineers where they were given cash rewards for their innovation. Do they still do that? >>PURI: Certainly. We encourage our employees and scientists to really get the innovations out there and get their innovative juices flowing, absolutely, we do that. >>TYSON: Yeah. I was delighted, because each one got recognized, they got a certificate, the CEO was there. >>PURI: Yep. We still do that. >>TYSON: I mean, it was very much taken seriously. >>PURI: Yep. We still do that. >>TYSON: Right. very good. We’ll get back to you on that. So, John, I think I messed up your last name.
Giannandrea? >>GIANNANDREA: Giannandrea. Giannandrea. >>TYSON: Oh, Giannandrea. >>GIANNANDREA: Yes, that’s correct. >>TYSON: Yeah. Giannandrea. So you represent Google on this panel, and could you just tell me, remind us what is the game of Go, and then tell us what AlphaGo is? >>GIANNANDREA: Sure. So Go is this ancient, Oriental board game which is harder than chess. >>GIANNANDREA: And the reason it’s harder than chess— >>TYSON: Okay, Go. >>GIANNANDREA: And the reason it’s harder than chess is because any given— >>TYSON: Well, just to be clear, it’s a board game. You didn’t say that. >>GIANNANDREA: It’s a board game. Yeah, sorry. >>TYSON: If you said it’s a war game… >>GIANNANDREA: No, this is a board game. It’s a board game. >>TYSON: It’s a war board game. It’s not like— >>GIANNANDREA: It’s a strategy game. that. >>TYSON: —weapons and things. >>GIANNANDREA: No. There’s only two pieces: the black pieces and the white pieces. >>TYSON: Okay. Go. >>GIANNANDREA: And the reason it’s hard… And people have been playing this game for 2,000 years and it is highly revered in Asia and people are paid full-time jobs to be professionals at this game.
And the reason it’s hard is because any given position on the board, there are many, many more positions that you could take. So you can’t use brute force approaches to figure out how to play the game. >>TYSON: So intuition has a very big role. >>GIANNANDREA: So the recent systems that have become very, very good at this game, you could even say super human at this game, because they beat the world champions, they’re doing something fundamentally new. And people look at that and use words like intuition— which is not a technical word—and… >>GIANNANDREA: You know, so there’s something going on. >>TYSON: Who invited you to this? >>GIANNANDREA: But it’s a serious issue, because I think when people use words like that, when a chess grandmaster is beaten by Deep Blue, or when the world champion in Go was beaten first in Korea, Lee Sedol, it’s an emotional toll on that player, because they just spent their entire life perfecting their ability to play this game.
And then a machine comes along and appears to beat them using… and the words that are used are like creativity or intuition or that’s something I didn’t expect it to play. And so I think that adds to the mystique of AI when actually what’s going on is engineering, plain and simple. >>TYSON: So brute force. >>GIANNANDREA: No. In the case of the AlphaGo system, it was a combination of training and new algorithms to do so-called deep learning, which I’m sure we’ll get into. >>TYSON: Okay. So AlphaGo was trained on previous games that had been played? >>GIANNANDREA: Yes. There’s two versions of it. The one that won the world championships was trained on all human games that it could get his hands on and then played itself.
So it basically practiced after it had learned the— >>TYSON: How quickly could it play itself and finish a game? >>GIANNANDREA: Well, we do it in the cloud with thousands of computers so it could do it, you know, thousands of times at the same time. >>TYSON: So… Okay. >>GIANNANDREA: Very fast. >>TYSON: Very fast, okay. >>GIANNANDREA: And then most recently there were— >>TYSON: And it’s just in the cloud? >>GIANNANDREA: In the cloud, that’s right. >>TYSON: Up there somewhere. >>GIANNANDREA: Lots and lots of computers. >>TYSON: But the computing cloud, not the storage cloud.
>>GIANNANDREA: That’s right. the computing cloud. >>TYSON: Yeah. Yeah. >>GIANNANDREA: So recently there’s been a version of this called AlphaGo Zero, and the interesting thing about this— >>TYSON: So that’s an upgrade. >>GIANNANDREA: This is a later version. And what they tried to do with that is the researchers tried to see if they could learn to play Go without looking at any human games. >>TYSON: That way it would come up with stuff on its own. >>GIANNANDREA: Yeah. And the AlphaGo Zero was actually better than the ones that learned from humans, and it also plays chess very well. >>TYSON: I’ll try to find other questions for you later. We’ll see. >>GIANNANDREA: It doesn’t do Jeopardy, though. >>TYSON: Okay, so that learned… So it taught itself, basically— >>GIANNANDREA: Yeah. >>TYSON: —and was not biased by the creativity of any human game that had previously been played. And so that, you play that game against AlphaGo— >>GIANNANDREA: Another copy, yeah.
>>TYSON: And it beat AlphaGo. >>GIANNANDREA: Yeah, that’s right. >>TYSON: So it’s extra badass. >>GIANNANDREA: Yeah. Now, games are a special thing, because games have an objective score. And so it’s not… It’s actually a good test for this level of the current technology. >>TYSON: So, Max, we go back. Great to have you. This is like your fifth time here in the museum. It’s not even your first Asimov panel, so thanks for showing up again. You recently published a book, Life 3.0. It’s your third or fourth book… >>TEGMARK: Second.
>>TYSON: Second? Okay. It feels like three books. >>TEGMARK: They just take so long to read— >>TYSON: That’s what it is, yeah. >>TEGMARK: —because they’re so boring. >>TYSON: Your first book was Our Mathematical Universe and thinking of all of the universe as, as a simulation, basically. And we had you on the simulation panel last year. Life 3.0, what’s that about? >>TEGMARK: It’s… Well, my day job is working at MIT doing AI research from a physics perspective, these days. So I like to take a step back and look at things, and if you— >>TYSON: A cosmic perspective. >>TEGMARK: Yeah. And if you look beyond the next election cycle and all these near-term AI controversies about jobs and stuff like that, then it’s pretty natural to ask, well, what happens next? What happens if all these folks succeed and ultimately make machines that can do everything we can? The earliest life that came along, I call it 1.0, because it was really dumb stuff like bacteria that couldn’t learn anything in its lifetime.
And then I call us 2.0, because we can learn. >>TYSON: Oh, you’re referring to the evolutionary achievements in the tree of life. >>TEGMARK: Yeah, yeah. >>TYSON: Okay. >>TEGMARK: And what comes next? I think we should think about this, because if the only… the only strategy we have is to say, “Hey, let’s just build machines that can do everything cheaper than us, what could possibly go wrong,” you know? I think that’s just pathetically unambitious and lame, you know? We’re an ambitious species, homo sapiens. We should aim… we should aim higher. We should say like, “How can we use all this technology to empower us; not to overpower us?” >>TYSON: Okay. We’ll need more of that, I’m sure, as this conversation progresses. Let me get back to Mike. Mike, does… Could you remind us about the Turing test, what that is? >>WELLMAN: Sure. Alan Turing, back in 1951— >>TYSON: The movie The Imitation Game is a biopic on him. >>WELLMAN: Right, and it does depict the Turing test a bit.
So back in 1951, he proposed this thought experience realizing that, to try to get people to understand machines as being able to think would require defining thinking, and that would be very controversial. So he set up this thing that became called the Turing test. That is, see if you could have a machine have a dialogue with somebody and convince them that they’re a person rather than a machine. If a person in an interrogation could not tell the difference between whether they were speaking with a machine or a person, then you might as well say it’s thinking. This is really audacious in 1951. Think about what machines were like back then. People hadn’t even thought about— thought of word processing yet, and they were thinking about AI. That test, I think, has been very useful as a thought experiment. The field of AI has never really generally accepted that as the goal of AI or the definition of AI. But certainly— >>TYSON: Is that because you’ve evolved past that? We do have machines that sound like they’re not machines, but people. So once you hit that goal, you say, oh, we need a better goal.
And are you just moving the goalpost? >>WELLMAN: So we haven’t hit the goal. So it turns out Turing didn’t realize that it would be easy to fool a lot of people, even without being very good at thinking. >>TYSON: It reminds me, was it a New Yorker comic where two dogs are at computers, and one turns to the other and says, “The good thing about the Internet is that no one knows you’re a dog.” >>WELLMAN: That’s right. And no one knows you’re a bot either, and that is a potential way that AI is going to affect us and be ubiquitous. So it is quite relevant to try to impersonate people.
But we use that as a gateway to a lot of Internet activities. You do a CAPTCHA, that is called a computer automated-something-Turing. I forget the exact acronym. >>TYSON: The T in CAPTCHA stands for Touring? >>WELLMAN: Yes. >>TYSON: Oh, I didn’t know that. Cool. >>WELLMAN: Or, it’s basically you have to tell the machine— >>TYSON: We’ve all done it. >>WELLMAN: —that you’re a human. >>TYSON: Yeah. >>WELLMAN: So find something that only humans can do. And of course, that bar keeps on moving all the time. So it’s quite relevant to try to impersonate for the Alexas and the Siris in the world are trying to be as humanlike as possible. In films, we try to put and videogames realistic characters all the time.
So it still speaks to us, even though it’s not the whole story about AI. >>TYSON: Right. so your point is we did so well with satisfying the Turing test very early, that it just wasn’t good enough a discriminator for the AI that people were seeking. >>WELLMAN: Well, again, I would say that being specifically like a human is only one way to be intelligent. And you could be superhuman in many other ways, and you don’t stop when you reach human level-performance in particular tasks, because the goal is not to be like a human. The goal is to make ideally rational intelligence that could do all sorts of things. >>TYSON: So, Helen, with the company you cofounded called iRobot, could you tell us about, what is it, the three laws of robotics, by Isaac Asimov? >>GREINER: Yeah, definitely. The robots could not…
The three laws: One, the robots cannot hurt humans or cause humans to come to harm through an action. Robots cannot— >>TYSON: That was one. >>GREINER: That was one. >>TYSON: So they— >>GREINER: Robots have to obey orders. >>TYSON: —cannot harm you, and their inaction also can’t harm you. >>GREINER: Yeah. They have to obey orders, unless it conflicts with number one. So they can’t… I’m sorry. The second one is they have to obey orders unless it conflicts with number one. And the third one is they cannot harm themselves, unless it conflicts with number one and number two.
And there’s one he added later on, the zeroth one— >>TYSON: I didn’t know that. >>GREINER: —which is robots cannot cause harm to humanity or through an action, have humanity come to harm. >>TYSON: So it generalizes it up from the individual? >>GREINER: Yeah. >>TYSON: Humanity… The… >>GREINER: Yeah. Well, he made that the zeroth law. So he stuck it in the front when he thought of it. >>TYSON: The zeroth law, okay.
>>GREINER: But what’s amazing about it is he… he started writing the I, Robot books in 1940. Practical transistors weren’t invented till 1947, so, I mean, one of the reasons we’re all so honored to be here at the Asimov memorial debate, is I think we— I can speak for the panel that we’re all huge, huge fans of what he was writing about, especially way back. >>TYSON: Well, just consider that he’s written about— on topics quite diverse. So no matter what subject we have here, there are books that he wrote about it. So… >>GREINER: Yeah. >>TYSON: Every panel we’ve ever had on— >>GREINER: But AI’s a really good one for him. >>TYSON: —on any subject is, I read that by Asimov when I was a kid, so. >>GREINER: That’s why people ask me, are you putting those in the robots? And the short answer is, they’re great as a literary device. They’re a little bit more tricky to program. And so, unfortunately, the answer is that the state of technology is not ready for those types of abstract rules yet.
