{"url": "https://www.youtube.com/watch?v=-Mi7ggw59EY", "page_url": "https://www.youtube.com/watch?v=-Mi7ggw59EY", "title": "How AI is Automating and Augmenting Work", "show_or_channel": "Future of Life Institute", "date": "2017-03-06", "duration_s": 2732, "type": "panel", "speakers": ["Mustafa Suleyman", "Andrew Ng", "Peter Norvig", "Daniela Rus", "Moshe Vardi"], "confidence": "high", "notes": "Beneficial AI 2017 (Asilomar) panel 'How AI is Automating and Augmenting Work', Future of Life Institute.", "people": ["mustafa-suleyman"], "id": "2017-03-06_future-of-life-institute-how-ai-is-automating-and-augmenting-work", "audio_duration_s": 2731.94375, "worker": "jc-w05", "transcribe_s": 37.3, "model": "large-v3", "segments": [{"start": 8.37, "end": 33.65, "text": "Let me ask you a bit more about having everybody learn computer science. I can understand you're the head of the computer science lab at MIT. Do you think that everybody in the country is going to have a job doing computer science? That seems a little ambitious. No, but I think that our dependence on technology is quite great, and it will get greater. So I really believe that digital..."}, {"start": 33.65, "end": 64.45, "text": "Literacy is very important. I also observed that there is no underemployment in IT, and so making it possible for more people to participate in IT will be better for everyone. It will be better for people because they will have better jobs, and it will also be better for companies because at the moment there is a shortage of labor. Ask me that question. Okay, yeah, go ahead, Marsha. Go ahead, tell us. You've got the mic."}, {"start": 64.45, "end": 92.86, "text": "A few months ago, there was an interesting experiment by OECD. I think you were involved in another session with them, weren't you? Yeah, and what OECD did, they take something called survey of adult skills, which you can think of SAT for the workplace, where they're focusing essentially on, mostly on, not the people, there are about 30% of the people who really mostly work with manual jobs are not involved in that. The people have some need to have..."}, {"start": 92.86, "end": 121.86, "text": "And they're looking at literacy skills and numeracy skills and problem-solving skills. These are the kind of generally the skill that one needs to have a job today in the world, in technology-rich workplace. And so first of all, it's important to look at how people are distributed when they take the skills. No, take a look behind you. So this is an example of one test in here. This is to give you a..."}, {"start": 121.86, "end": 151.18, "text": "Physical exercise equipment. And the question is, which piece of equipment listed received the largest number of ineffective rating? Now, the question that the workshop we're looking at is, can we design today software that will perform well on this test? Wait, before you, don't tell us what the workshop said. Maybe we should ask the panelists. Ask the same question to the panelists that was asked at the workshop. Yeah. So the question is, really, can we automate it?"}, {"start": 151.18, "end": 180.62, "text": "Those are the questions for us. Can we look, can you imagine a software that gets this test, SEs? This is the question up there. Which piece of equipment received the largest number of ineffective rating? This physical equipment, different types, rowing machine, gym bench, et cetera, different arm strength, leg strength, et cetera. You get rating, good, very good, average, ineffective. And the question was, which piece of equipment received the largest number of ineffective rating?"}, {"start": 180.62, "end": 210.46, "text": "And I don't think anybody here would have a challenge with just counting ineffective in every row and coming up with the answer. I think it's a gym bench. What is it? Some one of these. No, the dumbbells. But the question is, can you think of writing software to answer this question? This is one. Let's do the second one. This is a bit more challenging, partly because the resolution here is not very good. But I'm simulating the answer to the question we got."}, {"start": 210.46, "end": 240.58, "text": "But what you're seeing is a picture of packaged water bottles. And the question is, how many bottles are here? Now, this is now a bit even more challenging even for people because it's not just simply multiplying the two sides. If you can see, there are different, every alternating row has a slightly different number of bottles. So now think about it. Can we design today a system that will answer these kind of questions? Should we ask our panelists? You guys are AI experts."}, {"start": 240.58, "end": 267.6, "text": "What do you think? The question is today? Andrew, can today? Can we answer this question? It really depends on the breadth of the problem. I mean, so literally just hours ago, we just ran, for any narrow problem, you know, we could do it. If you want to count things in, count a lot of things, a lot of different contexts, you could probably do it. If the task is given an arbitrary image, do reasoning and figure out a whole bunch of stuff about the image and answer the question, then that's probably many years away."}, {"start": 267.6, "end": 294.7, "text": "So the parameter that we kind of define for ourselves, what does it mean to do it, is to take the whole test, not just, of course, one problem, and we give ourselves a budget of $1 million for one year. And so that means that you need to have a group of $1 million. You can think maybe four or five people working for a year, what can they do? As opposed to putting, let's say, for Watson, IBM put..."}, {"start": 294.7, "end": 324.02, "text": "But you don't know in advance that the test is going to have bottles. It could have sheep or cows. It could have all kinds of things. All kinds of things. Yeah, OK. You can all, you're welcome to Google OECD survey of adult skills. And there is a sample test. You can take the test and see how well you're doing. Mustafa, Peter, do you want to give an opinion on that? So I think the important question is, if you've got this group of four people, is this the best thing they should be spending their year on? But there's probably something else that's a more valuable task for them to work on."