| 0:33 | Intro. [Recording date: September 18, 2014.] Russ: I want to start by saying this is an extremely interesting and provocative paper packed with ideas and analysis but still accessible for the most part to a nonspecialist. We'll put a link up to it and I encourage listeners to check it out after listening to the conversation. I also want to say that this is one of many episodes of EconTalk where we look at the threat and opportunity posed by the increasing role of computers and smart machines in our economy. And I want to let people know that more are coming. So, enjoy. David, let's start with Polanyi's Paradox. What is that and why is it relevant? Guest: So, Michael Polanyi is a Hungarian philosopher and scientist. And he wrote a book--I believe it was published in 1961--called The Tacit Dimension, in which he sort of expressed or articulated the importance of tacit knowledge in human behavior. The quotation I give in the paper is "We know more than we can tell." And then he goes on to give an example of the skill of a bicyclist or an automobile driver cannot be simply articulated in words. Or more concretely, I could give you a day-long lecture on how to ride a bicycle; at the end of the day you wouldn't know how to ride a bicycle without having ridden one. And the point that Polanyi was making is that there are many things that we tacitly understand how to do or are capable of doing that we do not explicitly understand how to do and cannot articulate in terms of a procedure. So, riding a bicycle would be one example, but there are many, many others. We don't know the procedure for coming up with a new hypothesis or for making a persuasive argument or for that matter for recognizing people from a distance or even recognizing someone as they grow up that you haven't seen in 15 years and they've changed greatly--how do you know that's the same person? Or how do you choose the sort of locomotive path to navigate along, up a steep hill, a rocky surface, and catch your falls in real time. These are all things that we do as onboard equipment. Or another great example would be interpreting the nuances of spoken language, more than just the words individually. All[?] the things we know how to do, they are sort of built in, part of our hardware; but we don't know how to accomplish them explicitly. We don't know how to write procedures or describe the rules for doing those tasks. And so the point of this paper is, that makes an extreme challenge for computerization, for automation. That automation primarily works on taking the explicit procedure that we already do and codifying those steps so that a machine can do it in our place. So, when we have a computer program that does calculations or sorts through files or searches for words or helps us lay out a circuit board if you're a CAD (Computer-aided design) user, it's following a set of codified explicit procedures that we already understood and now we have laid out the steps. But hard to do that if we don't actually know the procedure for the thing that we're accomplishing. Russ: One of the things--I think I've probably told this story before, but one of the most beautiful and inspiring and moving videos I've ever seen is one that looks at Andrew Wiles's attempt, and finally successful attempt, to solve Fermat's Last Theorem. He had had this proof of the theorem that was knocked down. But for a while he was celebrated as one of the greatest mathematicians of all time. He must have been exhilarated for a while. And then it turned out: the proof's not right. And then he spent a rather horrible set of time, as you might expect, trying to resurrect that proof. And he couldn't. And to hear him talk about it, he says--and then one day, one day, I looked up or I put my head down and I looked up and he saw how to solve it again. And he can't explain what happened there. He has no idea of that intuition, that aha, that eureka moment. To me it's one of the most beautiful, mysterious parts of the human enterprise. Of course computer scientists wonder and philosophers wonder whether it's just a matter of time before we understand that process. You want to weigh in on that? Guest: Yes, absolutely. First, that's a great example you just gave. And let me emphasize that this paradox--the paradox being that we do things that we don't know how to do--applies as much to the mundane as to the sublime. It applies as much to proving a century-old mathematical challenge--that's one example. But something as simple as walking up a flight of stairs or looking at a garbled piece of text and figuring out what it says, which is of course what you do with these CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart) all the time. Russ: Or empathy. When someone walks into a room who has had a bad day and you recognize they need a little affection or kindness or a cold drink. Guest: Right. That's right. So, both of these things require, draw on a set of capabilities that we possess, but they are not in the kind of accessible, analytical, procedural set of steps that we would use to [?] have a machine do the same thing. In terms of our ability to get those things, or to automate those things--clearly, we've made progress on automating many things that initially seemed extremely challenging. So, people actually initially thought, like, teaching a computer to play chess would be just an extremely hard problem because we know that chess masters are geniuses, etc. And as it turned out, actually, that problem was solved relatively