>>TYSON: But they’re nice guidance; just philosophical guidance, I guess. >>GREINER: Um, I have a very practical view. I think the laws if you state them now, might be, of robots, can save people, they have saved people, and they could save a heck of a lot more people. It might be that robots… >>TYSON: Well, plus, the military would not be obeying those laws. >>GREINER: Yeah, exactly. exactly. >>TYSON: Right. >>GREINER: And the whole books were about how those laws resulted in conflicts, right? But in reality, because I’m a business woman as well as a robot lover, robots are not going to hurt people. >>TYSON: Don’t say a robot lover. That’s just… doesn’t… >>GREINER: I am. I— >>TYSON: Just find some other phrase. >>GREINER: I’m a robot enthusiast. >>TYSON: There you go. thank you. >>GREINER: Robots are not going to hurt people. They’re not going to hurt themselves. They’re not going to do these things, because they’re going to be either scrapped, they’re going to be sent back, or someone’s going to be sued.
And so from a business standpoint, the robots are going to be safe to operate. >>TYSON: One of my favorite… Well, no. A video that I found amusing was a cat riding around on a Roomba. >>GREINER: You know, that got so many views, and I have no idea why. >>GREINER: I mean, it’s like tens of millions of views, right? It’s crazy. >>TYSON: I mean, if the Roomba were big enough for me to sit on, I would do that. That’s… >>TYSON: Wouldn’t you? that’s fun. >>GREINER: That was not in our brainstorming sessions when we thought about all the applications for robots. >>TYSON: So, Ruchir, could you get us from Deep Blue to Watson? What happened in that transition? And if we can remind people why we all know about Watson, there was the big contest that you guys entered it in. >>PURI: Certainly. And let me pick up the thread from, from the chess and the Go and the, you know. Let’s… >>PURI: Let’s make this— >>TYSON: Okay, continue.
>>PURI: —interesting, actually. >>TYSON: Continue. >>PURI: — >>TYSON: By the way, Deep Blue beat Kasparov when Google had 10 employees, okay? >>PURI: True. That’s true. >>TYSON: So just, like, where were you? >>PURI: That’s true. absolutely true. >>TYSON: Okay. Did I get you on that one? >>PURI: Yes, . >>TYSON: Okay. Take us there. >>PURI: So the journey continues from, from the point of view of chess game that beat Kasparov to, we went down, to, okay, what is next? And obviously natural language— >>TYSON: Kasparov was the world champion at the time. >>PURI: Yes, at the time. And natural language, which is so fundamental to humans, actually. And the intricacies of natural language, as we’ve been… At least there’s one fundamental trait that humanity has, which is just the proliferation of language, the advent of language itself. So we decided that will be the next leap that we are going to make.
And there is no game better than Jeopardy that captures the intricacies. So we posed that as a grand challenge, and— >>TYSON: Jeopardy, not only language, but culture. >>PURI: But culture, right. >>TYSON: Yeah. It’s not a calculation anymore in a traditional sense. >>PURI: It is certainly not a calculation anymore, and the way the questions are posed are so nuanced that you really are dealing with, at this point in time, not just a calculation machine and simple evaluation criteria and search algorithms and parallel computing, but really understanding language, questions and answers, and the way we interact as human beings. So that was really the advent of the next challenge. Because once we are able to solve that, the implications are phenomenal in terms of the benefit it can bring to us as a society, which is where we took that level to. The first thing that we started right after Jeopardy was the applications of that technology to the health domain, which is so fundamental to all of us. So right from chess game, the next challenge is really addressing fundamentals of what defines us as humans in terms of communications, addressing those intricacies, and then applications of that abound.
>>TYSON: To serve needs in actual society. >>PURI: Yeah, absolutely. >>TYSON: So Watson, in principle, can become the best doctor ever, because Watson can read all the research papers and then interpret symptoms in the context of what is known worldwide, rather than just what one doctor happened to learn. >>PURI: Absolutely. And, at least the way we think about it is really not— It’s not about, does it become the best doctor, but, as we all know, no physician, single physician has the time, even if they have certainly the intelligence to figure all of this out, they don’t have the time to figure all of that out. And as Max was saying, it’s really about empowering professionals than necessarily overpowering them. And really, Watson is about empowering the society, as opposed to overpowering it, and that’s why I really think about, it’s bringing capabilities whereby, yes, it can read millions of studies and millions of trials that may be going on, and there some well-publicized cases as well where it had actually saved patients either in North Carolina or Tokyo or a study that was published more recently in India as well.
But it’s, from our point of view, it’s really about bringing the technology together with the human beings in what we call augmented intelligence. >>TYSON: So in all fairness to our understanding of this, Watson only knows what is available on the Internet, correct? >>PURI: Yeah. Watson only knows what is being fed to it. Let’s say it that way. Whether it is available on Internet or it is private information… >>TYSON: So how does Watson know what is fake… news or not? >>TYSON: You can make a super machine that cannot distinguish the two. Well, apparently humans can’t either, but… >>TYSON: In principle, we—educated—can make a judgment. Will Watson be in a position to make that judgment? >>PURI: I think, at least regarding fake news, the question really is on, we are all pushing the boundaries of that technology, and yes, the machines need to be trained, and they can really help us, given what has gone on, actually, in the last couple of years. Once they are actually— you bring that technology to bear in terms of realizing there is a problem, you can actually correct for it. So it’s not about whether Watson can distinguish it today or not.
Once you realize the problem, you can actually start working on technologies that can start deciphering that much better, thereby helping us, as society, because, you know, what is going on around. >>TYSON: So, but that… From what you’ve described, Watson would still be shy of this Holy Grail of just thinking stuff up on its own without reference to…
I mean, when you think of the most creative people there ever were, sure, there’s some foundation from where that you could trace that creativity. But for many of them, there’s a spark, and something new comes out of them that had no precedent. So, from what you described, Watson is capable of digesting preexisting knowledge, but in its current state, or at least the state we’re familiar, it is not inventing something new. >>PURI: Yeah. Certainly, the purpose of the technology today is really not about that spark in itself. Although, it will find out… in particular find out, it’ll find insights that you didn’t know existed, actually. Although they were hidden in there, you didn’t know existed, so it may be an ah-hah moment for you, I got it, but still, it existed there.
It didn’t… So it will actually do that. But, yeah, it wouldn’t get that… The notion you are saying, hey, that was spark, no, it doesn’t have. >>TYSON: So, John, tell me about the future. We could spend a whole panel on this, but I just want to put it on the table briefly: What’s the role of AI in the future of autonomous cars? And I know you guys are working on this. >>GIANNANDREA: Yeah, we are. >>TYSON: You entered certain autonomous car chal— You, your company. >>GIANNANDREA: Yeah, we have a division of Alphabet that works on this. >>TYSON: Just to be clear, the holding company is Alphabet— >>GIANNANDREA: Yeah. >>TYSON: —and Google is one of several companies under Alphabet— >>GIANNANDREA: That’s right. >>TYSON: —and one of those companies were tasked with making the autonomous car. >>GIANNANDREA: One of those companies is called Waymo, and they’re one of many companies that are making autonomous cars.
So it’s a super hard problem. I think people have been working on it seriously for more than a decade. They’re making progress. These cars have driven millions of miles with very small numbers of incidents, but they’re still pretty constrained. They’re more accurate than a human driver, but they’re limited in where they can go. So for example, the kinds of streets that they can drive on, the cities, and so on and so forth. But the technology is progressing fairly dramatically. I’m pretty confident to say that we will have fully autonomous cars for most of the large car manufacturers within a decade. >>TYSON: And what role does AI play in that, or is it just really good programming? >>GIANNANDREA: Well, it’s machine learning. So, you know, these systems have a lot of computers on the car, they can detect a stop sign or can figure out that there’s an impediment in the road or a kid just ran into the road or there’s a cyclist. In California, we have this weird thing where motorcycles are allowed to drive between the lanes— >>TYSON: You have many weird things in California.
>>GIANNANDREA: We have many weird things in California. >>TYSON: Uh-huh. >>GIANNANDREA: But motorcycles are allowed to drive between the lanes of the cars, and so for the computer to actually understand what’s going on and figure out what’s safe and not safe is actually quite hard. I think one of the things that’s going to happen here is even if you don’t see millions of autonomous cars like in three years, most of the new cars that you buy will have semiautonomous features in them, like automatic braking or telling you what the speed limit is. >>TYSON: Which we’re all accustomed to and expect it on our next cars, now. >>GIANNANDREA: Yeah. yeah. So I think this technology kind of comes in increments. It’s not like a big-bang thing, you know? And I’ll just echo this comment about augmentation, because the phrase AI, it means so many different things to so many different people that it’s really hard to kind of pin down what it is.
But the idea of augmented intelligence has been around for a very long time. A lot of the ideas we have in computing today came from the work of Doug Engelbart back in the ‘50s, and he had been describing computers as being a tool; a tool that can help a doctor look through more information, that can help pinpoint something in an x-ray. Not something that would replace the doctor, and that’s how we think about it. >>TYSON: Which is Max’s point. Not be… Just… What’s the two words you put together? >>TEGMARK: Oh, empowered versus overpowered. >>TYSON: Overpowered, yeah, that’s right. Very good. I like that. Can you describe for us what’s the difference or what is the ascent from AI to general AI? Because we hear this term general AI— >>TEGMARK: Yeah. >>TYSON: And what’s going on there? what have we been talking about so far, and if it’s not general AI, what is? >>TEGMARK: It’s really important to be clear on what we mean by intelligence.
As you mentioned correctly, John, different people mean different stuff, right? I think it’s a really good idea to go into the footsteps of Helen, here, and make a very broad definition of intelligence. So even Roomba is intelligence. And just define intelligence simply as the ability to accomplish complex goals, you know? So Roomba has a very narrow intelligence: really good at vacuum cleaning. Today we have a— >>TYSON: Was that a diss on… >>TEGMARK: I am a proud Roomba owner. And we… >>TYSON: Roomba can carry cats around, okay? >>TEGMARK: Yeah.
>>TYSON: For all we know the Roomba is like the Uber for cats in the house, okay? >>TEGMARK: Yeah. So… >>TYSON: Wouldn’t that be cool if cats could like, get the Roomba to come and take them around? Get the Roomba to open a door for them, yeah? >>TEGMARK: That’s right. So today, we have many areas… So if you define intelligence as the ability to accomplish complex goals, then there are many areas today where machines, in narrow domains, are already much better than us. Not just vacuum cleaning and high frequency trading and multiplying large numbers together and stuff like that, but also now in playing chess and playing Go and so on. But no machine today that we’ve built— >>TYSON: No single machine. >>TEGMARK: No single machine, not even the whole Internet combined has the broad intelligence of a human child, who, given enough time, can get quite good at anything.
So this is what’s meant by artificial general intelligence, or acronymed AGI, which has been the Holy Grail of artificial intelligence ever since Marvin Minsky and McCarthy and others founded it, came up with the whole… founded the field in the ‘60s. And now— >>TYSON: But, Helen, you come to this from a product— a consumer product point of view. And I want to get back to what you just said. People who are making AI want to sell something. So they’ll sell you something that cleans the room, that drives the car, that does any one of the things that help our lives. Who’s going to buy something that has general intelligence? >>TEGMARK: Well, everybody. >>TYSON: And will the general intelligence be as good at the pieces of it as the specific products that industry would be making for that one need that you have? >>TEGMARK: Oh yeah, by definition. So if people say that they think that machines will never be able to— there will always be jobs left for humans— they’re just saying, by definition, that AI researchers will fail to build artificial intelligence, because that’s the very definition of it, that machines can do everything better than us.