}, {"start": 324.02, "end": 345.1, "text": "Well, think of it as a metaphor for our ability to build agents that perform well in the workplace. This is, of course, nobody's interested in just doing this. I think for the next decade or so, we're not going to be addressing these one-off tasks. We're going to be solving bigger problems that..."}, {"start": 345.1, "end": 374.86, "text": "are more specified and have bigger applications. The challenge is not to pass the test. So we're not going to pass the test because nobody's going to try to do it. It's not worth the money. Nobody's going to spend a million dollars on it. But I think the reason that the OECD got these people together wasn't, of course, just to be able to pass the test. It was to get a benchmark for the ability of AI systems relative to humans. Would you say about 50% of humans passed the test? Okay, so now let me tell you about how people perform on it. So they classified by... Let me just turn away from the lamp."}, {"start": 376.08, "end": 402.51, "text": "Yeah, we can take it off now. So they look at five level of performance, one, two, three, four, five. Go ahead, have a seat. And 50% of the people perform at level one or two. 35% perform at level three. And 15% perform at level four or five. Now, here is what, to me, what ended up. And remember, this is just a bunch of people spending a day making..."}, {"start": 402.51, "end": 430.93, "text": "Educated guesses on what they can do in one year with $1 million. And our collective wisdom, and this is between one group I was in, and I think you were in another group, all together about 12 people, that with essentially today's technology, $1 million one year, it means today's technology, you can reach median level performance, which is a median level, to about 50%. So to me, on this set of skills, the assumption is that"}, {"start": 430.93, "end": 458.93, "text": "Today's technology can accomplish what 50% of the population can do in the workplace on these kind of skills. I'm not sure Peter and Andrew agree with your experts in the OECD, though. You gave us two data points. No, no, no. You have to spend time looking in the hole. This really seems challenging. I gave you here, I would say, one of the most challenging problems. I mean, you cannot judge us on this. Some of the problems are very easy."}, {"start": 458.93, "end": 490.0, "text": "Mustafa or Daniela, what are your thoughts? I obviously agree. I think counting is a very, very hard problem, and I didn't quite figure out the first task, but that seemed like a non-trivial inference. I mean, this seems way beyond what we're currently capable of doing. Daniela, do you want to weigh in? I don't have anything to add. I agree. No, actually, I was not. You're not, I see. Okay, I saw your name somewhere. I thought you were in the world. Yeah."}, {"start": 490.0, "end": 519.01, "text": "Anyway, I said it's an educated guess and a different group of people might come up with different estimate. You know, we spent two days looking. I mean, here we spent five minutes looking at two randomly selected problems that I was able to grab off the web. The full test one has to go through this and make a decision. Well, I think this is interesting. The Allen Institute for AI has been working on TOEFL problems and SAT problems, I think, which I believe are maybe simpler than this, but that's been taking years, right?"}, {"start": 519.01, "end": 548.35, "text": "Go ahead. You wanted to say a little bit? You've been solving some of these kinds of tests. So we've been working on SAT problems and science problems. And it is very much as a benchmark. So I do see analogously to what Moshe and this organization was saying. I think that there is a lot of value in it as a benchmark for AI. So I guess to your point, is it worth spending a million dollars on it?"}, {"start": 548.35, "end": 575.12, "text": "Paul Allen seems to think so, but I've bet my career on it, so I guess I do too. But the reason being that it forces us to take some of these more challenging problems that Andrew mentioned, that are more structured, that involve less training data, and figure out how to build systems that can tackle them. Can your systems answer these? I beg your pardon?"}, {"start": 578.16, "end": 588.78, "text": "Not very well. So we know that because we've had quite a bit more than that to work on the fourth grade."}, {"start": 588.78, "end": 617.36, "text": "Fourth grade science problems, the regents, New York science problems, and we've reached something like maybe over 50% over the entire test, which has diagrams, has problems analogous to the ones that Moshe did. By the way, you just showed the bottles. You didn't show the question, right? Understanding the question in the first place is an extremely hard problem. So just to make sure I heard that, right now you can perform it about the median level of fourth graders?"}, {"start": 617.58, "end": 647.34, "text": "On the science test? Yeah. On science tests that involve diagrams and so on, I'd say we're below the median. If they're just pure natural language multiple choice, we're approaching 75th percentile for fourth graders. I should also mention we did a Kaggle competition where we invited several thousand teams participated. They had a limited period of time. This was on these non-diagram multiple choice. None of the teams was able to do more than 60%."}, {"start": 647.34, "end": 677.57, "text": "Now again, not everybody participated. 60% of the questions are 60 percentile of the human population. 60% absolute score on an eighth grade science test. And how do eighth graders do? Great question. I haven't calibrated exactly. Because the problem is it's a subset of the test. So eighth graders, median eighth graders do substantially better because they tackle problems with images and so on and so on."