quickly. And now, any inexpensive chess-playing computer piece of software can pretty much beat the world's best chess players. And in fact this observation that many things that initially appeared hard turn out to be simple, and things that appear simple turn out to be hard, has a name. It's called Moravec's Paradox. And many things that--artificial intelligence thought it would be easy to have robotic servants that would, you know, empty your dishwasher-- Russ: See Sleeper by Woody Allen. Guest: Exactly. That's right. But it would be hard to get the computer to play chess. Actually the opposite has turned out to be true. And there are two reasons, or two--so, there are two reasons why we've gotten good at some of these things and not others. One is that it turns out that there's a well-known procedure that you could use. So you could algorithmatize how to play chess and find the optimal solution for problems. You could write that out as a set of equations. Now that's not true for chess, actually. Chess doesn't have a sort of closed-form solution. There is no one dominant strategy in chess that anyone knows about. But instead what computers do is they use a lot of processing power to basically iterate through multiple steps, multiple levels of moves, and choose after calculating many thousands or hundreds of thousands of possible sequences of boards, they then choose what appears to be the best strategy, given that kind of forward looking search. And they may also do some kind of database lookup of prior games. What computers do when playing chess is probably pretty different from what Grand Masters do, in the sense that they are using the kind of comparative advantage of computation in doing lots and lots of quick calculations and storing the information accurately. Whereas chess Grand Masters are probably much more likely to be using a kind of mixture of intuition and recall of previous games-- Russ: Pattern recognition. Guest: Exactly. Russ: Elegance. Guest: Yeah. So the question is: Will we get better at all these other things the same way we got better at chess? And I think the answer is: Probably in the long run. Very few people doubt that in the long run, over 10 years or 20 years or more likely 30 or 40 or 50 years, many of these problems will make substantial progress. It's more a question of a). how long it will take, how hard the challenges, and b). how we'll do it. Will we actually have machines that will do what we do, in terms of having intuition or figuring out problems in this kind of nonprocedural--as far as we know--way? Or will they in fact, will we recast the problems or recast the technologies so they do it very differently from what we do and yet still successfully. |
| 9:47 | Russ: Talk about the driverless car. Which we've talked about a number of times on the program. There's a really beautiful de-romanticization of it in your paper. Guest: Okay, good. So I [?] lay out two sort of central strategies for dealing with what I call in the paper nonroutine problems--problems that we don't know the explicit procedures. One of those strategies is what I call environmental control. Environmental control means simplifying the environment so that something that is far less flexible than a human being can manage in that environment. So, you know, just a very literally concrete example: If you think about modern automobiles, they are reliable, they are fast, they are safe. They are extremely competent. And yet, actually they require smooth, even surfaces with shallow grades and turns that are not too tight. Something that would never occur in nature. And so, what do we do? We change the environment to make it car-compatible. So, it's estimated that an area the size of Ohio in the United States is covered by impermeable surfaces, most of which are roads. So that's why I said, a 'concrete' example. So, environmental adaptation is applied in lots of areas. Assembly lines, basically make things very predictable and consistent; that's why it's easier to use robotics, robots in an assembly line, where there's a very narrow scope of activity, than in an uncontrolled home or work environment. Russ: [?] Guest: Exactly. So, I make the point in the paper that actually this environmental control is also applied very frequently to computerization problems--that you basically make the environment predictable in a way that allows machines to adapt to it. And the example I give, of the Google car, is: Many people think of the Google car as being flexible, like a human driver, right? Russ: Smart. Guest: It goes--yes, smart. And it is smart, in a way. But it's not--if you were to put the Google car in a--just drop it down in the middle of a city, it had not been prepared for--and I'll explain in a second what prepared means--it would have to stop. It cannot in real time recognize a road, figure out where the traffic lights are and the stop signs, determine what the routing is, determine what the speed limits are, and so on. It's not that adaptable in real time. Instead, the Google engineers basically go through their mapping software and then hand-curate the maps on which the Google car, for the roads the Google car will navigate, identifying all the stop signs, all the traffic lights, all the routing and speed limits. The Google car still has to be adaptive in the sense it has to, you know, recognize objects in its way--other vehicles, pedestrians. It needs to tell whether a light is red or green. Etc. But it doesn't have to figure out the entire environment