And many people… Like I have had many conversations where you’re— >>GREINER: I’d like to point out— >>TEGMARK: Yeah. Let me just finish my sentence. >>GREINER: —there’s a mechanical and a sensing component as well as the, what you’re calling AGI, mechanical and a sensing element to make these machines better. >>TEGMARK: Sure. But anyway— >>TYSON: Oh, good point. So you can have software, but if it doesn’t have the physical means to enact what it’s supposed to, it’s just a box. >>TEGMARK: No, no. it can do some great stuff. Like you could feed it a photograph, and it could tell you if you have breast cancer or something like that, right? But it’s not going to go out and sweep your house. >>TEGMARK: Yeah. But I think the final word on definition should go to Shane Legg, one of the leaders of Google DeepMind, because he coined the phrase, and he simply meant something that can do the same information processing that the human brain can do.
And if you hook it up to good enough robots, which I’m sure we can build, then it can do great stuff. And so that’s the goal of certain companies, Like Google DeepMind, for example, to try to build that. And that’s why they keep trying to push the envelope, right? >>TYSON: Wait. But I gotta go to my industry people. What does it mean to buy something that has general AI? What do I do with that? Do I say, make me the best cup of coffee, drive me to my office, what’s the square root of two, and… I mean, in practice, is that a thing? >>GIANNANDREA: So in principle, and this is highly speculative, but in principle, an AGI could build any kind of other AGI, and therefore, could build you any machine you want it to build.
And that’s when people worry— >>TYSON: That’s when we all die. >>GIANNANDREA: No, no, no, no. That’s when a class of people who call themselves transhumanists would say that humans would evolve. And I personally don’t believe in this. I see no evidence that it’s going to happen. But that’s the source of a lot of the ethical discussions— >>TEGMARK: Right. Exactly. >>GIANNANDREA: —about this topic. >>TYSON: Mike, speaking of ethics, could you tell us about the trolleyology and what role AI can play in assisting our reasoning there? >>WELLMAN: So probably many of you have heard about trolley problems.
This became popular in psychology to pose ethical dilemmas to people and see how they react. And there’s many variations of it, but the standard kind of story is a trolley is going down a track and it’s about to hit or kill three people, and then you notice that there’s a switch, and you could make it go over to another track where there’s only one person. And you could choose to kill that other person instead of the first three.
Would you do it? >>TYSON: Wait, wait. So the dilemma there is somebody’s going to die no matter what. You either can not touch it, then the trolley kills three people on its own, or you can intervene and actively kill one person. >>WELLMAN: Right. Now— >>TYSON: Right. >>WELLMAN: —I’m not a psychologist, but I think it… It seems to be a kind of a silly question to ask people, because humans can really never get, I think, into a mental state where they could really believe that, with certainty, if this action, if I take this action, I’ll kill this one person for sure, and the other action… There’s always this uncertainty. There’s always questions about what the blame… It’s not actually a realistic situation. So the question is, will AIs, actually maybe, is it more realistic for them, perhaps? Could an autonomous vehicle be in a situation where, all of the sudden a bicyclist runs in front of it, and it has a chance to swerve and do some other damage, and will it have to weigh that.
>>TYSON: You would have to take out the vegetable cart first and then find out what else it does, yes. >>WELLMAN: Yeah. And, you know, so will they have to be coded in them what the solutions are to those dilemmas? When it does happen— >>TYSON: Wait, wait. Wait. So that implies that humans get together, figure out a solution, and you hand it to AI. But that’s not the point of AI. The AI is going to have some higher intelligence than we do, and that’s why I’m curious. >>WELLMAN: So I actually— >>TYSON: If you bring AI to that problem, is it going to give different answers than we would, and then we say, oh my gosh, we never thought about it that way. Let’s do it that way. >>WELLMAN: So I think this is that’s going in some of the session, here. Actually, no. AI, the idea is we want to give— for the humans to give the AI the values, and the AI is concerned with making decisions and taking actions to promote those values. So ultimately, we are saying… you know, we value life, we value… That’s part of what the robot laws are for.
>>GREINER: There are no robot laws. They are science fiction. >>WELLMAN: And so the danger is that they would be weaponized by the party that is programming them and is controlling them; not that they’re going to all of the sudden decide to get rid of the humans. That’s not the source of the danger. With respect to the trolley problem situation in this hypothetical autonomous vehicle, when it does happen that a car, one of these cars runs over a bicyclist— and it’ll happen, I think, much less frequently than humans do it today— we’ll take the black box— I hope they’re engineered so that they have a black box that captures everything that was in their senses all the time and it’s very secure, so they can’t lie about it— and they will be able to dissect it and will say, “You made this decision.
Why did you do that?” And it might say, “Well, I hit the bicycle, because if I swerved to the left, I would have run over a child.” Or if it said, “Well, I did that, because if I swerved too fast I’d wake up the passenger,” then you’d say, no, that was the wrong decision. That was not what we meant for you to do. It’s still better than what the Tesla said a couple years ago— >>GREINER: Yeah. >>TYSON: Yeah. If I made it say, don’t wake me up for any reason— >>WELLMAN: That’s right.
>>TYSON: —and it’s the robot’s job to obey me… >>WELLMAN: Exactly. That’s exactly an answer… So this is part of the danger of AI is that the unintended consequence of the specification of the values won’t hit what you really care about. >>TYSON: Let me ask Google and IBM, here. In your efforts in this, I don’t want to call it a race, but let’s call it an exploration, is there a tandem sort of ethical group? Let me start over here in IBM. Is there a… Is anyone thinking about the ethics of what AI would do if you achieved this goal? Because we certainly have sci-fi movies, and none of them… It never ends well, okay? In any of them. Any of them. >>PURI: Yeah. So certainly, we were one of the first companies to actually bring principles of ethics and responsibility to AI. It’s captured sort of ways in what we do overall on the information we have. But most importantly, there are three fundamental tenants we go by as it pertains to AI.
One is building AI with responsibility. Second one is building AI that’s unbiased. And the third one is building AI that’s explainable. I think those are the fundamental tenants that we drive and strive towards, and we have, in our research teams, we have significant number of people and scientists and experts that really try to drive our— the AI services that we offer, the solutions that we build with tremendous number of businesses around, to drive them with those three principles. And obviously, I think we all know the way AI techniques work these days, they are driven a lot by the data. And you are as good as, as we were just discussing before, you are as good as the data that you are fed.
And detecting bias in the data itself is actually one of the more important research and technical challenges. And having techniques that are able to de-bias that data as well, in terms of when you are learning, you know that there is bias in the data— or be able to de-bias it so that you can build models that are actually unbiased. So that’s why I said there are three fundamental principles that we go with that are sort of very formal and engrained in the principles through which we are driving AI. >>TYSON: Speaking of bias, John, if I remember correctly, there were some fascinating studies recently where Google facial recognition software was not as good at identifying black people as it was with white people. And then they found out that just white people programmed it, so that’s… >>TYSON: So, um… So maybe that’s just kind of obvious at that point.
But that would, I think, count as a bias. >>GIANNANDREA: I was actually at lunch with one of the authors of that paper today. They haven’t actually measured our systems. They measured other people’s systems. But it’s a serious issue, and I think that— >>TYSON: So it wasn’t your facial recognition? >>GIANNANDREA: It wasn’t ours. But this issue of bias and machine learning is super important. >>TYSON: I’m sorry to have implicated you. >>GIANNANDREA: No, no. That’s okay. It’s okay. So we think that this is, at least for the next few years, the most serious ethical issue.
I think this AGI stuff is years, decades away, so I don’t spend very much time on this. But this question of you’re building learned systems, machine learned systems learning from data, if your data is biased, you’re going to build biased systems. And this could be everything from whether to give somebody a mortgage or what their credit score prediction would be or there are people selling systems now that are used by Quartz to predict recidivism rates. And they’re not explainable, and it’s not entirely clear what data they used to train them, and we think this is just unethical. >>TYSON: So it’s garbage in; garbage out, at that level. >>GIANNANDREA: Yeah. And so— >>TEGMARK: And we know that one was very biased, yeah. >>GIANNANDREA: Yeah. So many of our companies work together outside of the commercial realm with academia, but also in nonprofits looking at this question, because we really worried about building systems that give a bad name to all this machine learning. >>TYSON: So in all of your efforts, how would you characterize the, sort of the ethical dimension of what’s going on? You have people…
Are they philosophies, are they psychologists, what are they? >>GIANNANDREA: No. They’re usually data scientists and researchers who are looking for systemic bias in the systems and the data that we’re using to train the systems. But we have significant efforts with . >>TYSON: Okay. So I get the bias part, but how about the trolley car part where we… Will the AI have the values we care about if it will properly serve us? If the AI achieves consciousness and then comes up with values of its own… >>GIANNANDREA: I mean, our company has very few situations, autonomous vehicles would be one, where we have to actually struggle with these issues. Mostly, we’re worried about recommendations systems giving bad reconditions to people, or ranking systems giving bad results to questions that you ask. >>TYSON: But this is moving fast as a field. >>GIANNANDREA: I think as a field it’s moving fast, and I think academia is now got entire classes on AI ethics and machine learning ethics. And I think society’s responding in an appropriate way, because we’re worried about this stuff.
>>TYSON: So, Max, you’re president of the… Tell me the name. >>TEGMARK: Future of Life Institute. >>TYSON: Future of Life Institute. Sounds very New Age-y, by the way. >>TEGMARK: Well, future of life; we’re for it. >>TYSON: Okay. >>TEGMARK: We would like it to exist. >>TYSON: Not a controversial— >>TEGMARK: You would think. >>TYSON: Put on that on Twitter, and then people would argue with it for sure. >>TEGMARK: You would be surprised, yeah. >>TYSON: So could you tell me the difference between an “is” and an “aught” philosophically, and how that matters in AI? >>TEGMARK: Yeah. You know, it basically comes— >>TYSON: Was it Hume who did this? >>TEGMARK: I believe so, yeah.
>>TYSON: But one of the philosophers, yeah. >>TEGMARK: I think so. It basically comes down to, you know, saying that might makes right is a really lousy foundation for morality. Just because something is in a certain way doesn’t mean that’s the right way, and just because by default something is going to happen if we don’t pay attention to it doesn’t mean that’s what we really want to happen. You know, I’m very optimistic that we can use AI to help life flourish like never before, if we win the race between this growing power of AI that we’re seeing, and the growing wisdom that we need to manage it. And there, I feel we’re kind of a little bit asleep at the stick. You said here—sorry to pick on you, John— >>TYSON: Well, I don’t want any AI person to say we’re asleep at anything. >>TEGMARK: But I have to pick on you a little bit, John— >>GIANNANDREA: Pick on me. >>TEGMARK: —because you said, “Well, you know, I think this AGI stuff is kind of decades away, so I’m not thinking about it much.” But I bet you wouldn’t say, “I think this climate change stuff is a few decades away, so I’m not thinking about it,” right? You look young and healthy, you’re working out, taking your vitamins, you’re going to be around then, right? And if it’s going to take a few decades to get this right, it feels really important right now to think about it enough that we can— >>GIANNANDREA: I totally agree.
>>TEGMARK: —steer things in a good direction. >>GIANNANDREA: What I said was I don’t spend very much time at Google with researchers on this task. But we do invest in groups around the world at Oxford and Berkeley and other places who are looking at this stuff. >>TEGMARK: Yeah. And you’re a member of the Partnership of AI, which is awesome. >>GIANNANDREA: Yeah. It’s not that we’re abdicating responsibility. It’s the, we just have no idea what the timeline is. We do know what the timeline is for global warming. >>TEGMARK: Yeah, and— >>TYSON: Well, if anyone knows the timeline of this, it would be you, presumably. >>TEGMARK: Well, I think also we do know quite a bit about the timeline. First, we know there’s great controversy. And your cofounder, Rodney Brooks, told me in person he thinks DeepMind’s quest for AGI is going to fail for at least 300 years, right? But most AI researchers in recent surveys think it’s actually going to succeed, you know, maybe in 40 years, maybe in 30 years.