}, {"start": 677.57, "end": 704.43, "text": "Look, as I said, it's an estimate of two days, and one could argue, and a group of people will come up with different things. Our conclusion was that with that technology, we can perform with the right investment at the median level, and with 10 years, we think we can reach level 3. Level 4 and 5, for example, these questions of counting bottles, this is at the high end. Level 4 and 5 seems, at this point, we thought, out of reach."}, {"start": 704.43, "end": 733.62, "text": "But remember, I mean, level one and two is the median level of performance. Level three is the, where are we at level three? We're at the 85th percentile of the working population. So, Mustafa, you've been working on a whole set of questions. You've had some huge successes in some areas. And, you know, the cover stories in Nature and all. But I'm sure you've also had failures. Tell us about some of the things that you found especially hard to do. And what have you learned from..."}, {"start": 733.62, "end": 750.22, "text": "What's easy to do versus hard to do? Like Daniela put up the little quote from Mark Zuckerberg that their whole class of problems are in some ways similar to each other that seem we're making a lot of progress on. Where have you been most disappointed or feel like we need to make more progress?"}, {"start": 750.22, "end": 778.94, "text": "Yeah, it's a good question. I mean, one of the areas that we haven't made a lot of progress on, which I think is very important, is this question of algorithmic transparency. So for lots of reasons, it's really useful for us to be able to understand where in the network a particular representation for a class label sits. Knowing where that label sits or where in the network that representation sits allows us to bias training, for example, towards a specific outcome that we intend and therefore have"}, {"start": 778.94, "end": 792.37, "text": "More control over the algorithms that we train for specific purposes. And it's also useful for debugging and problem solving as an engineer. But in some ways, more importantly, I think..."}, {"start": 792.37, "end": 818.16, "text": "People want to understand why a particular network is making an important classification or a judgment. What is the underlying training data that has been used to produce a particular class output? Or what is the kind of similarity comparison that is being made when, for example, a credit score is being issued or a decision is being made? And so the more..."}, {"start": 818.16, "end": 840.69, "text": "Give me some practical examples of some things that your systems are doing or where they're having more difficulty. That was a good sort of abstract description. What are some of the systems that you have working now or that are in process that maybe took longer?"}, {"start": 840.69, "end": 862.54, "text": "Well, one of the practical examples would be around radiology. So we've been working very hard in DeepMind Health on a bunch of different radiology areas. One is ophthalmology. So there's clearly a lot of value in us being able to parallelize and standardize the human ophthalmologist. Where do they look at the eye and diagnose?"}, {"start": 862.54, "end": 886.99, "text": "What concrete are they doing? We basically take a three-dimensional eye scan and we try to label the pathologies that have been identified by a human expert. We haven't published on this yet, but in the next couple of months we have a publication coming out that looks very, very promising that we can essentially perform."}, {"start": 886.99, "end": 910.62, "text": "At human performance on a whole range of pathologies in 3D. At expert pathologist level performance. Right, right. And so this is clearly a hugely valuable problem and I think will speed up and standardize early detection of a lot of really critical diseases like age-related macular degeneration or glaucoma."}, {"start": 910.62, "end": 940.54, "text": "Even diabetes, which is an enormous problem. But clearly we would also like to be able to understand why the network was making that diagnosis, be able to interpret it, and be able to provide as much clinical guidance around that final recommendation to a human expert as possible. And so that means that we have to design slightly different networks that are able to reveal where in their representation... Right now it can make a recommendation, but some percentage of the time it may be wrong. It's not perfect."}, {"start": 940.54, "end": 962.26, "text": "And then the human is stuck with the situation of, well, the system says to do X, and do I trust it? Right. That's exactly the case. And it can't explain why. Yeah. Okay. To give an example, I think consumer loans is another one. Today, we use machines to make decisions on whether to approve or deny a loan application."}, {"start": 962.26, "end": 985.73, "text": "And if you deny loan applications, you kind of owe the consumer an explanation for why. So we actually struggle with that when these decisions are driven by neural network. Well, picking up on the same question I asked, some are related. When we talked, I don't know, six months ago, you said there's sort of a rule of thumb for the kinds of things that machines could do well today. Well, go ahead. Why don't you tell us what you see as some of those criteria?"}, {"start": 985.73, "end": 1001.14, "text": "Sure, you know, so many months ago at Baidu, we had a lot of product managers that are trying to struggle to understand what these new exciting AI capabilities that many of you developed, what can we now put into our products, what we use for consumers."}, {"start": 1001.14, "end": 1019.04, "text": "Gave them this, frankly, vastly oversimplified rule of thumb, which is anything that a typical human can do with less than one second of thought, we can either now or in the near future probably automate with AI, such as look at a picture and tell me if there's a face in this picture, or look at an email, tell me if it's spam."