in real time; and in fact if it encounters a substantial deviation from the environment it's anticipating, it needs to stop. So, if there's a signal man in the road--so if there's changing the traffic routing--at that point the Google car has to seek control to the human driver. It's not that adaptive. So, what I think, the paper somewhat is a little unfair: Although the Google car appears as flexible as a human being, in some sense it's more like a train driving on invisible tracks. Now, that's a little bit of an overstatement, because of course a train doesn't recognize pedestrians and cars and come to a stop. And a train can't swerve out of the way. But nevertheless the Google car is, the tracks are sometimes laid out for it, and then it needs to only react to deviations from the tracks it hopes to be driving on--obstacles in it's way. Russ: Yeah. It's a fascinating example. I guess the question is: As our knowledge advances of the brain, are there going to be limits? Now most--Robin Hanson on this program, long, long time ago, basically said, 'you know'--and a lot of people agree with him--'it's just chemistry; just a matter of time before we figure it out.' And that might be 10, it might be 20, as you said, might be 50 years. The real question is: Is it going to be possible to ever understand what that Grand Master is doing, what Andrew Wiles is doing, through a chemical analysis of what's going on in the brain? I don't know--it's an unanswerable question-- Guest: Yeah. I don't think the answer will be through a chemical analysis. I mean, in the sense that, even if you thought, you said, I'm going to understand what a computer does through an analysis of silicon. Russ: Yeah, exactly. Guest: Right? That wouldn't actually be informative. Because it's actually about information. It's about symbolic processing. Right? So, the physical structure is actually somewhat divorced from the kind of meta-information-processing structure. That doesn't mean the brain can't be understood. I just don't think that a chemistry set is going to do it. It's really understanding--yeah. |
| 14:53 | Russ: No, I think that's exactly right. And I think what's fascinating about all these examples--and we're going to get into this a little bit more and then we are going to turn actually to data rather than speculation, which is probably a good idea. But what's fascinating to me is going back to a recent EconTalk episode with Paul Pfleiderer, where he talked about the tendency of economists to employ the Milton Friedman 'as-if' hypothesis. So, we don't know what managers really do. Or we don't pay attention, actually to what they really do. We posit something. And then we say, it's as if they do it this way. And I think we're doing this with this smart machine stuff. We are saying: Well, we don't know what a Grand Master does playing chess. But I'm going to act as if it's what the computer's doing when it sifts through thousands and thousands of different opportunities. But that's not what the Grand Master's doing. It may be somewhat predictive of what a Grand Master might do most of the time. It's not the same thing. You can't then go the other way and say, So we could just develop a computer, then; since we can develop a computer that acts as if it's a Grand Master, then we can develop a computer that is a Grand Master. They're not the same thing. Guest: Yeah. There's a big debate. You know. There have been many decades of setbacks and some successes in artificial intelligence (AI). And there's a lot of debate among our artificial intelligence researchers about how you should go about trying to master human activities with machinery. So, one school of thought is we need to learn from biology itself. That basically, if you look at the way the human brain recognizes objects--visual recognition starts about two cells back in the eye. It's not like it's just a data receptor call and then it goes to a central processor. There's a whole mechanism that's sort of built in for pattern recognition that is sort of fundamental to the hardware itself. And so some people think we have to learn from the example of how biology does it. Others say: No, no; actually all we need to do is have a conceptual model of the world, and then the machinery, based on the conceptual models, can process the information it receives and figure out what's what. It doesn't have to look, doesn't have to work--an airplane doesn't have to flap its wings to fly. And then there's a third school of thought--and there's many variants in between--that says: We don't need to do any of that stuff; we just need machines to learn to behave like us by basically learning from the world. So the idea of machine learning-- Russ: Talk about that. Guest: So machine learning is the idea--so Polanyi's Paradox is we know what we can tell; we do things but we don't know how we do them so we can't explain them. This gives rise to the idea of machine learning: instead of trying to write down the procedure that we don't understand for doing something, why don't we have the machine, a machine, that looks at examples, correct and incorrect answers, and then infers what is the procedure or a set of statistical connections that makes that the right answer. This is--a kind of iconic or at least highly discussed example was the Google cat recognition software, hardware thing--came from Google X Labs and it used 16,000 processors to parse through a database of millions of pictures of things that were cats and were not cats. And without any