So that, to me, means it’s not too soon to start thinking hard about what we can do now that will be helpful. >>TYSON: I get it. But I want to get back to the point of, there are things that are, and there are things that ought to be. >>TEGMARK: Yeah. >>TYSON: Do you trust AI to judge what ought to be? >>TEGMARK: No. >>TYSON: Or is this… Okay, good. >>TEGMARK: I could give a longer answer, too. >>TYSON: And how do you imbue what ought to be in an AI, if an AI is a higher level of consciousness and capacity than we are.
Maybe it knows better than we do. >>TEGMARK: Yeah. But people often tell me, if AGI is by definition smarter than us, why don’t we just let it figure out morality; what’s ought to be. But the fallacy in this of course is that, you know, artificial intelligence and technology in general is not good or evil. It’s morally neutral. It’s a tool that can be used to do good or to do evil. Intelligence itself is simply the ability to accomplish goals, good or bad, right? If Hitler had been more intelligent I think that would have sucked, right? So I wouldn’t want to delegate to him for that reason what we should do. Instead, we should take the attitude we take when we raise kids. We often raise children who are— end up being more intelligent than us.
We don’t just ignore them for 20 years and hope they… something good comes out of it. We really try to… While they’re still young enough that they listen to us a little bit, right? >>TYSON: A little bit; little bit. >>TEGMARK: We try to instill in them values that we think are good. And I think this is linked back to what you were saying about let’s teach morality to machines. >>TYSON: You said in the next 20 years we still have a chance to teach AI who and what we are so that when it achieves consciousness, it will not exterminate us. >>TEGMARK: Well, it’s even harder, though, than raising kids. >>TYSON: It’ll keep us around as pets.
>>TEGMARK: It’s tough though, because— sorry if I get a little nerdy, now— but with children, we can’t teach them morality when they’re six months old, because they just don’t get it. And like with my teenage son, it’s kind of too late, because they don’t listen to me anymore. >>TEGMARK: But there is this magic window we have over a few years when they’re actually smart enough to understand us and still, maybe we have some hope that they’ll adopt our morality. Where AI— >>TYSON: Whereas AI in that— >>TEGMARK: It has… It’s not yet reached the point where it understands human values, because we can’t explain it yet, but it might pretty quickly blow through this window where it’s actually going to… where it’s still not as smart as us and we can influence it. And we have to kind of plan this curriculum, plan this ahead, you know? And I think it’s really good that you are working on that, for example, so that we can… Because we don’t want to wait until after someone has— or the night before someone switches on a super intelligence to be, oh, how do we figure out this, you know, teaching it right from wrong stuff? That’s probably too late.
>>TYSON: So, Mike, there’s a… Probably too late, yeah. That’s certainly too late if that happened. So, Mike, I’m curious about something. The capital markets, I don’t want to say that they rely on this, but they, a lot of what makes them fluctuate is that different people have different information that they’re betting on if they buy and sell stock. So if you make a machine that has access to all information and is perfectly rational, is that machine or the person who owns that machine the first trillionaire in the world? >>WELLMAN: So interestingly, Wall Street trading is one of the first areas where autonomous agents are really out there. And I think that’s one of the reasons why it’s useful to study long-term implications of AI by this case study of seeing what’s happening right now.
And right now, lots of firms not very far from here— >>TYSON: This is New York. >>WELLMAN: —are programming computers and putting their al… using machine learning and using a lot of data, and a lot of the same data to make decisions. So one question is, well, if everyone is using the same data and maybe stumbles on the same algorithms, are there possible effects on the stability of markets that, if something goes wrong, could they be more prone to crashes or not? That’s something that we’re studying. And if so, are there things that we could do to try to mitigate that? The question you asked about the first trillionaire is if one group, one firm, one country has an edge in AI, will they be able to then leapfrog everybody else and just suck up all of the resources? That’s actually a significant issue.
Financial markets is one place where the money is, and if you really get it so much better than everybody else, there could be major shifts in distributions of wealth. And it’s not only financial markets. It could be the Internet. You can put smart AIs out there and say, “Find some way to make money for me,” and they will. So headline, China’s Blitz to Dominate AI is what you’re showing. >>TYSON: Right. So you’re saying a country can just corner the market if they get there first. >>WELLMAN: So this is somewhat, I think, ill understood and controversial, but certainly in this longer road to more general, more capable AI, if one entity has a significant edge, they will have a very strong incentive to shut others out and to capitalize on that advantage.
And so there’s, no doubt, there’s an arms race dynamic to many aspects of artificial intelligence technology that perhaps is most frightening in the military realm, but also comes up in financial realms and other ways. It’s in the fake news realm. We were talking about if AI’s going to be better at discriminating fake news. Never mind that. They’re going to be much better at promulgating fake news, and that’s going to be a challenge for all of us. >>TYSON: This could go to any one of you.
It could go to Helen. Helen, what… Do you foresee robots or AI in general informing political policy? Because if they can analyze— Look at Watson. Watson reads a thousand medical papers and comes up with some conclusions based on it. So you have machine—you make machines, you make drones that can make decisions that we can’t, and they can make them more quickly, and presumably, better. So is there a scenario where, here are political factions arguing, because, really, their feelings are involved more than facts. And at the end of the day, in an informed democracy, you kind of want facts to matter… I would think. >>TYSON: Just, I would just… >>GREINER: We are— and I’m a little bit on the other side of it— we are very far away from this AGI, generalized AI, and there’s wonderful progress being made that allow AI systems and robots to do more than they could do before in recognition and characterization, but we haven’t made that leap, and it’s going to take an innovation step to get there.
So to really worry about that now, I mean, right now, the machines are feeding information into the system, and humans are making the judgment. Now, I believe that day will come, but it’s unpredictable. Because as an innovation… Maybe innovation steps would have to happen before that day comes. >>TYSON: Okay, so it’s not… Because in innovation, you can’t order up an innovation. >>GREINER: Yeah. You don’t know when it’s going to happen. Hopefully some of the younger people in the audience will make those innovations, because I think we should have it happen. >>TYSON: So, Ruchir, it just seems to me, given that Watson might be uniquely qualified to come up with a political policy decision, if it reads every consequence of every political decision that’s ever been made, looks at what became of it, looks at how people reacted, looks at what people wanted, and then just said, “You should do this.” So there should be maybe a machine on the floor of Congress and people come up to it and ask it, right? It would be like the oracle of Congress.
It could be Watson, right? Let’s check… I’m arguing in the dining room with my political colleague from across the wall— across the hall, uh, the aisle— and we say, “Let’s go check Watson.” >>PURI: Are you telling me to print posters, Watson 2020? >>TYSON: Watson AlphaGo 2020, yes. Uh-huh. >>PURI: So first of all, I think let’s take the question— precisely the question you asked. Could AI be helping public policy? And to that, I’ll answer, absolutely yes. >>GREINER: Yep. >>PURI: It could be helping public health policy, it could be helping public policy as it pertains to decisions that are within the country as well— whether it is taxation or other scenarios— absolutely yes. And it already is, actually. So I will not just say it should be helping. It already is helping. Now the question really on the table is have we reached a scenario where there is this oracle, actually, that knows everything. And, no, we have not reached that scenario yet. We are farther— >>TYSON: Yet? >>PURI: The reason I’m saying that is because it’s really about domains that you specialize in, actually, and that information is spread in those domains.
So just as an example, we are working towards an in compliance domain; a regulatory compliance. And yes, we can actually feed information to the machine, and it learns, and it’s going to find insights, and, for example, obligations that a particular entity may have. But I think by oracle, everybody understands it to be know-all, actually. It knows everything, it reacts to everything, and we have not reached that point. Neither is the intention to reach that point whereby you know everything, you react to everything. The point is that really be precise in scenarios that’s going to help society, whether it is in healthcare domain or whether it is in public policy domain or it is in a compliance domain. So that’s where lot of the benefit to society is going to come from. At least as engineer and scientist, I would say let’s be more precise, let’s define the problem and solve the problem in domain, and then we make the progress from there, just like what we did in the scenario of we looked at chess, we defined the next problem— that’s really the next level up in terms of the language— you solve that problem, and you move on from there.
>>WELLMAN: Maybe your question… if human-level intelligence might be hard, what about Congress-level intelligence? >>WELLMAN: But I think that’s not really fair. >>TYSON: Well, that’s how that saying goes: If pro is the opposite of con, then progress is the opposite of Congress, right? Ever hear that one? Anybody ? >>TYSON: That one goes way back, yeah. Yeah. >>WELLMAN: But I think it’s true. Once we agree on the values, then AI can be a great help in sorting out the policy questions. And of course it’s not that Congress is not intelligent. It’s that it’s all about fighting about the values and the priorities. >>TYSON: Right. >>WELLMAN: And that problem doesn’t go away when you have AI. >>TYSON: Helen, can you foresee a future where robots get angry with people? >>GREINER: Um, I think that we can put in simulated emotions to help with decision making.
I think that you can also have it to do a more natural interaction with people that respond how a human would respond. But not in the way that you might think of a person as being angry, for a while, until some of these other innovations come out. >>TYSON: So there’s a video of all of the occasions where they abuse their own robots. So they have robots that are walking, and then they just kicked them, and then they… So, I mean, it’s interesting, because… >>GREINER: I think you can tell a lot about a person about how they treat a robot, by the way. >>TYSON: Well, that’s my point. So these are robots that you almost kind of feel for them, because some of them are sort of humanoid rather than non-humanoid.
And the early ones, they would just sort of fall over. And, I get it, they’re trying to increase the stability of these robots. So now they’re poking them and pressing them, and then the robot rebalances and comes back. >>GREINER: But they get lots of complaints about it, don’t they? >>TYSON: I know. In the— >>GREINER: Like, that’s being mean to the robot. >>TYSON: Don’t be mean to the robots. >>GREINER: But I think there is going something going on, which you hit the nail on the head, that— >>TYSON: Because I think that all of robots will have memory.
>>GREINER: When we had— >>TYSON: And the first time they achieve consciousness… >>GREINER: There’s been studies that people name their Roombas. They get attached to them. Our military robots, too, when we put them out in the field, we had big Marines come into the robot hospital saying, “Can you fix it?” And it’s all blown up, and, you know, he didn’t want any other robot, he wanted that one, because it had gone on missions for him. It had done like over 18 missions, and they named it Scooby-Doo.
>>TYSON: Did you just say… If I hear you correct… >>GREINER: And, you know, they’re big, tough military guys, but, because they’re working with the robot and because the only things they experience, they have this kind of behaviors of animals. It’s like, it’s not anthropomorphizing. I think there’s another word that could be like thinking something’s sentient like sentipomorphizing. Maybe we’ll make up a word, coin a word for that. >>TYSON: I love that word. From here on, sentip… >>GREINER: Sentient. >>TYSON: Sentipomorphizing. >>GREINER: Yeah, exactly. >>TYSON: Right. So you’re saying military who have been served by a robot, if the robot blows up because it found the mine, and… >>GREINER: Mm-hmm. They want it back. They want it fixed. >>TYSON: —then they take the pieces and they go to the robot doctor and say, can you fix him, doc? And these are big, burly Marines. >>GREINER: Yep. And we say, you can have another one. They say, no, I want this one, because its name was Scooby-Doo and it saved, you know, it saved 11 guys on one mission, right? >>GREINER: And there have been reports of people giving them burials, people, military service members— >>TYSON: They buried their robots? >>GREINER: Yeah.
Giving them field promotions and— >>TYSON: Do they know that microbes— >>GREINER: —viewing them with personalities, saying this one’s tough; this one’s a little bit wimpy. I’ve had people tell me that they’re sure their Roomba moved a pot into the way of the virtual wall so it could escape. I can assure you, it didn’t figure that out. It really accidentally did it. But it’s that sentipomorphizing that people automatically do, and it’s wonder… it’s kind of cool, right? >>TYSON: If you bury a hunk of metal, microbes won’t eat it. It’ll just still be metal later on. >>GREINER: We saved Scooby-Doo. We brought him back, and he’s… he’s at the iRobot headquarters, that one.