}, {"start": 1019.04, "end": 1048.37, "text": "Counting those bottles took a little more than one second, I guess. Yeah. And so what happens is a lot of complex AI components are made up with a lot of these simple components, which only take a second of thought. So for example, the way that we do a conversation-powered computer, right? So you talk to a computer and say, please order takeout delivery for me or something. There's an AI component that does a speech recognition, which you could do in less than a second. And then there's an AI component that figures out this is a food delivery query. There's another."}, {"start": 1048.37, "end": 1076.24, "text": "So a lot of these components, each of which a human can do with less than one second of thought. And we string them together in a clever way with a complicated way for how to collect data and so on. And this turns into a very valuable product. Now, this one second rule is clearly imperfect. It makes both type one and type two errors. But we found it a useful heuristic. So we do have many product managers running around looking for things that they can do in less than one second of thought and trying to automate that. That's actually productive."}, {"start": 1076.24, "end": 1093.01, "text": "This sounds like it could be very useful. What do the other panelists think? Daniel, you put up a chart of a whole bunch of things from McKinsey that they thought were automated or not. I gathered you sort of roughly agreed with it when you were talking about it. Right. So I guess the criteria were repetition."}, {"start": 1093.01, "end": 1108.75, "text": "And predictability. And with robots today, predictability, a predictable environment is absolutely critical. A slow moving environment is absolutely critical. That is why you can do self-driving in..."}, {"start": 1108.75, "end": 1137.36, "text": "Simple environments where you don't have too many things buzzing around and you have fairly slow-moving vehicles that can keep up with the perception and control systems of today. So with robots, this kind of predictability is critical. And with data crunching, the patterns are really what the system is all about. Actually, you mentioned about self-driving. You have this project going on in Singapore. Is it a controlled environment? Tell us a little bit about how you made that work."}, {"start": 1137.36, "end": 1159.62, "text": "So we've been working on self-driving vehicles in Singapore since 2010, primarily because Singapore opened the roadways and their data sources to us at a time when that was not possible elsewhere. And we have golf carts, the ones you have seen. We also have some electric vehicles and these scooters and wheelchairs."}, {"start": 1159.62, "end": 1185.68, "text": "And they are part of the ecosystem of the National University of Singapore. So in fact, they move on all the roads at the National University. There's pedestrians walking around. There's pedestrians. And in fact, these vehicles have these big red signs that say, danger, robots. But the students have seen them before, have been. How fast do they go? Well, they go up to, let me make the conversion. So it's."}, {"start": 1185.68, "end": 1207.98, "text": "We can do kilometers. 25 kilometers an hour. So they can go up to 35 kilometers an hour. And we run them between 15 and 35 kilometers an hour. And the students just know that these are vehicles that are part of their ecosystem. How many fatalities so far? They're moving slowly. So the critical point is that..."}, {"start": 1207.98, "end": 1236.24, "text": "In moving slowly, you can actually process the visual feed and the laser scan feed, and you can see that there is an obstacle in front of you, and you can compute an alternative path. Are you saying the computers aren't fast enough yet? Well, I am saying that our ability to do perception is not reliable nor fast enough. Would a few more ticks of Moore's Law just automatically solve that problem? I think we need advances in the algorithms."}, {"start": 1236.24, "end": 1252.91, "text": "Honestly, I feel like perspective on self-driving cars, and this is our role strategy in China, which is I don't think the way to build self-driving cars is to try to make it drive exactly the way a human does. If you run across a construction worker, a construction worker does that, you should stop."}, {"start": 1252.91, "end": 1282.19, "text": "They do that, you need to go. This is a critical decision you better get right. I don't know anyone that has a computer vision team that can reliably distinguish this hand gesture and this hand gesture at the, you know, having sufficient accuracy needed to make this critical decision. So I think Daniela said something very insightful, which people around the car got used to it. So I think the roadmap for self-driving cars is not to try harder to make a car do everything a human does. It's to change the regulations and change the expectations of the people outside the car."}, {"start": 1282.19, "end": 1304.24, "text": "How do you fit this into, give the car context so you can drive it safely? Another perspective is part of our project with Toyota where the objective is to make a parallel autonomy car not a series autonomy car. So in other words, the perception system and control system of your car watches what you, the human driver,"}, {"start": 1304.24, "end": 1328.34, "text": "I heard Google wasn't so thrilled with that approach. Is that right? I guess we want to be sure."}, {"start": 1328.34, "end": 1357.82, "text": "I think there's sort of this uncanny valley type problem that if you have an assistant that does a little bit of work for you, that's great. It's very useful. And if you have autonomy that does 100% of the work for you, that's also great. But if it does 90% to 99%, you're in a lot of trouble because we just couldn't get people to pay attention. We told them you had to be ready to take over, but once it passes a certain threshold of reliability,"}, {"start": 1357.82, "end": 1384.37, "text": "They don't. They're not willing to take over. We got Gil. Gil wants to... So Peter, this is different than... So to clarify the answer both from Waymo's point of view... This is Gil Pratt just to identify your... Oh, sorry. Gil Pratt from Toyota Research Institute. There's actually two different things you guys are talking about. The Guardian Angel is a parallel system that only intervenes when the person is about to have a..."