specific model except being told this has a picture of cat, this doesn't, and circles the cat. It would then attempt to look at new pictures where the cat was not circled and say: Which of these photographs or drawings contains a cat? And it was--if you look at the pictures which it recognized--one example is included in my paper--it's pretty good at mostly, it recognized a bunch of cats. One thing it viewed as very likely to be a cat turns out to be a pair of coffee cups next to one another. Russ: And coffee cups don't need a litter box. Okay. Guest: That's right. So, what is it doing? It's basically--it has no model in mind of the world. It doesn't say, a cat is a feline, is a biological organism with four legs, etc. It just says, here's pictures of things that you've told me are a cat, and I'm going to look at their statistical properties and then try to predict or infer in other pictures what else might be a cat that fits that description in some sense. The description you haven't given me but only just shown me examples. So, that's what someone would call a scorcher or brute force approach to learning. Russ: Big data. Guest: It's a big data approach. And it also has strengths and limitations. Strength of course is it requires only processing power. It doesn't necessarily require a huge amount of analytical infrastructure to build a model to teach about thinking about a cat. Its disadvantage is it may be fairly brittle and not very general in the sense that without reasoning about what the object is or the thing you are trying to recognize, you could get things that would be statistically unlikely and yet clearly incorrect. So, like the coffee cups. And that no human, no 4-year old kid would make that error, would look at that picture and say--even a kid with 16,000 processors would look at that picture and say, oh, those coffee cups are cats. They would see the difference. Perhaps the reason they would see the difference is because their reasoning about what makes a cat is much more sophisticated than the simple statistical features. Someone who understood a cat as an animal would say, well, it has to be an organic creature. What would be [?] being organic? Well, it wouldn't have a smooth ceramic surface. But of course that's a step, many, many more steps down the line of reasoning than simple recognition. It requires some knowledge of what the object is in the world. Which is a much harder problem. So, I give the example in the paper--in fact I'm quoting from a paper in the computer science community called "What Is a Chair?" And it talks about the difficulty of teaching a machine to recognize a chair. Because chairs come in a huge variety of sizes--not just sizes, features. For example, does it have to have a back to be a chair? Well, no, of course not; we know there are backless chairs. But then how do you distinguish a chair from a table? Well, you'd have to sort of look at the dimensions and sort of thing about: does it look more table-like or chair-like? What does that mean? And then--the example given in this paper that I quote in my article is: If you looked at a traffic cone and a toilet seat, you would say, well, they both look somewhat chair-like. They have a base; they have a top. Russ: Height's about right. Guest: Yeah, exactly. However, if you reasoned about human anatomy, you might think to yourself, well, a traffic cone wouldn't be that comfortable to sit on. So probably that's not a chair. That requires reasoning about what the object is for, not simply what physical features it has in a very basic sense. So that's a harder problem. So there's a divide in the computer science community--not a civil war, but a divide: Can we just do this by brute force--i.e., pattern recognition? Or does that in some sense, what I heard someone say, one of my MIT (Massachusetts Institute of Technology) colleagues is: Well, it gets it right on average and misses all the important cases. So, I think there is a great deal of uncertainty about what routes will prove most productive. I think all of them will, to some degree. So it's more a question--I think a lot of the debate is how fast things will progress. How quick will this move? Will these challenges be quickly surmounted as some people [?] believe? Are we still decades away from having, you know, domestic robots like Woody Allen's Sleeper? |
| 23:17 | Russ: So, let's move into the less speculative part of this issue, which is: Regardless of where we are going to end up--which is an interesting question to discuss, and that's what we've been discussing in this opening--regardless of where we end up, it's clear that computers have done a lot of substituting for some human tasks, because computers do it well and do it cheaply. And this has people worried that even if all the jobs aren't going to be eliminated, many or most of them will. We have computers diagnosing cancer, replacing potential high-paid doctors. We have computers doing all kinds of things they couldn't do 5 years ago, 10 years ago, certainly 20 years ago. So let's look at the impact of this so far. And you point out in the paper of course, this is very old worry, that technology is going to replace human employment. I just find it fascinating that the world right now kind of divides among pundits and even some economists: Well, people have always worried about it before, but they've always been wrong. Technology has been good for human beings; it's created other kinds of jobs at the same time. And the second view that says: This time's