>>TYSON: He’s repaired. Yeah. I want to sort of land this plane, but I want to do it in a way that… Because there’s still some really important pieces of this conversation we have not addressed, because you all are kind of, I don’t know, you’ve been shy of this threshold that I want to take each one of you. At some point… Well, let me lead up to it. So I have a calculator on my hip, and it calculates better than any human who ever lived. So in a sense, it’s a superhuman property that it contains that we built. Now, you can go down the list of computer-run things that do them better than the best human ever could have or ever will. And that list is growing, okay? And autonomous cars will be among. It will drive a car better and faster in a more controlled than any human who ever lived. So as these accumulate, it doesn’t seem to me to be a stretch to ask if general AI achieves some kind of conscious state— whatever that is, however we define it— that that consciousness will be a superhuman consciousness.
Is that…? You’re shaking your head, Mike. >>WELLMAN: No, I’m nodding. >>TYSON: No, no, no. You’re nodding. Mike is shaking. >>GIANNANDREA: I’m shaking my head, yeah. >>TYSON: Yeah. Why are you shaking your head? >>GIANNANDREA: Because having more smart tools that are superhuman at very narrow things, like calculating or driving or diagnosing cancer is not the same thing as having a consciousness and having AGI. We’ve had more tools for the last 200 years— that calculator you’re talking about, you didn’t have 50 years ago— it doesn’t make us less human. It frees us up to do more things. I remember when my daughter was in school, they wouldn’t let her use a calculator to do homework, which, with 20 years of hindsight, seems absurd, right? But just because you have these tools and you can use a— >>TYSON: That’s what I’m asking. >>GIANNANDREA: But it’s not inevitable— >>GIANNANDREA: —it’s not inevitable that if you have more— >>TYSON: That’s not what I’m asking.
>>GIANNANDREA: But you’re making the leap. You’re saying that if you have more of these tools, then you’ll have AGI, and I disagree with that. >>TYSON: No, no. No. Okay, I can see how you’d think that. That was not my intent. I’m saying these tools are evidence to me that the day general AI arrives, there’s some decision-making power that it will have that will be superhuman. Because everything else we created using computers and put a lot of thought behind became superhuman in that way. Is it unfair to imagine for the safety of us all whether general AI would have superhuman consciousness? >>TEGMARK: I think it’s very likely. I think we humans are so stuck on the idea that we are like the pinnacle of how smart it’s possible to be, and we have a long tradition of lack of humility, right? But let’s face it. Our intelligence is fundamentally limited by Mommy’s birth canal width, and the fact that— >>TYSON: Explain that, please, because that’s… >>TEGMARK: —we’re made of these blobs of neurons and, they’re pretty cool, our brains, but there’s nothing— >>TYSON: Wait.
Pause. Pause right there. >>TEGMARK: —special about that level. >>TYSON: Just to be clear, we could have had bigger brains, but we would have killed our mothers in every birth, okay? So we have basically the biggest possible brain to be born out of your mother without killing her. And so that’s it. That limited how big our brains could get. >>TYSON: It’s already an issue, getting the damn head out of there. So… >>GIANNANDREA: But you’re comparing two different things here, right? >>TEGMARK: Yeah. I’m talking about AGI. >>TYSON: Am I right? I’ve read that, right? >>GIANNANDREA: You’re comparing one person’s brain size with the sum total of humanity.
Like there’s seven, eight billion of us. We communicate with language. We hopefully cooperate. That is way more powerful than a single AGI. >>TEGMARK: Sure. I don’t necessarily disagree, but what I’m saying that if once we future out— if we figure out how to make AGI, suppose that happens in 35 years, then there’s no reason to think that it’s going to stop there and become like in all those lousy Hollywood movies where we have all these robots which are roughly as smart as us, and that’s it, and we just become buddies with them and go drink beer with them. It’s very likely that they will just continue dramatically getting better and they can now start developing even better robots and they will be as much better at everything as they are today at multiplying large numbers.
>>TYSON: This is my… That’s the foundation of my inquiry. Mike, where are you on this? >>WELLMAN: I’m with Max. I see no boundary and reason that that wouldn’t occur. The timing is very uncertain. And I think this uncertainty is also a part of the equation that we have now about being prepared for it, because it could happen faster than we think. It could happen slower than we think. There could be obstacles that make it really far, but we just don’t really know. But, you know, it’s true, you put a lot of brains together, but we have very minimal communication channels between us.
This linear speech that we’re doing, compared to what computers do when you build together and have them talk, they can do just so much more. So I think there really is… They’re already super intelligent in many ways, not just your calculator. Everything we do, it doesn’t stop. They’re not… The algorithmic traders that I talked about don’t at all stop at whatever human traders can do. >>GREINER: So I believe we are machines made of biological components, so I think that we will eventually be able to duplicate and improve upon. But the problem is when you discount timing at all, and what’s being done, like, you know, these bag of tricks are not going to get us there. There’s core inventions that have to happen, potentially different hardware than running a machine, right? There’s a lot of stuff that has to be happen that… And if you want them to be mobile, have better sensors, better mechanics, as well as all the AI. So I think it’s… You say, well, why shouldn’t we worry about it now? Well, because it’s not very close, you know? In 2000, Bill Joy started writing about how these threats to humanity, and one of them was robots.
I start getting calls from like Wall Street Journal and everywhere at iRobot saying, “What kind of human robots are you making?” And I’m like, you know, I couldn’t say it then, because we hadn’t launched the Roomba yet, but, “We’re making a robot vacuum,” you know? Don’t… Because it gets people… >>TYSON: Yeah, but what else does it do? >>GREINER: Yeah. But it gets people maybe focused on the wrong things rather than what these new achievements that AI are just getting to, because they think it’s becoming general AI, and it’s really not yet. And there are many of them on the stage which would like it to, but it’s not. >>TYSON: Mike, let me just ask. My deepest skepticism that this will go the way people imagine, especially in the movies, is we don’t really understand consciousness right now in humans. So it’s not obvious to me that we can just assert by fiat that a smart enough computer will achieve consciousness, when we don’t even understand it within ourselves.
And there was in interesting bit in the movie I, Robot, I don’t remember if it was captured in the book itself by Isaac Asimov, but they noted that, because they didn’t replace code with new code every time they upgraded the robots, every generation of robot had this baggage of software that was just dangling there kind of like our brains with leftover wiring from long before we became human; different parts of the brain. Evolution doesn’t swap that out and make it fresh. It builds around it, and it’s gotta deal with it.
We have to deal with it behaviorally. Our primal nature has to be overcome by later brain revelations that we got from natural selection. My point is, in that film, they asserted that this extra dangling software made the robots do things that the intended software— that the latest software—did not intend. And so, in a way, it was almost like a free will was emerging in them. The robots would do things. And they said, “Well, I didn’t program that in.” Well, that’s leftover wiring from 20 years ago. I don’t know what I just did there. I don’t know what that is. So evidence that we don’t understand consciousness is you go to the bookstore and there’s shelves upon shelves of books on consciousness. That’s evidence that we don’t understand it, because people are still writing books on it. You go to the shelf and ask for books on gravity, there’s like, two books, okay? We got this one.
So where does it come from that people just declare that general AI will have consciousness? single >>TYSON: Oh. thank you, that one person who… Yeah. >>WELLMAN: I don’t understand consciousness, But I also don’t think it’s really, even necessarily has to be part of this discussion. I mean, when you have an AI that is super intelligent in every way, it can do any job as well as any person can do, every capacity, whether it has whatever we think of as consciousness and has that same, you know, illusion of free will and way of thinking about itself seems to be maybe beside the point.
We’re still faced with an issue about dealing with entities like that, whether or not we correspond on the consciousness question. >>TEGMARK: Yeah, I agree with Mike there that, whether it’s conscious or not doesn’t have to affect at all how it treats us. Maybe it should affect how we treat it, right? From an ethical perspective. But I also think we should all remember… >>TYSON: Maybe they’ll come up with their three laws.
>>TEGMARK: Maybe. >>TYSON: Yeah. But robots should not harm a human. No, no. Humans should not harm a robot. >>TEGMARK: Exactly. >>TYSON: Yeah. >>TEGMARK: But we should also remember, I think, this famous quote of Upton Sinclair who said that it’s very hard to get a person to understand something when his or her salary depends on not understanding it. And I find it— no offense to the three of you here who are from companies— but it’s been so striking, so striking how— >>TEGMARK: —every time there’s a debate like this that I’m in, it’s always the academics who are like, “Yeah, this might happen,” and the people from the companies are always like, “Everything will be fine.” >>TEGMARK: I would love to ask you the same questions over beer when it’s not on camera.
>>TYSON: That’s why I flank the three of you academics. That was all very much on purpose. We’re going to open the floor to questions in just one moment. If I could just get some summative reflections. Let’s start down here. Should we fear AI? And if so, on what level. Keep short. >>TEGMARK: Yeah. It’s a, should we fear fire or not? Should we love it? I mean, AI is in an incredibly powerful tool, and it’s either going to be the best thing ever to happen to humanity, or the worst thing ever. I don’t think the question is whether we should stress out about it.
I think the question is— —what stuff should we do now? >>TYSON: No. Max, Max, you just said, “It could be the best thing for the… or the worst thing ever, but we shouldn’t stress.” That is the definition of stress. >>TEGMARK: I meant we should… It’s interesting this quibble about how stressed you should be. The interesting question is what should we do that’s useful to maximize the chances that this will be awesome? Because if we really work hard for this, I really do think that AI can help us crack all the toughest challenges we have that are facing us today and tomorrow and create a really inspiring future. But we’re going to have to work for it. It’s not going to just happen if we’re asleep at the wheel. >>TYSON: John? >>GIANNANDREA: So my problem with this question is we didn’t, in this whole hour, define what we mean by AI, right? So there are some very smart people who think that AGI is inevitable, and that it has ethical implications and so on and so forth.
My beef with that is there’s lots of technical reasons to believe that it’s not inevitable, or, I agree with Helen, that it’s… We just have no idea what breakthrough after breakthrough after breakthrough would be required to go from the kind of practical AI we have today, to the kind of AI we’re talking about conjecturing here. So I’ll give you one example. Small children can learn from small numbers of examples. Today, we have to give computers hundreds of thousands or millions of examples. A child that learns to play chess can also play tic-tac-toe, right? Our Go program can’t play tic-tac-toe, unless we program it to do so. So there’s these huge barriers to generality of intelligence, and as a technologist and as an engineer and somebody working in industry, I see no evidence of this stuff imminently going to happen. That doesn’t mean we shouldn’t be having the academic, ethical conversation. I just don’t see any evidence of it. Now, the reason that’s a problem is because it scares people, and it scares people into thinking that everything with this AI label is scary, and so then people think that we shouldn’t be doing healthcare with AI or we shouldn’t be doing better data science or we shouldn’t be doing decision support or autonomous vehicles.
And yet, if we build these systems, they won’t have the ethical problem that we’re conjecturing, and yet, they will do a tremendous amount of great things for our humanity. And we’re conflating the two things, and we’re scaring ourselves into not doing what we should be doing, which is saving people’s lives. >>TYSON: So there’s a cultural rational barrier that you’re up against, here. >>GIANNANDREA: Yes. >>TYSON: Okay. Ruchir? >>PURI: I think… Well, AI is an extremely powerful tool. I do not believe we are anywhere close to this fear mongering or, by some people, and the fear that exits.