}, {"start": 1384.37, "end": 1410.18, "text": "Blunder. And I think there's a whole lot of neat things that AI can do as in many fields, actually, as a guard against human error. And it would be neat to chat about those sorts of things. The idea of a handoff back and forth where the person has to watch over the system and the person guards the AI and has to make sure that that doesn't make a mistake. I think that's the thing that Peter was mentioning there. So the mirror image of that in a way. Correct."}, {"start": 1410.18, "end": 1434.88, "text": "But I think we are kind of, in my opinion, we are kind of a bit digressing because I think Andrew made a very, very important point, which is, and this is in some sense the flow in this experiment we did with can a computer pass this test. It kind of assumed that computers would perform the task in a similar way to the way humans did it. So imagine that we said, let's try to automate tellers. So it says, okay, they have to have a conversation with the customers."}, {"start": 1434.88, "end": 1464.32, "text": "So they need the speech recognition, speech synthesis, machine vision, and a whole bunch of skills. And the answer is, of course, ignoring the issue whether we have more tellers or less tellers, we re-engineer the process to make it amenable to machines. And this is generally, if you look today, you look at, for example, the famous picture of the Tesla factory, which is robots only, they do it differently than humans would do it. And I think this is the right answer. If you want to automate something,"}, {"start": 1464.32, "end": 1492.99, "text": "Don't think, how do I emulate a person? Re-engineer the process. And that's where we are much more effective than just trying to emulate people. So I would just like to make the observation that today when we work with machines, we have to adapt to machines. But it would be wonderful to develop the field so that machines adapt to us. So that we do not require biologists and physicists and English and philosophers."}, {"start": 1492.99, "end": 1521.62, "text": "Specialists to learn how to use the computer, but we get the computer to naturally support the different disciplines, different problem solving in the language of that domain. And this is not something that we know how to do, but it's an interesting challenge. But I wanted to go back to the point that you started discussing with Daniele before, which was more people study coding. So remember that right now in the United States, 50% of the people perform at level 4 and 5."}, {"start": 1521.71, "end": 1549.1, "text": "And the natural thing to say is, okay, let's improve our education system and move more people up there. And the answer is if you look internationally, I think Finland does the best. And you look at people between 25 and 34 in Finland, they do the best. And there you have 25% of the people performing at level four and five. And so this is kind of the high watermark. And so the idea is with education, we are somehow miraculously are going to move all these people up."}, {"start": 1549.3, "end": 1577.36, "text": "It's a bit unrealistic. I want to talk about education. To set a broader context, the issue of ethics has come up a lot of times in this conference. I'm going to give my personal opinion. I personally think here are the three most pressing ethical issues facing AI. I think number one is job displacement. Number two is openness and transparency. And number three is bias. That's my opinion. I'm open to changing my mind. I would love your feedback on that."}, {"start": 1577.36, "end": 1606.64, "text": "But just to add just number one for now, you know, it turns out that in the workflow of machine learning, we're actually very good at, we're relatively good at automating things that people can't do. There are a lot of reasons for that. If it's a task that a human can do, like ophthalmology, we can get labels from the humans. And then there are various reasons why, oh, and compositionality. You know, the brain is compositional, neural networks are compositional. Yoshua was talking about this. Jan's talked about this. So maybe the architectures kind of match. So when you look across the machine learning landscape,"}, {"start": 1606.64, "end": 1622.62, "text": "And I think to quote something that Eric once said, this means that a lot of work in AI is building substitutes for human labor. And I think the job displacement thing is coming."}, {"start": 1622.62, "end": 1651.98, "text": "What is the solution? I think there are multiple pieces. I've seen a lot of corporations in the past invest in education and not get a huge ROI. I think that's just been observed. But I feel that the game of education has changed. And if you look at the bleeding edge of how we educate people, I feel that we now do have the ability to educate people at scale far more effectively than anything done before. You know about my involvement with Coursera. I think Coursera is just part of the solution. I think we actually need much more than Coursera."}, {"start": 1651.98, "end": 1664.91, "text": "I feel like we now understand human motivation or human engagement much more than ever did before. And some of these things aren't implemented in Coursera or any of the root platforms."}, {"start": 1664.91, "end": 1694.19, "text": "Just as it's a huge challenge to build the K-12 educational system and the university educational system, I feel like we now understand human education well enough to build a new type of educational system. And this will take massive societal will, just like it did before when preparing our society for the type of work that is available to do today. But I think that as the job displacements comes, I think it's incumbent on all of us who are creating some of these job displacement problems to help solve this."}, {"start": 1694.19, "end": 1715.94, "text": "What we know today of how to build these scalable and actually effective educational systems, I believe it is possible for us to do this over the next large number of years. And just to hit the point really hard, you think that we can build a system to keep up as fast as the AI systems are affecting different kinds of tasks and jobs so we can have a kind of an equilibrium where people are learning new tasks?"