different, because this is going to get rid of all the jobs. Or almost all the jobs. So tell us what we know about what's actually happened, because there's some very interesting patterns. And you're the only person I know of who has actually looked at what's going on and at a more micro-granular level. Guest: Well, many people at Google[?] have worked on this at this point. But, let me sort of make three points. One, there is a long history of concern about the impact of automation on employment. The most famous example that people give is the Luddites, who were 19th century weavers who rose up against the power frame, the power frame loom, because they were afraid that it would it would reduce employment and earnings. And they very well may have been right for themselves. Because it did take scarce artisanal skills and basically substitute them with basically machines and children doing those jobs. In the long run, of course it didn't reduce employment. But it probably had significant distributional consequences. But more recently I don't think most people are aware of this: This concern, again, arose in the early 1960s under the Johnson Administration. And there was a Commission set up to investigate the productivity problem. And the productivity problem was that productivity was growing too fast. And the concern was that would mean there wouldn't be enough jobs. And in fact the U.S. Department of the Interior was in on this, talking about the potential 'leisure crisis'--the crisis being there would be too much leisure. Department of the Interior, I think they--they run[?] the national parks. And so they would have to deal with all these leisure seekers. So, the concern--so, I think those concerns, obviously we had a bunch of good decades and we didn't feel that we had any leisure crisis. But the concern has re-arisen. So, if you look at the Chicago Booth Poll of economists, the vast majority are saying in the long run there is no evidence that talent[?] technology has reduced employment. But then if you ask about sort of stagnating wages over the last decade, and the role of information technology, a kind of plurality of economists think that there might be a direct connection there. So, my work--with many coauthors including Frank Levy, Richard Murnane, and Larry Katz, and David Dorn and Gordon Hanson and Daron Acemoglu--I could go on, a whole list of notables but people who have been extremely valuable and insightful as coauthors. We have sort of pointed out the role that computerization has played in changing the occupational structure. In particular in displaying routine codifiable tasks, going back to the beginning of our discussion, which would mainly mean jobs in clerical, administrative support, some degree sales, and also many production and operative positions. All of which are skilled work that use a lot of codifiable, procedural activities. And those things, even though they are educated tasks, increasingly are relatively automatable. And the one consequence of this is that you sort of see this what someone would call a polarization of employment, that, on the one hand we don't have computers substituting for people who are doing professional, technical, and managerial tasks--you know, things that require intuition, creativity, expertise, and a kind of a mixture of fluid intelligence with technical knowledge. So, if I'm a scientist, I have lots of technical knowledge but I also need this kind of fluid intelligence for developing hypotheses. If I'm--same as if I'm in sales and marketing, if I'm an attorney, if I'm a medical doctor. So on the one hand, creating an increasing role for these very highly educated skill jobs that are not only not directly substituted, but substantially to an important degree complemented by information technology. Because information processing is kind of input into these occupations. On the other hand, it also leads to a relative growth in many in-person service jobs, like food service, cleaning, landscaping, personal care, home health aids, security guards. And these require tasks that have proved very difficult to automate, as per Polanyi's Paradox. But the irony is that the supply of workers who can do them is quite abundant. Right? So, although it's hard to develop a domestic or restaurant-serving robot, it's not hard for a person with their full physical faculties to have that job and do it productively, with very little training. Many people, it can be in the course of a couple of days. So we have a simultaneous growth of high-education, high-wage jobs, and relatively low-education, low-wage jobs. And those jobs are low-wage for a reason; and the reason being that the supply of workers who can do them is potentially extremely elastic. So it's hard for wages to rise in those activities unless essentially there are better opportunities elsewhere in the economy that you have to bribe people not to take. That's the sort of economic explanation: Why would barbers' wages rise over time even though barbers don't get any faster cutting hair? Well, the answer is because they need to be compensated for not doing something else in which, where they would have rising productivity. So people have to be willing to pay more and more over time for those type of jobs. So, we talk about this polarization phenomenon. And this has been documented across not only in the United States, but at least 16 European Union, EU, economies and the work, recent work by Alan Manning, Maarten Goos, and Anna Salomons. And it appears to be a pretty broad, widespread phenomenon, this decline of many of these middle-skill office and production jobs and relative growth of both high wage, high skill and low wage, low skill jobs. |