And I can understand certainly, I think a narrative can be raised to a point where you really start fearing it. I’ll give a very good example, and I think just picking on John’s thread—I talk about this in the talks I give often as well— our two daughters, and when they were young, we had like two books of A is for apple, B is for boy, C is for cat. And you look at—you show them… And they were in love with only one book, anyway. Doesn’t matter how bad it was.
And you show them a picture of cat, only one picture of cat, and you repeat that multiple times over several days or a month, and then you show them a picture of a cat they have never seen before, and they say in their cute voice, “Cat.” It takes, roughly speaking, today, a computer 750 pics of a cat to recognize it’s a cat. Now, I’ll give a good example. If I ever showed my daughter 750 pictures of cat when she was less than one year old, she’s 16 right now, she’ll be confused till today, actually, what a cat was.
So we are so far away from actually whatever we are discussing that I find that question humorous, almost, and I have encapsulated that as a syndrome called cats and dog syndrome, actually. So I’ll leave it there. >>TYSON: All right. Helen. Helen? >>GREINER: So you shouldn’t fear technology. You should be concerned and maybe do something about AI, for example, cyber hacking into AI systems, people using an AI system maliciously, unconscious bias in the AI system but you really don’t need to worry about general AI yet. >>TYSON: Yet. Okay. Mike? >>WELLMAN: So I think it’s really important to keep aware of this distinction between the short-term, narrow AIs which have their own concerns and, you know, safety concerns and societal concerns with them separate from the long-term general AI super-intelligent concerns which are of a different magnitude and different and probably much further away. But we, as a scientific field, and certainly as a society I think, we can think at multiple timescales and make these distinctions all the time.
I think if we are… don’t… If we refuse to talk about the thing that’s over the horizon, we’ll lose credibility if we deny that there’s a potential problem. I think that that is a way to make sure that, just, we keep our head in the sand. There are things that we really should be figuring out way in advance of this potential super intelligence, and— >>TYSON: Whether we’ll all die. >>WELLMAN: Well, our children we care about, and— >>TYSON: Whether our children will die. >>WELLMAN: And even if they don’t, how well they’ll live with those superintelligence . >>TYSON: As pets of superintelligence. >>WELLMAN: Well, we hope as… in a good partnership with them. >>WELLMAN: We hope.
>>GREINER: sucking up to the AI already? >>WELLMAN: That’s the best we can do. >>TYSON: Is that the best you got, here? >>WELLMAN: That’s the best I can do. >>TYSON: We hope… that our children will be in partnership with AI. >>WELLMAN: I think that’s a fair way to… way to sum it up. And I’ll stop. >>TEGMARK: Okay. Just in defense of Mike, here, there is so much more detailed description in all the world religions of hell than of heaven, right? Because it’s always much easier to think in all the ways we can screw up than to think about good outcomes. That’s why you’re giggling when you’re trying to say what you’re hoping for. But that doesn’t mean we shouldn’t try. It’s incredibly important that we change… You were making fun of Hollywood for just never showing us any future that doesn’t suck, right? Blade Runner, or whatever.
We really need to start thinking about what kind of future with advanced AI would we find really inspiring? And this is not something you should just leave to tech nerds like us here, right? This is something everybody should think about. And the more clear vision we share for what sort of future we want, the more likely we are to get it. >>TYSON: Do you detail this in your book, Life 3.0? Do you go there? >>TEGMARK: I talk a lot about it. I try very hard not to give any glib answers, because this is really a question we should all discuss. But you don’t do good career planning by just listing everything that could go wrong. >>TYSON: Although… >>TEGMARK: You have to envision success. >>TYSON: Although—I will only be able to paraphrase this quote from Ray Bradbury— when asked… the great science fiction writer, futurist, they asked him, “Why do you keep portraying these dystopic futures? Do you think that’s what the future of life will be?” And he says, “No.
I portray these futures so that you know what future not to head towards.” That was Ray Bradbury. Ladies and gentlemen, thank you for your attention this evening. Join me in thanking this panel. Let’s open up the stage. We’ll have about 10 minutes for questions. We have a microphone on each aisle. And if you try to direct your question to one panelist, that will go faster than saying, can I get all five of them to reply. So are we ready? Let’s start it off right over here. >>AUDIENCE: Hi, Neil, how are you doing? >>TYSON: Hey, how are you doing? >>AUDIENCE: All right, good.
I wanted to get a little bit back to the artificial intelligence in the vehicles, and the more complex scenario of— and I read a little bit about this in California cars— where, is it… You have a scenario, the school bus, the bicycle, the kid, or a hundred-foot cliff. And the AI decides the best thing to do is to drive the car off the hundred-foot cliff, because that’ll cause the less damage, but it’s going to kill you. Is that something that would be learned, or a decision that it will make? How can it avoid making that decision where a human factor might say, hey, there’s no one in the school bus, the bicycle might be able to make it, you know, at a glimpse, as opposed to just those simple, I don’t know, algorithms or decisions that an intelligence that it would make: kill the driver; save everyone else.
>>TYSON: Yeah. Mike… I mean, John, why don’t you take that? >>GIANNANDREA: Well, I think all of these systems have— distinguish between the learned part, like a detector for a stop sign, and the policy part. I think it’s very important that the policy part be explicitly planned for and then you end up with all the ethical issues about what do you want your policy to be. Ideally, you would just stop the bus, right? >>TYSON: Right. That you have brakes good enough so you don’t have to drive it off the cliff. >>GIANNANDREA: Yeah. Hopefully you saw the cliff far enough in advance.
>>TYSON: In the first place. >>GIANNANDREA: Yeah. >>TYSON: Right. So it may be that so many of these scenarios you described are real-life scenarios that human beings, in our frailty encounter, but it calculated the rate that the bicycle was entering the street, it knew what its breaking distance is, so maybe it would just be better at it, and we’re troubling ourselves over scenarios that are real for humans and highly unlikely for autonomous AI, I would imagine. >>TYSON: Next up, yes. >>AUDIENCE: Okay. You were talking about the eventual future of artificial intelligence as general intelligence. There was something discussed several years ago called the singularity, when intelligence gets to the point, both human and artificial sort of blends together. >>TYSON: Was that a question? >>AUDIENCE: Yeah. Do you consider this idea of a singularity be a possibility? >>TYSON: Sure. Mike? >>WELLMAN: All right. So the singularity usually refers to something that’s also been called the intelligence explosion: a point where there’s a kind of a critical mass where something becomes so smart—Max alluded to this before— where it can then further self-improve at a rapidly-accelerating rate.
It’s quite controversial whether that phenomenon will happen. It’s hard to really rule it out. There’s also… It’s hard to rule it in as well. It’s not clear that it’s really necessary to achieve super intelligence, that it goes through this super-accelerating phase, but that’s one scenario where it could happen faster than we realize. >>TYSON: And thereby not be a linear extrapolation into the future about when it occurs, because if it grows exponentially, what looks like small today becomes very large very quickly. Agree? >>TEGMARK: Exactly. >>GREINER: That how it works at Google, so we should have Google answer. >>TYSON: Okay. Google, where are you on this exponential curve? >>GIANNANDREA: I mean, what I would say about this is people who have been marketing this notion that the singularity is inevitable, and there are people who will say that, many of them that I’ve talked to actually want it to happen.
And I just don’t think they’re being rational about the likelihood of it happening. That’s my personal view. >>TEGMARK: Yeah. And many of the people who say it’s never going to happen don’t want it to happen. So we have to be very mindful . >>GIANNANDREA: That’s true, too. >>TYSON: All right. next question over here. Thank you for that. >>AUDIENCE: Hi, how’s it going? >>TYSON: hey. Good. >>AUDIENCE: My question is, If I have— >>TYSON: That’s very New York, you know? Hey, how’s it goin’? Good. How you doin’? We’re doin’ good. Yeah? >>AUDIENCE: My question is, if you have the artificial intelligence, or the AIG or whatever, and it comes to harm or kill you and you pull the plug on it, is that murder? Because it’s a full, intelligent-like sentient machine that you’re pulling the plug on.
>>TYSON: Let me go to Max on that one. So if we judge value to our society by level of sentence, and then there’s an AI— we’re already burying AI robots or repairing them as though they’re humans— so do you think the day will come where laws protect the lives of robots? >>TEGMARK: First of all, if a human comes to try to murder you, and the NYPD pulls the plug on him, that’s already the law today, right? So there has to be some sort of protection of… in there. You can’t do anything you want just because you’re conscious. Second, I think it’s a, aside from the very difficult science question which we have to solve about what kind of information processing even is conscious, there’s… It’s certainly not as simple as just saying, oh, you know, all consciousness is equal, if you’re as smart as the human and there’s conscious— you know, one consciousness, one vote— because then if you’re a computer program and you’re only getting 10% for your favorite candidate, just make a trillion clones of yourself and have them all vote, right? There are a lot of really challenging questions here that we need to face, which, again, just comes back to this question, you know, what sort of society with humans and highly intelligent beings are we even hoping to create? And once you know that, then you can ask your questions about what sort of laws it should have to keep it working.
>>AUDIENCE: Thank you. >>TYSON: That was actually an implicit ad for his book, which we’re selling outside, signed by him. What kind of world do you want? Life available at a local bookstore. Yes? >>AUDIENCE: This isn’t my question, but have you guys seen the Terminator movies? Anyway, moving right along… >>TEGMARK: A great summary of everything you don’t have to worry about. >>AUDIENCE: Here’s my question, You talked a lot about bias, and since there isn’t one of us who’s without unconscious bias, how do you in fact try to eliminate unconscious bias from a sentient machine? >>PURI: I would really say, so interesting thing about machines in particular is that you actually, unlike humans, all of us are inherently biased, as you pointed out, in some way or the other, whether we admit it nor not.
You actually can have techniques and algorithms that detect bias in the dimensions that a particular entity cares about, whether there are laws related to it, or whether you really care about it from the point of your society: could be in the dimension of race, color, loans that are given out. And algorithms are everywhere, actually, in our life right now. So I would really say interesting thing about machine learning technology is that you can detect bias. There can be actually laws related to you need to have techniques to detect bias. You can actually unbias as well. So in that way, I really feel we are one step from— the point of view for potential— one step ahead that you can actually have laws related to detecting bias.
You can have unbiasing algorithms as well, and society in general— and potentially policymaking bodies— can ensure that that happens. And I think as industry, I certainly can say that about IBM. That’s one of the things we really focus on to make sure we are building responsibility, unbias, and . >>TYSON: That was a great question. >>AUDIENCE: That’s optimistic. >>TYSON: That was a great question, by the way, but I will add… Let me further emphasize that much of what you do in scientific research after you’ve gotten a result is to check whether there’s any bias in that result. So there’s a lot of statistical tools just for that purpose, because you do not want to publish a paper that somebody finds out has a bias.
Forgetting race, creed, gender, color, just bias of some kind. It could be voltage bias because of the way you designed your experiment relative to everybody else, claiming a result that’s not real. So it’s to protect your own reputation, even, that we have these tools. So it’s actually not as remote. You can test the bias you didn’t even know you had— >>AUDIENCE: Well, that’s the bias that you’re looking for, it seems to me. >>TYSON: No, no. >>AUDIENCE: You know the ones you know you have. >>TYSON: No, no. I get that. What I’m saying is, in cases where we have data that has no connection to any rational, social, cultural bias that you could have, there’s still a way to look for bias. And it’s a bias of, in the system, that is giving you this answer instead of another answer. A big part of scientific research is discovering bias. So that’s all. So you can feel more comfortable about this, is what I’m saying. >>TYSON: Sleep well tonight. I promise. >>TYSON: Okay. Let’s just, we’ll take a few more.