}, {"start": 1715.94, "end": 1738.18, "text": "Yeah, so I think that, yes, Sean says yes. And I believe it's possible to teach creativity, for example. I believe it's possible to teach entrepreneurship and intrapreneurship. And there are so many jobs that we just can't find people to do, and we seem to be creating new jobs all the time. And all of us are so busy, right? I wish there were more people to do my job, frankly. So I would love to set an educational system to do a lot more of these things."}, {"start": 1738.18, "end": 1760.9, "text": "But look at the list of jobs, nice list of jobs that Daniela presented. And think of all the truck drivers that are going to lose their jobs. And how many of them are going to be moving, become data visualizers? Well, so the truck drivers are not yet losing their jobs. The truck drivers are getting support to be safer drivers for now. And we have some time to think about, we have..."}, {"start": 1760.9, "end": 1780.1, "text": "Actually, let me just ask you about that. The miners are so happy that look how they voted."}, {"start": 1780.1, "end": 1797.52, "text": "A number of you have been working on self-driving cars. There was a recent prediction that 20% of the vehicles in China will be highly autonomous by 2025 and 10% will be self-driving by 2030, according to a group of experts commissioned in China. Do you agree with that estimate?"}, {"start": 1797.55, "end": 1817.76, "text": "So I find it hard to make predictions. I like saying what I would like the world to be. And if you think about the fact that five, six years ago, nobody was talking about self-driving cars. And look at the frenzy that we have today and all the advances. I can really see our progress."}, {"start": 1817.76, "end": 1846.26, "text": "Picking up pace and getting to the point where in some short amount of time we could provide increasingly more autonomy. So you're talking, you're quoting 10 years to 15 years. Yes. And it's really possible that the technology could get there. Anybody agree or disagree with that? It feels very challenging but plausible. Plausible but challenging, okay. But we should separate the technical issues where 15 years seems plausible."}, {"start": 1846.26, "end": 1876.3, "text": "But imagining the transition from one system to a whole other social system, I don't think any of us is good at it. Let's pick up on that a little bit. Andrew mentioned it was the number one issue, as he saw it. Mustafa, we had a little conversation. What are some of your thoughts about the responsibility that people have to consider that as they're developing the technology or not? Consider the effects on the people's livelihoods. I think one of the problems with the way the question is framed is that every..."}, {"start": 1876.3, "end": 1904.35, "text": "A question like this needs to be bounded by some time constraints. What's going to happen over 10 to 15 years, 20 to 30, and then beyond? I think beyond 30 or 40, it seems, at least intuitively to me, quite likely that there will be very significant displacement. And I actually think we'll struggle to adapt and retrain and upskill, reskill people at the scale that we're displacing them."}, {"start": 1904.35, "end": 1925.65, "text": "The level of certainty or confidence in that assertion is very, very low because who knows, but I think it's significant enough that that warrants really rethinking a lot of quite fundamental things like, for example, our approach to taxation and redistribution, the extent to which we have people involved in some of these. But you're talking 30 years."}, {"start": 1925.65, "end": 1949.39, "text": "Plus. Plus, yeah. I don't think that's a near-term question. Near-term, I'm quite happy with Daniela's 10-year hyper-optimism around augmentation and working with machines. That's great. I mean, I think that's quite likely, but over a 30- to 40-year period, I think that's a bit ungrounded, to be honest. So, Mustafa, let's assume that we introduce mandatory computer science training. Our high school graduates in..."}, {"start": 1949.39, "end": 1973.87, "text": "10, 12 years will have all the tools to participate in the IT industry. And this is totally possible. In fact, I can quote the example of one charter school in Boston where 90% of the students are Afro-American and Hispanic. And the school has the highest performance in all the standardized tests. It's also the only school that has"}, {"start": 1973.87, "end": 1997.89, "text": "Consistent computer science education starting with grade one. And it's amazing. I think if you challenge kids, if you give kids interesting and exciting problems, they will rise to the expectation and will learn the material. Of course, the jobs will be very different, but they will have the kind of tools to participate in an IT economy. I feel like until humanity runs out of wants,"}, {"start": 1997.89, "end": 2028.59, "text": "There will be new job opportunities. The main challenge is whether or not people have the skills to pursue those opportunities. We can look at how we have performed over the past 25 years. We can look at manufacturing as a test case, and we can ask, you know, the standard thing is there will be new jobs, and we can look at this 8 million people or so that lost their job, and where are they going? In fact, there was just an article, I think just yesterday or the day before that, about what the new jobs are."}, {"start": 2028.59, "end": 2054.38, "text": "And the answer is the new jobs are many of them in services and health aids. And women take these jobs. Men rather stay home than take these jobs. We can argue why it is the case, but this seems to be the case. That's why we have seen now men labor for participation for a prime age male, dropping for the past 50 years. And so we can say, well, we teach them all computer science, but we better do it very quickly."