| 30:18 | Russ: So it raises a bit of a specter. And Tyler Cowen's book, Average Is Over, and the episode we did with Tyler on it. It creates this vision of the future if these trends continue which seems a bit ominous. A lot of successful, highly paid, skilled people, and [?] a much larger group of unskilled and poorly paid people. We've been talking about the employment data so far. Talk about what we've learned from the wage data. Guest: Well, so, the wage data are--it's a great deal more complicated. Though, there are several mechanisms by which that changes--so there are changes in task demands, translate into changes in wages. The three of them that I outline in the paper, one is: are you directly substituted or are you more likely to be complemented? Right? So, if you are a person who is an accountant who can only add and subtract, and you are a bookkeeper, clearly automation devalues your skills. You can just do that more cheaply with a machine. On the other hand, if you are an accountant who understands sort of the conceptual basis of the business and spots problems and creates, you know, valuable kind of record-keeping ideas or ways of organizing information to augment operation, then you are complemented by computerization, because of course you can accomplish more of the things you are good at in the same amount of time, because you have all this hardware to help you do it. Right? So in general I think it is vastly underappreciated in discussions of automation that automation generally complements us by substituting for the things that are time intensive and allowing us to focus on the things in which we have value added. Right? Russ: This goes back, by the way, to Adam Smith, when he talks about the division of labor being limited by the extent of the market, and he talks about the application of technology making people more productive. And it's very important. It has a long history. Guest: That's right. Adam Smith mostly invokes on the division of labor as sort of a way of increasing productivity. I don't think he fully appreciated the role of capital augmenting labor. Russ: I think he did, actually. Well, I'll send you an excerpt. We'll put it up. He talks about the kid doing some mundane task thinking about a way to mechanize it and thereby not have to work as hard; and making the thing more productive. It was primitive. He wasn't thinking about it. Guest: Okay, I stand corrected. That's fully[?] fine. Russ: It's surprising. Guest: So, I think it's easy to see the many ways in which machines substitute for the things we used to do. And then what's harder to see, typically, is: how is that complementing us? But of course you sit back and say, could you and I actually be having this conversation, could you and I have a podcast? Could I actually do significant[?] research as an economist without all this sort of hardware increasing my output per hour? The answer is: Not very well. So, on the one hand is the: Are you directly complemented, versus substituted? Second factor that affects how that automation affects your earnings in a given activity is sort of, the elasticity of final demand--so, in other words, if we get really productive at something but there's a fixed amount of it that people want, then eventually they just buy less and less of it. So you see, for example in agriculture, the vast increases in productivity in agricultural stemming from the green revolution and so on, have eventually reduced employment dramatically in agriculture. And the reason is that, all evidence to the contrary, there seems to be a finite amount that we can eat. Russ: It's a great example. So, food is incredibly cheap, which is a glorious thing; but it doesn't lead, therefore, to: Oh, there will be more farmers. There are fewer. Guest: That's right. That's correct. And other places, that's not the case. So, in medicine, for example, we get, we are much more productive at medicine than we were 50 years ago or 100 years ago where we mostly harmed people. Now we do lots and lots of useful procedures, and demand seems to be extremely elastic. So we spend more and more of our money on medical care because it becomes more and more valuable as it becomes more productive. Russ: Well, we've subsidized it also, so it's a little complicated. Guest: Well, right. There are many factors. But--okay. Russ: But we do want it. Guest: Right. That's right. And then--so, this again, many of the professions, people seem to demand more of them as they get better. Medical care or a lot of professional outputs. And then the third factor that I think is extremely important is kind of elasticity of labor supply. So if there's an increase in labor demand for medical doctors, or they become more productive, so people are more willing to go to the doctor, I can't just read about it in the newspaper and say, Oh, great, all this demand for doctors, I think I'll start being a doctor tomorrow. Because it takes years and years to become one. So, productivity increases in those occupations generally will translate into wage gains, because you won't have very rapid numerical increases in supply. Russ: But it's not just the time. It's the fact that not everybody is capable of doing it. Which is I think also very important. Guest: True. That's right. Russ: Which I think means those wages will be [?] Guest: You could argue the supply is even less elastic because it's not just how long you have to go to school, but who is suitable for that work. Now, let's take these three points and bring them over the last side, the low-education side of the labor market. And say, we have, you know, automation, how is this affecting employment in, say, housekeeping? Well, gosh, there isn't really that much substitution of machinery for housekeepers. But there isn't that much complementarity, either. You could imagine that technology would somehow increase the productivity, the amount of housecleaning someone could do per hour. But it's hard to see where that actually happens. Russ: The vacuum cleaner is already invented. So that was good. That was complementary. Guest: That's right. That was complementary. Absolutely. Russ: But it's over. Guest: Exactly. The second point: well, what about the elasticity of demand? Well, it turns out that actually, there's not a lot of evidence that those personal services are very price-elastic. But they are elastic to overall societal wealth. So, when income goes up, people spend more on those types of those things. So economic growth can certainly be beneficial for those type of activities. But now let's imagine, take the best case scenario: so we have, let's say there's some productivity increase, and then there's economic growth so demand for these personal services rises. Well, what happens? Well, supply of labor to those activities is potentially very elastic because there's almost no barrier to entry. People can do them and be productive really rapidly, and they don't need specialized skills or training. That means that it's hard for wages to rise in those activities quickly, because labor supply will tend to dampen that rapidly, especially when people are being kind of displaced from middle-skill jobs. So, we have seen growth in personal services, but only in the 1990s when labor markets were extremely tight. And otherwise, even though employment growth has been rather polarized with growth at the top and growth at the bottom, wage growth in the 2000s and in the 1980s was not polarized. It was rising more at the top and falling more at the bottom. So, the point I make in the paper is it's easy to understand how these technological changes affect the shape of employment growth--what activities are demanding more and less labor. The actual implications for who earns what are mediated through other general equilibrium forces that, you know, tend to benefit the high skilled and don't seem to be nearly as beneficial for low-skilled workers or low-skilled occupations--low-education occupations, low education workers, even as numerical employment in those activities rises. |
| 38:20 | Russ: So, that's kind of what we'd expect. Right? The idea would be--you go back to earlier times of technological change. When the car comes along, blacksmiths don't make so much money any more. Like, zero, all of a sudden. And they have to turn to something else. And for many of them it's late in life; it's hard to tool up. What we'd hope would happen, in the current scenario, and we've seen a little of it, is that it is the high-end jobs, the jobs that require more education are paying a lot more that would draw people into high end skill acquisition. It's been somewhat surprising though, as you point out in the paper, that there hasn't been more of that. And I want to add two things to that and then let you react. One is: there's a sort of--it's all well and good to say we need more STEM (Science, Technology, Engineering, Mathematics)-trained people. But maybe people have limits in how many people can do STEM stuff. And so, this problem is just perhaps going to "get worse." It's not obvious that--and then, people who do go on to college don't always study those things. Either they can't or they don't like them or they are not thinking about it enough, or don't want to think about it; maybe they shouldn't. But it seems to me that this application--so, I'm trying to summarize what you are saying here and then I'll let you react. That at the high end, we have a pretty healthy situation. And we see that high-end skills situation, we see that--the unemployment rate's very low, the wage growth is healthier. The bottom 75%, or whatever the number is--maybe 25%--but somewhere in that 25-75% range, it's not going so well. And it's not obvious that that's going to change through natural responses. That I think is the worry. Guest: Yeah. I think that it is the supply response to the rising return to education in the United States over the last 35 years, really. It started rising in 1980. And it's basically risen almost continuously to the present day--it hasn't grown as much in the last 5 or 8 years. But it's applied[?] to an extremely high level. The supply response has been surprisingly slow and weak. It's been much, much stronger among women than among men. And you know, women vastly outnumber men in college education at this point. And also are outnumbering them in the professions. And this has been a big puzzle. I don't think anyone knows the answer for why that is true. We did actually see, in the 2000s, we did see an increase in high school completion rates, the first time we've seen that in decades. And an increase in college attainment. So, it's not that the message isn't getting through, but it's getting through very, very slowly. Now, I think part of the reason--this goes back to your question of why aren't more people doing STEM--it's now pretty