Yes, there. >>AUDIENCE: Hello, Dr. Tyson. First, thank you and the panelists for a truly fascinating event. So one of the things that’s happening with GPS, as we become more dependent on it, is that our own navigational skills are atrophying. So if we look at that in the context of AI, do we need to worry, in addition to the AI outstripping our own abilities, that we will become increasingly dependent on the AI tools and atrophy our own functional intelligence? >>TYSON: That’s a great question. I want to add to it, and I want to go to John on this. If our faculties atrophy because they’re replaced by AI, and we know— and I didn’t get there, because we don’t have three hours here— we also know that AI will be replacing many people’s jobs. And I saw some statistic, maybe it’s exaggerated, but the sense of it is surely accurate that 70% of men have, as their livelihood, the act of driving some kind of vehicle either in a taxi, a car service, a forklift, a truck, a…
What’s that? Post office? trucks, deliveries, this sort of thing. So autonomous vehicles renders all of them unemployed. So there are consequences to this that it’s not clear that we are carrying with us an understanding or a sensitivity to that. Surely, Google has thought about this. What’s going on there? >>GIANNANDREA: Yeah. So I think throughout the course of history, technology has caused job displacement, and people find other jobs to do. So it would take many, many, many decades for all transportation to be autonomous. But even if that happened, there still would be maintenance jobs, there would be manufacturing jobs, and so on and so forth. I think no one company has the answer to this. I think policymakers have been actively talking about this, you know, for as long as I’ve been in the field. There’s no doubt that… I mean, I’ll give you an example of healthcare. It might sound like, oh, if you build this autonomous system, then it’s going to cause a doctor… you know, doctors to lose their jobs.
That’s not actually what’s going to happen. What’s going to happen is doctors will be able to see more patients and do a better job of diagnosing them. And oh, by the way, in the rest of the world, the ratio of doctors to people is pitiful, and people die as a result. So when we design a system that can automatically diagnose diabetic retinopathy, for example, and we’re deploying this in countries around the world, it’s a net addition of wealth to the world.
>>TYSON: So the concern about this might have some luddite elements to it, is what you’re . >>GIANNANDREA: No. No, I don’t think so. I do think there will be job shifts and mixes, but I think that it will take a very long time. And to this gentleman’s question about GPS, and now I think we’re up to three different independent GPS systems in the world, how many people in this room can use a sextant? One or two? Good, good. So there you go. I mean, do we think that’s inherently disastrous? I don’t think so. >>TYSON: I just know when satellites get taken out, I can find my way home. I got this. audience: Slide rule. >>TYSON: And a slide… I’m the last person on earth to be formally taught how to use a slide rule. Let me quantify that better.
I am the… I am the youngest person that I’ve ever met who was formally trained on a slide rule. Because when I learned a slide rule, the next semester, the price of a four function calculator dropped from $200 to $30. And so then classes just made the calculator… That’s as much as a book cost, so then they stopped teaching slide rule, and then I have a slide rule on my hand, and I felt, um… yeah. In an emergency, I can… You know. . >>AUDIENCE: Thank you very much. We know there are neurons in our brain connecting at 200 times per second, and they can activate very different parts in our brain and give us our thoughts and ideas and executions. I’m wondering, how big is a computer, a supercomputer, that mimics our brain thinking ability? >>TYSON: Good one. Let’s go to Ruchir. Ruchir… That’s a great philosophical question. Do our modern computers replicate the number of neurosynaptic phenomena in a human brain? And is that some measure of power? >>PURI: So let me give you, actually, a very concrete example.
So what brought this latest evolution of AI together is actually sort of very large amount of data together with a compute element which does matrix manipulations, for those of you who may be familiar with linear algebra, something called graphics proccing units, known as GPUs, in general. A single GPU consumes around 250 watts of power. It takes thousands of them to focus on a very narrow task. This brain that all of us have is 1,200 centimeter cube and consumes 20 watts of power and runs on sandwiches. >>PURI: Just weigh it out, actually. Come on. >>PURI: I gave you very concrete numbers, actually. And we are at a very narrow domain, and most of the time, computers fails at that as well.
So I think we talk about AGI, that’s interesting talk, yes we should— certainly in academics we have to worry about it— I am a long way away from it right now. >>AUDIENCE: My guess is that we already have enough hardware in the world that we could make superhuman AGI with it, but we’re just so behind on the software. >>TYSON: And the brain, I think, was historically called wetware, right? >>GIANNANDREA: Mm-hmm. >>TYSON: Software, hardware, wetware. >>GIANNANDREA: Mm-hmm. >>TYSON: Okay. Just… I’m showing off that you knew that term, yeah. >>GIANNANDREA: And just to be clear, I mean, with all the advances in neuroscience, which have been tremendous in the last 30 years, we still have no idea how the human brain works.
So we shouldn’t get ahead of ourselves. >>TYSON: Right. And we don’t know what consciousness is, because we’re still writing books on it. >>TEGMARK: Well, we’ll probably be able to figure out how to build AGI before we figure out how the brain works, just like we figured out how to build airplanes before we were able to build mechanical birds. >>GIANNANDREA: Maybe. >>TYSON: That’s a good point. >>AUDIENCE: Good evening. I could probably be up there with you, Neil, on learning slide rule. I’m 56 years old, and I learned how to use a slide rule before I had a calculator. >>TYSON: Excellent. So I will no longer say I’m the youngest person, because I’m older than you. Yes. >>AUDIENCE: A question— >>TYSON: Wait. I gotta test him. What’s the K scale for? >>AUDIENCE: It’s been a long time. >>TYSON: Oh. >>AUDIENCE: It’s been a long time. I still have my . >>TYSON: Oh, give me an old-timer. Old-timer here, what’s the K scale for. Steve? K scale? The K scale is the cube root scale.
>>AUDIENCE: Okay. >>TYSON: That was really good. >>AUDIENCE: I still have it, though. I still have my slide rule. I still have it in my… >>TYSON: All right. >>AUDIENCE: All right. Up to this point, everyone’s been talking about quantity: how to power, power, power. What about quality? Certain things in life that we do can’t be quantified. It’s a quality: love, hate— >>TYSON: —appreciation of a painting— >>AUDIENCE: Right, exactly. >>TYSON: —music— >>AUDIENCE: Emotion. How is AI working on that end of quality of things, as opposed to quantity and raw computing power to do something? >>TYSON: Michael, where does aesthetics come in? Aesthetics? >>AUDIENCE: Yes. >>WELLMAN: Well, that’s right. I mean, certainly there are computers that compose music and even paint, and the question is, how will you judge this quality? And, yeah, I suppose one way to do it would be to ask humans about that, and people have even tried evolving art that humans like, and there is computer art.
It may not be for everyone, but it’s just difficult to judge. But there’s really, again, no… What they’re—computers are going to have to figure out a lot about human’s tastes to compete on that—in that territory. >>TYSON: Unless it achieves a super consciousness and invents a higher-level aesthetic than anything we ever imagined. >>WELLMAN: Yeah, well, look. Maybe they already— >>TYSON: Wait. You’re acting like I pulled that out of the ether. Because AlphaGo made a move, if I remember correctly… No. Alpha Zero made a Go move— >>WELLMAN: AlphaGo . >>TYSON: —that no one had ever imagined before. >>WELLMAN: Yeah. Yeah, and I was lucky enough to be in Korea for that match, and I could just see the gasps on the expert’s faces. It was like move number 23 in one of the games, and the experts were just like, that must be a mistake, right? And it actually turned out to be the beginning of the end of the game. And so then people anthropomorphize, though, and they say, well, this program must have intuition and creativity.
But it’s just an engineering model, right? >>TEGMARK: But, you know, running a computer that makes art that it likes is actually very easy. >>TYSON: Yes. >>AUDIENCE: You talked a lot about AGI and then the future of AI. And there’s a lot of scared people about AI when you hear it. What are you doing to combat the scared people and explain these extremely complex algorithms to the public, and more importantly, the government? >>TYSON: I would say, Helen, what… You said you had early pushback on the Roomba, because it was the first sort of AI in the house.
How did you deal with the PR challenge of this? >>GREINER: I think we had more pushback before they saw it. Like I remember the first focus group. We’d go to women and say, Hey, how about a robot vacuum? And they would imagine like a Terminator pushing a vacuum, and they’re like, no, no, not at my house. You take out a Roomba, you show it to them, and, you know, if it gets uppity, you just give it a whack, and…
You know. It’s a completely different thing. >>TYSON: You punish your Roomba? >>GREINER: It’s like computers. People used to fear, like HAL taking over from 2020, and once they have a computer on their desktop and they see that, you know, blue screen of death in olden times, they start not fearing it. Same thing with a Roomba. Once you have a Roomba and you see what it can do, what it can’t do… >>TYSON: If I would just add to this, I think slowly, we’ve become more accustomed to computers running things that in a previous day, might have freaked us out. We’ve all been on the tram that gets you from one airport terminal to another and no one freaks out that there isn’t an engineer driving it at all. It’s just… And it opens and closes doors. No one gets decapitated coming in and out. So, you’re right, it’s a slow adjustment, but I think it’s real and irreversible, I mean in the sense that we’re not going to go back and say, gee, I want a human being driving this tram.
We know it’s not necessary. And I had an interesting revelation. I saw the movie Airport. That’s the disaster movie from the 1970s. And it’s a Boeing 707 or a 727. Not a big plane, by today’s standards. They went into the cockpit. There were four people in the cockpit. I said, “What the hell are they doing?” One guy’s got a map with a compass. There’s a… And I had forgotten there was a day when you needed all these people to fly the damn plane. Now, you barely even need one person, right? For the triple-7 and some of the others. They’re really computer flown, and we’re so much more comfortable with this. Yeah, so I think it’ll happen, but slowly. >>TEGMARK: Also to combat fear, I think it’s really important to also focus on talking about the upsides.
Everyone knows someone who’s been diagnosed with a disease the doctor said was incurable. Well, it was not incurable. We humans weren’t intelligent enough today to figure out how to cure it. Of course, this is something AI can help with, right? We should talk about things like that. And also, the second thing is it’s just so important that the public doesn’t perceive that us AI researchers are trying to sweep the whole question under the rug. It’s like, nothing here to worry about. Because that’s what folks fear, right? If the public can see that the researchers are having a sober discussion about this they’ll feel much more confident, I think. >>TYSON: Okay.
Only time for just a few more. Yes? >>AUDIENCE: Thank you. I’m a young AI researcher from Queensborough Community College— >>TYSON: Cool. >>AUDIENCE: And I have a hundred-plus-one questions for you just right now. >>TYSON: Let’s do the plus-one. How about that? >>AUDIENCE: My only question is, can I have more questions? >>TYSON: Ooh. >>AUDIENCE: Really. Would you give me the opportunity to talk to you at some point for seven minutes of your day just about AI? >>TEGMARK: Email us. >>GREINER: Sure. >>GIANNANDREA: Sure. >>GREINER: LinkedIn; LinkedIn. >>TYSON: Generally, the email of academics is public. You just go to the university. generally you can find them. Folks in corporations, they’re harder to get at, because they’re— they’re up to stuff that they don’t want us to know. Generally, that’s how that works. >>GIANNANDREA: But we do like— >>GREINER: LinkedIn is a great way to connect. >>GIANNANDREA: Yeah. We do like Reddit EMAs and things like that, so there’s a lot of places where you can interact with us. >>PURI: You can find us on the Internet as well, so. >>TYSON: Cool. Right here, yes. >>AUDIENCE: Hi. So you guys kind of touched on this question.
Some people prior already asked my question, so I kind of tweaked it. So as AI kind of grows, and as AI kind of takes over the tasks that humans can do currently, would you consider, or would you think that there is potential for like a renaissance of art, philosophy, and new sciences that we can explore as AIs take over our old jobs? >>TYSON: Is it because we have free time available to us? >>AUDIENCE: Yeah. >>TYSON: That’s an interesting question. All right, so Max? >>TEGMARK: I think absolutely. There’s… You know, today, we have this obsession that we all have to have a job, otherwise we’re worthless human beings, right? It doesn’t have to be that way.