}, {"start": 2054.51, "end": 2082.21, "text": "Because they're simply dropping out. One of the reports that was mentioned earlier about coming from the White House was exactly about these men who were disappearing from the job market. I feel like some of the skepticism about a new educational system is because maybe they haven't articulated a concrete proposal yet, and in the absence of something concrete, it's hard to say yes or no. So let me actually try to describe the very broad outlines of something that could be the beginnings of something that I think could work."}, {"start": 2082.21, "end": 2106.75, "text": "So first I think the data on the ROI of education is very clear. Multiple studies done at all levels of education showing that people learning new things has a very high ROI for the individual and in most cases for society as well. So let's take as a given that we can train people economically that there will be a high ROI. Now we also know that participation in learning activities, formal learning activities like attending universities or taking courses or going to school."}, {"start": 2106.75, "end": 2114.46, "text": "That actually drives learning. I think one of the biggest challenges facing education today is motivation or keeping you engaged."}, {"start": 2114.46, "end": 2143.76, "text": "Now, there's also been a lot of discussion about basic income, the safety net, which I support. But I would love for society to gravitate toward a version of basic income where we pay you not to, quote, do nothing, but we pay you with an expectation that you will study so that there is a high likelihood that you will gain the skills that enables an unemployed person to reenter the workforce and earn a living and contribute back to the tax base that in turn is contributing to basic income."}, {"start": 2143.76, "end": 2167.06, "text": "It's a huge restructuring of how we think about redistribution in the educational system, but I feel like something like this could be the beginnings of a new deal. People have talked about an unconditional basic income, UBI, but this is really a conditional basic income that ties it to... Would you tie it to time or specific accomplishments, or are you open on that?"}, {"start": 2167.06, "end": 2196.05, "text": "It really details TBD. In the educational world, there's a lot of attraction and competency-based education where we don't care. Instead of measuring seat time, how long you sit there listening to whatever, we measure your ability to demonstrate the actual learning. These are the questions kind of at the bleeding edge of where the education field is thinking, but I think we should be open on the question you asked. Go ahead, Danielle. I really like the idea of tying the unemployment checks to taking classes to education."}, {"start": 2196.05, "end": 2226.59, "text": "Thank you. What do you guys think? Do we have a proposal here? Go ahead, Peter. I think it's important to distinguish labor displacement versus job displacement. So labor displacement is great. I get a tool that helps me be more productive. But job displacement, I no longer have a paycheck. And the job provides productivity for society. It provides income for me."}, {"start": 2226.59, "end": 2251.94, "text": "It provides my sense of purpose, but I think in some ways, most importantly, it's mutual insurance policy, where I say working 2,000 hours a year at the same place may not be my most productive use of my time, and for the company, they may not need me for all that time, but it means neither of us has to worry about it."}, {"start": 2251.94, "end": 2282.54, "text": "I think we're seeing, not just through automation, but also just through communication, we're seeing more of this gig economy. So Uber doesn't have self-driving cars yet, but they're displacing taxi drivers because of communication, not because of automation. So whether or not there's battle between inventing the new jobs versus displacing the old ones, no matter which way that goes, we're gonna see much higher variance."}, {"start": 2282.54, "end": 2311.25, "text": "And that people are going to have up and down years. And so we need some insurance policy to replace that. And I don't know if universal basic income is it or income tax credits or guilds. But there has to be some insurance policy to make up for this no longer stay at one job for multiple years to decades at a time."}, {"start": 2311.25, "end": 2339.78, "text": "Because that's inevitable no matter how much automation happens. Can I make a comment on that? Let me just say, I'd be happy to take some questions from the audience. I think there's a couple of them here. Maybe we get some mics over here. Is that Eric? Josh Green, Harvard. I'm trying to adjudicate in my mind between the optimists like Daniela and the pessimists. And I have a question about a way of thinking about this and wondering, is this a good way to think about it? And when you think about it this way, does it change?"}, {"start": 2339.78, "end": 2368.34, "text": "Your answer. You can imagine an economy, let's say, hypothetically, in which everybody gives each other massages. And one hour of massage giving results in one hour of massage receiving. And you can see how that scales. By contrast, if someone makes an app, say it takes someone 500 hours to make that app, but it generates thousands of hours of use. Given that however many new wants may be created, you're not creating more time for people."}, {"start": 2368.34, "end": 2397.66, "text": "If you think about the tech jobs of the future in terms of the ratio of the amount of time spent on labor and the amount of hours of service provided per hour of labor, if it turns out that there's a lot more hours of service provided by the most desirable products than the amount of labor time that goes into them, does that inevitably lead you to the conclusion that no, there's just no way you can keep everybody busy making apps or whatever the new technological employment world"}, {"start": 2397.66, "end": 2416.0, "text": "So I'm wondering, does this way of thinking about it make sense? Has anyone done this kind of analysis looking at the amount of labor time put in versus service time put out? And if you do look at it that way, where do you come down on the question of whether we could have a sustainable economy of knowledge workers?"