Making the Most of A.I.’s Potential
Advancements in artificial intelligence continue to spread through our world. This past spring, a robot surgeon beat a human at stitching up a pig’s intestines, and self-driving cars are now answering some Uber calls in Pittsburgh. Computers’ ability to see and make sense of the world has made great strides in recent years, while improvements in machine learning could bring greater autonomy and utility to the A.I. we already use.
But where could this field take our society? Will the warnings of Elon Musk and others, suggesting that superintelligent A.I. could escape our control, bear fruit? How do we ensure that developments in A.I. are as beneficial as possible?
The new Center for Human Compatible Artificial Intelligence at the University of California-Berkeley proposes to focus on ensuring that A.I. is as compatible as possible with our needs, and that robots are kept under our control even as they become more advanced. Berkeley computer science professor Stuart Russell is leading the center, and says that teaching computers our values could be the key to humans and computers working toward the same goals.
Meanwhile, at Stanford, a 100-year study on the social and technological challenges and potentials for A.I. is underway. The AI100 project’s first report, released this month, looks at the potential that properly-channeled A.I. could have for the year 2030 in sectors ranging from transportation, to education, to neighborhoods. Study panel chair Peter Stone, a computer science professor at the University of Texas-Austin, explains these findings.
Dr. Stuart Russell is a professor of Computer Science and the Director of the Kavli Center for Ethics, Science, and the Public at the University of California, Berkeley in Berkeley, California.
Peter Stone is a professor of Computer Science at the University of Texas-Austin.
IRA FLATOW: This is Science Friday. I’m Ira Flatow. Now, for a topic we’ve been talking about since the dawn of Science Friday 25 years ago– artificial intelligence. The field is on the move this year with some striking examples of what the future could hold. We have a robot surgeon beating a human at stitching a pig’s intestines. Self-driving cars are now answering Uber calls in Pittsburgh. And as robots get smarter, they learn more about how we work, start doing more things for themselves.
We’ve heard warnings from futurists like Elon Musk and Stephen Hawking that super intelligence could escape our control. How do we ensure that AI developments are as beneficial as possible? Well, a new center launched by University of California at Berkeley proposes to focus on ensuring AI is compatible with our needs, while keeping robots under our control even as they get more advanced.
Berkeley computer science professor Stuart Russell is leading the center. Welcome back, Dr. Russell.
STUART RUSSELL: Thank you for having me.
IRA FLATOW: And based at Stanford, another group is studying the achievements and potential of artificial intelligence for the next 100 years. It’s the AI100 project’s first report, which is now out. And it looks at the potential properly channeled AI could have for the year 2030. The lead author of that report, Dr. Peter Stone, is a computer science professor at the University of Texas at Austin. Welcome, Dr. Stone.
PETER STONE: Thank you.
IRA FLATOW: And if you’d like to talk about how do we make most of AI? How do we take advantage of it, AI’s potential? Give us a tweet. Our tweet is @SciFri, again, the tweet is there. How do we make the most of AI’s potential? Tweet us @SciFri. Dr. Russell, how do you define human compatibility AI?
STUART RUSSELL: So what we really want to do is make sure that AI systems are provably beneficial, specifically to humans. And so I need to unpack that a little bit, if you don’t mind.
IRA FLATOW: No, go right ahead.
STUART RUSSELL: So first of all, I think we all understand that AI is about making machines intelligent. And that’s basically a good thing. Dumb machines are not all that useful. And we already see many of these benefits that you mentioned.
And as we approach human level AI, the ability of machines to solve a wide range of problems and learn all kinds of new subjects, then I think that will really represent a dramatic change in the progress of history. Because history is really the result of our minds operating in the world. And if we have access to substantially greater intelligence that we can use as a tool, then that’s going to change the direction of history– hopefully for the better.
IRA FLATOW: So but why do we need a center for human compatibility? Can’t we just say that AI is already human compatible, does stuff for us?
STUART RUSSELL: So, yes– so far, we’ve focused on a very restricted tasks like playing chess. But if you’re Garry Kasparov, was it a good thing that Deep Blue beat you? From Garry’s point of view, probably not. And maybe that illustrates the issue.
IRA FLATOW: But they’re good at sweeping the floors too. We got those robots that sweep floors, what have you.
STUART RUSSELL: And they’re good at driving us around, until they crash into the side of the truck and kill the driver. So the real issue, I think, if you ask an AI researcher, what do you mean by AI? Not just the lay version, making machines intelligent, but the technical version– it means making machines that can optimally achieve objectives.
And then you say, well, what objectives? And the AI researcher will say, well, that’s not my job. That’s your job to figure out what your objective is and put it into the machine. And as long as we’re putting in small, simple objectives, like win the game of chess or drive to the airport, there isn’t too much of an issue. But if you put in objectives like, make my company as profitable as possible, or end hunger for humans.