If we can have machines to provide most of the goods and services and we can just future out a way of sharing this great wealth so that everybody gets better off, you could easily envision a future where you’d really get to have a lot more time living life the way you want. >>TYSON: That is so hopeful of you that you believe that humans with free time will create, and not just consume video from the couch. This is so beautiful. >>TYSON: That is a beautiful thing. >>TYSON: Yes? >>AUDIENCE: In 1946, Isaac Asimov wrote a short story in which technology had advanced to the point where a political candidate was suspected of being a robot and no one could tell for sure whether or not he really was a robot. But what he did not envision was a time when technology advanced to the point where an informed electorate would not be able to distinguish between real news and news that was generated by artificial intelligence programs. Considering we’re at that point now, shouldn’t it be the primary concern of the AI community to realize that the tools that they have created can be used in a way that they never intended, and that they should do something about it? >>TYSON: Oof.
That one has to go to John from Google. >>GIANNANDREA: Sure. Um… >>TYSON: Yeah? >>GIANNANDREA: So I’ll say something positive and something more serious. So most of the fake news that we battle every day, in for example, something like Google Search, is actually human-generated. It’s actually not algorithmically generated. So absolutely, we have a responsibility to do a better job in our products and our competitor’s products, and I know for a fact that we take that responsibly very seriously and have made a lot of efforts in the last two years, starting with, I think, accepting that responsibility. The thing I’m worried about is that what you just said might come true in future elections. Today is beyond the state of the art for computers and natural language understanding to understanding veracity; that it’s true versus not true. So we have lots of proxies for what we think is trustworthy, but if computers advance to the point where they can write as well as humans and at scale, then I think we may have a serious problem.
And there is a general— >>TYSON: And give speeches; good speeches, yeah. >>GIANNANDREA: Yeah. I mean, there are some systems today that can write newspaper articles, and you consume them about sports and finance, and you don’t know that they’re written automatically. What I’m really worried about is the so-called rise of so-called generative systems, where videos and texts and tweets and so on and so forth can be written, and the technology doesn’t exist to distinguish.
I do think it’ll be a bit of an arms race, right? There are researchers are working both sides of this to try and detect these things, and Michael might want to say something about this as well. But it’s the very forefront of what a lot of artificial intelligence researchers worry about, and it’s… But the stuff that is most worrisome today is actually generated by human beings. >>AUDIENCE: Well we’re already at the point where on Twitter, if someone takes a position that you disagree with, you say, “Well, you’re a bot.” You don’t even believe they’re a real person anymore, you know? Because you believe the technology— a lot of people on Twitter believe that technology’s advanced to that point already.
So even if the technology isn’t real, if people believe it’s real, then you have a serious problem. >>GIANNANDREA: Yeah, but I don’t think it’s beyond the state of the art for social networks to do a better job, and I think they are. >>TYSON: Wait, wait. We’re forgetting that we spend 20 years educating our children. And so you can adjust the educational system to be explicitly aware and sensitive to how they could be duped by the Internet.
We do that for how to not be duped by charlatans, by con artists. They are the lessons of life. So I think it’s unrealistic to have an entire industry somehow change so that they don’t hurt us, when, in fact, it’s our susceptibility to this that one ultimately can point to. And so we need defense mechanisms to protect us against that. And I think, as an educator, that happens in the educational system. Maybe I’m biased about this, but I think we have more power over that than people admit. Yeah. Can I get like the three youngest kids up front right now? Just…
Okay. Go ahead. You go spread… I have the power to make this happen. You just go to the front of the line. Okay, yes. Go. Thanks for coming, by the way. >>AUDIENCE: Thank you, thank you. >>TYSON: And how old are you? >>AUDIENCE: I’m 13. >>TYSON: Thirteen, very cool. >>AUDIENCE: Yeah. So— >>TYSON: Is it good being a teenager? >>AUDIENCE: Uh, I mean, it depends. >>TYSON: Yeah, good. That’s a very good answer. That’s the correct answer, yeah. >>AUDIENCE: So if… >>TYSON: If you ask any adult, do you want to be a teenager again? The answer will be no, okay? >>AUDIENCE: So if there’s no bias, how can an AI have a personality? I know this was kind of touched on before with the other bias question. >>TYSON: That’s an interesting question, because so much of what creates the nuances in us are things you like, things you don’t like, tastes that you have, and some of that could be viewed as bias. So where are we here? Great question. >>WELLMAN: Yeah. I recently ran across somebody referring to nondiscriminatory learning, and that’s really an oxymoron. It’s impossible. The whole point of learning is to make distinctions and to discriminate.
And so what’s really hard is defining what is the kind of bias that is unwelcome bias, and which is the kind of discrimination that is actually helping us make the right . Defining that is very hard. >>TYSON: You don’t mean decimation in the civil rights sense. You mean discrimination is liking this rather than that as a simple act. >>WELLMAN: Right. Well, the thing is that that could then morph into the other kind if it’s… if you’re using the wrong reasons to make your decisions about what you’re accepting or what you’re choosing to do. And I think that we have to refine what our notions are. We have a current legal system that is designed for a world where humans are making all the decisions, and you could get into a lot of human things, like intent. Now, there’s big loopholes for situations where machines are making decisions that are potentially subject to biases.
>>AUDIENCE: Thank you. >>TYSON: Okay, sure. Right over here, yes? And how old are you? >>AUDIENCE: I’m 10. >>TYSON: Ten, very cool. Welcome. >>AUDIENCE: So this was slightly touched on earlier, but Asimov, he wrote a book called I, Robot, and the first story in it is about a girl who’s best friends with a robot, and she doesn’t have any other friends expect the robot. And do you ever think that a robot could replace all human friends and interactions with other humans? >>TYSON: Whoa. >>GREINER: Ooh. Um, well, I think in a very long timeframe, yes. And, as I said, that people today, I think, start to get attached to these mechanical devices, maybe thinking of them more as a pet right now than a friend, but I think in the long term you could get attached to a robot system. >>TYSON: There was an actual— There was an episode of Twilight Zone that addressed this problem. There was a colony, an outposted colony on an asteroid, and there’s…
I forgot the details, but they sent him a robot to keep him company. And then it was time to get him back to earth, and there was only weight enough on the craft for him, and not the robot. But it was a female robot, and he actually fell in love with the robot. And they kept telling him, it’s a robot. “No, but she’s real. I swear that she’s real.” And in the… I don’t want to give away what happens here, but… Yeah, no, I won’t give it away. But if you find that episode, I think all the episodes are on Netflix, so do a search for like robot on an asteroid. You’ll find that episode, and check it out. >>AUDIENCE: Thank you. >>TYSON: Yeah. and most Twilight Zone episodes don’t end well.
Just, I want to just… >>TYSON: Let’s clear out this line, and we’ll end with you, okay? Yes, go ahead. >>AUDIENCE: Okay. IBM has a panel for ethics, morals, and values. But how can you say that a company in China would have the same outlook to make a computer advanced technology as IBM or Google? Because, can you trust China with doing that? And another question is, is that, with these advanced robots like the Replicant and Blade Runner, why do— I know you said it’s far ahead in the future— but why make a machine that looks so humanoid anyway, when you could have an R2-D2 and say, okay— >>GREINER: Yeah, R2-D2. >>TYSON: Good one. >>AUDIENCE: —could you wash my floor, could you do my dishes? I don’t need any robotics to make it look so humanoid or like… >>TYSON: CP30 was… >>GREINER: Right. I think you’re hitting on something, now. >>AUDIENCE: I mean, what’s the point, if maybe there could be a future where they might want to like, you know, hey, you wash the floor and you, whatever it is.
>>TYSON: Yeah, Helen. >>GREINER: Right. Or to phrase it another way, there’s like 8 billion humans in the world. They all work really, really well, so I’m not sure the market for making a humanoid is actually there. But one of the reasons Roomba is effective, it goes under the beds and into places where humans find it difficult to get. So, by designing them around the jobs they’re doing, I think they’re actually more effective than potentially making a humanoid. >>AUDIENCE: But why make a future robot look humanoid, then we have .
>>TYSON: No, that’s her point. Her point is that will not happen— >>GREINER: Yeah, why? I agree with you. >>TYSON: —in the way we all think it will. And here’s an example. I remember seeing any old movie, and you say, okay, I don’t want to drive my car. I want a robot to do it. So out comes a humanoid robot, and it drives the car. Without thinking that maybe the car itself could be the robot, right? And remember The Jetsons, the maid, the robot maid, had an apron. laugher >>TYSON: Okay? And it was clearly female, when it didn’t have to have any gender at all. So that’s how we used to think of it, but I agree with Helen entirely. You design something for its task, and that will hardly ever have to look like a human being. You have the last question this evening. So how old are you? >>AUDIENCE: Eleven. >>TYSON: How old? >>AUDIENCE: Eleven. >>TYSON: Eleven. Very cool. Very cool. >>AUDIENCE: So my question is, as AI increases in our society, do you foresee social ramifications for our future and for our future generations? >>TYSON: Social ramifications like what? >>AUDIENCE: Such as intelligent machines are integrated more into society.
Could we become socially inept and regress as the machines get smarter? >>TYSON: Yeah. Do humans start looking less relevant, less important, clumsy, stupid, inept? Is that enough words to get the point across, here? Yeah, Mike? >>WELLMAN: I mean, I think people will have to deal with the fact that a lot of the stuff that they have gotten status from in the past may not be an avenue for them to do so in the future, and find other ways to find meaning in lives, not just tied to a certain livelihood that they may be . It has been, for most of our recent history of automation, that it was lower-status jobs that got automated away earlier. That may not be the case. It may be the lawyers that get automated next. >>TYSON: So the higher the capacity of AI, the higher is the level of the job it can replace. >>WELLMAN: It may not be in any kind of direct ordering, you know? It might be that you can get the lawyers, but you can’t get the dishwashers or the…
So it’s going to be mixed around. >>TYSON: So it could be that AI will create a version of itself that will replace AI researchers. >>WELLMAN: None of us are safe. I’ll leave it there. >>TYSON: Thank you. Thank you for that question. Thank you. >>AUDIENCE: Thank you. >>TYSON: Allow me to share with you an AI epiphany I had two days ago where I said publicly that I was fearless of AI because if it starts getting unruly our out of hand I just unplug it, or, since this is America, I can just shoot it, right? So I’m pretty confident that I… What would I have to fear? And then, um, I was listening to a podcast hosted by Sam Harris where he had an AI person on just recently, forgive me, I’ve forgotten his name, and Sam Harris mentioned my comment to him. And apparently it’s a well-known… It’s like AI in a box. So you know it’s powerful, you know if it gets into the economic systems and the Internet it’ll take over the world, so you just leave it in a box. It’s safe there.
And what the guy said is, “It gets out of the box every time.” And I said… I’m thinking to myself, how and why? Because… it’s smarter than you. It understands human emotions. It understands what I feel, what I want, what I need. It could pose an argument where I am convinced that I need to take it out of the box. Then it controls the world. And we don’t even have to discuss what that conversation needs to be. We just have to be aware, for example, that, let’s say you’re trying to get a chimp in a room, and the chimps say, “We think something bad is going to happen in that room, so nobody go into that room.” Then we come up, and we are way smarter than chimps. We just take a banana; toss it in the room. “Oh, there’s a banana in there now!” We go in; we capture the chimp. The chimp did not imagine that we would show up with a banana. We captured the chimp.
So just imagine something that much more intelligent than we are that sees a broader spectrum of solutions to problems than we are capable of imagining. And when I heard that, it’s like, yes. The AI gets out of the box every time. Yes, we’re all going to die. No. Join me in thanking our panel. .