}, {"start": 2416.0, "end": 2432.14, "text": "So you say that like it's a bad thing, I mean, that we have more wealth with less work. The question is, it would be good for the people who are making the apps that everybody loves and spends thousands of hours on. What happens to the people who put in $500 hours on an app? And so this is getting to the issue of"}, {"start": 2432.14, "end": 2454.06, "text": "Replicability and scalability that you, Eric, discussed in your talk. So I'm just looking for general reactions to that. So I think that once upon a time, I think we used to wonder if everyone needed to be literate. And then eventually we thought, well, maybe only a few people need to do the writing and everyone else just needs to read. Because surely with a few best-selling authors, everyone just reads what they write. Everyone else doesn't need to learn to write."}, {"start": 2454.06, "end": 2472.64, "text": "But what we found is that now that most people are literate, we now write for audiences of one. Like I send my wife an email, my wife sends me an email, and that creates value. Right now, CS talent, computer science talent is so scarce that almost all the CS talent is dedicated to what the applications that attract the attention of mass audiences."}, {"start": 2472.64, "end": 2498.67, "text": "I can't prove, but I believe that if CS literacy becomes widespread, which Daniele was pushing for, and I think is a fantastic thing, we'll be able to have a mom and pop store where the husband and wife decide to write code for a specialized display for their little grocery store, and they can program something just for their store. You send your wife an app, she sends you back an app. You send your wife an app, she will send you back an app. Yeah."}, {"start": 2498.67, "end": 2526.5, "text": "Mustafa, you thought a little bit about whether we need new governance models for this? I think the point is not that we want widespread literacy. Of course, everyone would argue for universal access to education. The question is, is there going to be enough dollars attached to that that we can actually fund very, very large populations to write or to build apps? And I think that's the point of the question, is that there's sufficient uncertainty around that answer that we should..."}, {"start": 2526.5, "end": 2556.38, "text": "For the sake of building an insurance policy, we should be skeptical about it. And that's just an approach. It's not pessimist versus optimist. It's just we do need to think about insurance policies, especially as we see this colossal rise of the populists who, rightly or wrongly, whether it's connected to the substance or whether it's just a feeling that they're being hard done by, are very frustrated and are increasingly pissed off and are going to be directing more and more of their frustrated energy at elites like us who..."}, {"start": 2556.38, "end": 2581.86, "text": "wield incredible authority and power. And I think that's something that we have to take on very seriously and consider with respect to policies like more progressive taxation or more transparent governance. We're just about out of time. Peter, Wayne, then I'll have a quick question for all of you at the end. I think there's some things for which we just prefer to interact with people. And maybe your massage is one of that. There could be a robot massager that's just as good, but you'd rather get it from a person."}, {"start": 2581.86, "end": 2609.81, "text": "You know, you'd rather get your likes to your posts from a real person, even though I could write an app to generate 10 million likes. You're not going to be satisfied with that. Right. We watch sports with real humans, and even if a machine could do it faster. So we have to wrap up, I'm afraid, but let me just ask you all a closing question. So imagine that you find yourself in New York City next week, and you're riding up this gold-plated elevator, and in steps the president-elect."}, {"start": 2609.81, "end": 2637.95, "text": "And to your surprise, he's read the White House report that just came out about AI and the importance of it. And he says, you know, you seem to know something about this. You have one minute as you go up to the 60th floor. What's your one piece of advice or how are you going to use that 60 seconds of policy advice? You want to go first? Massive investment in computer science education. And secondly..."}, {"start": 2637.95, "end": 2662.32, "text": "Thinking about algorithmic transparency and understanding how we can build and deploy machine learning systems and use them in high-stakes real-world environments and get around the regulatory framework to do that. Andrew? I'm going to say what I said just now again, just emphasize. I think number one ethical problem, jobs. I think we need education, maybe taxation basically comes on to solve that, but jobs."}, {"start": 2662.32, "end": 2685.26, "text": "Number two, openness and transparency. And I think number three, bias. I really think those are the top three ethical problems facing AI today. So I think jobs in the social security net that goes with it, either for long term or short term, because if you don't keep people happy, they'll vote you out like you did to the last guy. Am I alone with him?"}, {"start": 2685.26, "end": 2709.3, "text": "What's that? Yes, you're alone. I'm alone. I'm not your advice. Advice will be my top priority. But I would emphasize, rather than try to run into a solution, I would emphasize, because I don't think so far he's internalized it, it's not about trade, it's about automation. Okay, great. Daniel, you have the closing words."}, {"start": 2709.36, "end": 2725.3, "text": "Mr. President, American kids are falling behind. We need mandatory computer science education for all. And by the way, we also need significant investment in the science of autonomy and the science of intelligence. Well, thanks very much, guys. We are out of time. Appreciate you all coming up here."}]}