Well, there are different ways of doing that. We can end hunger by ending humans. If you ask King Midas, he’ll tell you, no, it’s very easy for humans to misspecify the objective. He said, I want everything I touch to turn to gold. And he got his food and his drink turned to gold. And he was very unhappy, and he died miserable.
So the real issue is that AI is only human compatible if the objectives are consistent with what humans really want. It’s too late when we put in an objective like end human hunger to find out that the machine is basically exterminating humans so that they won’t be hungry anymore.
IRA FLATOW: I hate it when that happens.
STUART RUSSELL: Yeah, it’s a real bummer. So what we’re trying to do is change the definition of AI so that getting the right objectives actually becomes part of the job of the field. And it isn’t just left to the user to specify.
IRA FLATOW: Peter, maybe you can help us hone in on some of the objectives that you see. What are we looking at as you look past and future about AI and human compatibility, what are some of the biggest areas of opportunity right now?
STUART RUSSELL: Yeah, that’s a great question. And that was one of the main focus areas for the AI100 report that we came out with. We were given the charge to think ahead to the year 2030 and come up with the likely influences of artificial intelligence technologies on a typical North American city by the year 2030.
And before I go into the areas that we focused on, I do think it’s important to say that, in contrast to what Professor Russell was just saying, one of the leaping off points for the report was to observe that at least in the present day, there is no sort of one artificial intelligence. AI is not one thing, rather it’s a collection of technologies that require special focus and research towards every application.
So there doesn’t exist right now, and we don’t anticipate in the foreseeable future, at least in the near future, a machine that could be given any goal. But rather there are technologies that are focused on particular areas. And so the areas we looked at in our reports were, starting by the one that’s probably front and center in everyone’s attention right now, is transportation.
It’s clear that there’s progress in autonomous vehicles and in other areas of transportation where AI technologies have a huge potential to make an impact in the relatively near term. But there’s also great opportunities in health care, home and service robots, education, public safety and security, employment in the workplace, and entertainment. And in each of these, there’s great potential for AI technology to improve our lives, as Professor Russell was saying.
But there’s also barriers, and both technical challenges in each of these domains that are unique to each domain. And also social and political barriers that need to be overcome, if they’re going to achieve their full potential.
IRA FLATOW: Is that why we should bring social scientists in to this conversation about deciding where AI should be going?
PETER STONE: Yeah, I think this is a conversation that needs to happen among technologists, but also policymakers, and politicians, social scientists, economists. And on our panel, we had people there representing all of those areas. Because I think, for example, there is a lot of concern– I think legitimate concern– that AI technologies will affect the employment landscape and the job landscape.
And probably not by replacing lots of jobs in the near term, but rather changing the nature of certain jobs. For instance, physicians will certainly not in the near term be replaced by AI technologies. But there will be tools, and new tools and skills, to build that they’ll have at their fingertips and be able to use.
But they’ll need to be able to and we’ll need to be able to make sure that there is enough transparency that the physicians and the patients are able to trust the recommendations that are made by the new technology. And when it comes to the economy, there’s a lot of potential for greater productivity as a whole by society. But there is also danger that there could be even greater concentration of wealth in a smaller number of people’s hands. And if we need to start the conversation now about how to make sure that the wealth that’s created by AI technologies is treated fairly and equitably.
IRA FLATOW: In fact, there are some tweets coming in about that. Mohammad Athar writes, “most programmers are still wealthy white men. How do we make AI better for all people?” Stuart, you got any suggestion, maybe curricula change?
STUART RUSSELL: Well, I think there is a very big issue with the demographics of Silicon Valley. I wouldn’t say white. I would say white and Asian. But I think the technology is spreading throughout the world, and different populations are going to be adopting it and becoming expert in it.
But I do see a lot of people saying, well, it’s great that robots are going to do all the work, because then we can have a guaranteed basic income. And anyone who doesn’t want to work doesn’t have to work. To me, this is just a recipe for disaster, because what you end up with is a society of a thin layer of extremely rich owners of technology, a layer of people who serve their personal needs, and then the rest of the population is fed and housed and detained, and pretty much left to their own devices and doesn’t have a real function in society.
That’s not a vision of the future that I, for one, would welcome at all. So we really have to think very hard about what the future shape of an economy is going to be when the material needs of the population can largely be met by fully automated systems.
IRA FLATOW: What’s the best way of teaching a robot?
STUART RUSSELL: There are a lot of different ways. Peter, for example, works in an area called reinforcement learning, which means that essentially you give the robot a virtual lump of sugar when it succeeds, or at least partially succeeds, in a task. And you wrap it on the knuckles when it doesn’t do it right. This is reward and punishment.
And that process can be very effective in getting a robot to learn various tasks, assuming that the human knows when to supply the sugar and when to supply the punishment.
IRA FLATOW: Peter, you agree with that.
PETER STONE: Yes, absolutely.
IRA FLATOW: What about having robots watch what we do?
PETER STONE: Yeah, so there’s many different ways of communicating the teaching signal. So one is by just a reward signal that Stuart was talking about. There is also an area of research called learning from demonstration, where the robot watches examples of successful behaviors. And this is an active research area as well.
There’s a channel of reward signal. There’s a channel of demonstrations. There can also be a device. You’re watching a robot or a program do a task, and provide it more of a richer feedback than just good robot or bad robot. But more next time you’re in that situation, try doing x instead of y.
IRA FLATOW: I’m Ira Flatow. This is Science Friday from PRI, Public Radio International. Here broadcasting from KQED in San Francisco with us, Stuart Russell of the UC Berkeley and Peter Stone of University of Texas at Austin. Do we want to teach robots morality? And whose morality do we teach?
STUART RUSSELL: Yes, and in some sense, everyone’s. So one of the objectives of our center is actually to get robots to have the right objectives so they don’t go off and do things that make us unhappy after the fact. And there’s a process which is related to the idea of learning from demonstrations that Peter just mentioned. It’s called inverse reinforcement learning.
It’s sort of the opposite of the reward and punishment idea. And so what that means is you observe a human’s behavior, and you try to figure out what objectives the human is trying to achieve. So for example, if I get up in the morning, and I struggle downstairs with my eyes half-closed, and I press buttons on a funny looking machine. And it makes grinding and hissing and steaming noises, and then this brown liquid comes out. And then I feel happy.
The robot should be able to figure out that my objective in all this is to get a cup of coffee. And then perhaps the next morning, it will make the coffee for me and bring it to me in bed. So that’s a very simple example.
We would like this process to extend so that eventually machines understand the full spectrum of human objectives. So that, for example, when we ask it to end hunger, it doesn’t go off and exterminate everyone so they stop being hungry.
IRA FLATOW: But as you were saying before, we don’t expect the machine that makes us coffee to be the same machine that’s going to take on hunger for the whole world.
STUART RUSSELL: So I think it’s a spectrum. One of the kinds of systems that people are very interested in building right now– within the next five years, I think we’ll see lots of products, and we’re already seeing partial products, is the digital personal assistant. So that’s an AI system that looks after your day to day life, your calendar, books, your travel. Some executives have very well-trained, inexpensive assistants who do all this stuff for them. But in fact, we could make this available to pretty much every human being on the planet for $0.99 or $0.29 a month or whatever.
So imagine what that system is going to do. So you want to go to a meeting. And you say, OK, just book me into the closest hotel. So does it book you into the closest hotel where the only room available is the $2,000 a night presidential suite or does it put you in the Hampton Inn down the road for $115? Well, it depends on your preferences. I
IRA FLATOW: Not that there’s anything wrong with that.
STUART RUSSELL: So you might have one preference or the other. And out of the box, the system doesn’t know much about you and your preferences. So one of the main things these systems will have to do is to learn the preferences of the individual in order to make good decisions on their behalf. And also to make decisions that don’t negatively affect other people.
So you might say, I am very jealous of my calendar. I don’t like too many people making appointments. But then, I’m a professor. That means there’s a long line of students who can’t get their form signed and can’t get into classes and so on. So you have to take into account the preferences of other people.
It starts to get very complicated. And those are the kinds of problems that we will see in the near term, let alone the problems of exterminating the human race by accident.
IRA FLATOW: Peter, the AI100 report warns us against unnecessary fear and suspicion of AI. With futurists like Elon Musk and Stephen Hawking warning us not to trust AI, how do we trust AI? How do we get to trust it?
PETER STONE: Yeah, there are various ways to instill trust. One is transparency. So making sure that people understand how something is programmed, and if it’s trained, or if there’s learning involved, and what data it’s trained on. Going back to the point on morality, and this is one of the reasons that the center at Berkeley that Stuart’s leading is very timely, is that as autonomy as cars come into existence, decisions that never had to be made explicitly are going to have to be made explicitly. And that’s where the morality comes in.
People talk a lot about the trolley problem. If a car has to make a choice between saving a passenger inside the car versus pedestrians outside the car, that’s never been something that we’ve had to explicitly decide about. But with an autonomous car, somebody’s going to need to program that morality. And we need to decide as a society, what are the right decisions? But another way for–
IRA FLATOW: I got to break– my programming inside of me says I got to say goodbye. So I want to thank both of you. It’s a great story. We’ll get back to this. Dr. Stuart Russell, computer science professor at University of California-Berkely, Peter Stone computer science professor, at U Texas at Austin.