clearly documented that your college major matters a huge amount for your subsequent earnings, and STEM workers do earn more. And sociology and clinical psychology majors earn a lot less. And that information is known: why aren't people switching? And I think one reason--and partly it's just taste; some people just really don't like doing those things. Part of it also is poor preparation. The United States has a very weak STEM education in secondary, at the high school level, and even the middle school level. And so we're way behind other countries. And so that makes it much, much harder for people to enter those activities. Even when you teach--we have Ph.D. students in economics, [?] teaching my Ph.D. students in economics at MIT, but the people who go into the fields of theory, sort of extremely mathematically intensive, they are basically--many, most of them are from other countries, and they have very deep math backgrounds by the time they are through high school. And there are very few U.S. students. So, if we get a Liberal Arts student who enters MIT's economics program, it's very unlikely they are going to become an economics theorist. Now, I'm not saying that's a great loss to the world or to them; but that road is already blocked to them because they didn't have the foundational preparation when they were younger. So, they'll do other things in economics, very useful things. I'm not a theorist; I'm not sad about that. My point simply being that the foundational skills in STEM, particularly in the mathematics and analytical training, need to come pretty early. So I think many U.S. students are not prepared to enter the fields that in fact would be more remunerative as a result of shortcomings in our primary and secondary education system. Russ: We've talked a lot about that recently, and we'll continue to talk about it. That's another issue, obviously. Russ: So, but is correct to say--I want to make sure people understand the underlying economics here. Is it correct to summarize what you've been saying as follows: At the upper ends of the skill distribution, technology complements--that means, enhances--one's productivity. At the lower end, not so much. So that to some extent the wage effect is not going to be as large. And is, therefore, the middle being reduced dramatically, the employment opportunities of the so-called middle--and in the wage distribution? Is the wage distribution polarized? Is it bi-modal in any way? Guest: No, it's not as bimodal because wage growth has been so weak at the bottom. So that was the point I was making earlier, that the sort of the decline to the middle, sort of cascade downward, because people who are in these middle skill activities can easily move into personal services. Right? So if they are displaced, they are going to put pressure on wages in lower wage activities as well. So, even though the employment is fairly polarized, wages were polarized in the 1990s but otherwise pretty much it's been, looks like a downward escalator. But I guess--so, a couple of points on this that I want to emphasize. I think that one should not assume that polarization, even of employment, will go on forever--that the middle will just collapse to zero. I think that that's--it's always dangerous to just take the current trend and just forecast, extrapolate linearly to the vanishing point. That's not likely to happen. A lot of the possibilities for that substitution may already have occurred. And when you look at what's left, "in the middle," actually the jobs become more skilled again. So, we have many fewer typists and filing clerks than we used to, but the people who are clerical workers have more skilled jobs than they used to. They are people who organize travel and work out logistics and deal with hard problems--like, how to get reimbursed. And, you can also--a nice place to look is in the medical world. So, there are lots of medical technician jobs, some of which don't require a college degree. But they virtuously combine a set of technical skills, these sort of fluid intelligence skills, so being a nurse, being an x-ray tech, being a phlebotomist. And those things pay pretty well and arguably will be growing. I think partly because of the aging of the population. I think there will be--there certainly are going to be highly paid, good career jobs in medical, in medical technical jobs, in the skilled trades, like for example construction or electricity or plumbing; in skilled repair. And sometimes in this maniacal focus on college for all, we've sort of forgotten that there is a whole set of skilled vocations that again I think are complemented in the sense that they combine expertise and technical knowledge with these very-difficult-to-substitute human capacities. Right? The complementarity is there. |
| 46:34 | Russ: So, let me bring us back to our earlier discussion. I can easily imagine a world where a robot--and I know people who literally work on this now--where a robot would give me--I'm in the hospital, God forbid, and I need post-op (post-operative); first of all the operation, clearly technology right now, has incredible complementarity with surgeons using robotic devices. But we could imagine a world where I don't even have a surgeon. The robot takes up my kidney or whatever it is--it's imaginable. But this is much more imaginable: in the post-op, an arm, a robot arm gives me my, dispenses the right amount of pill for me that I need. Maybe not just dispenses the right kind of pill but knows my history and does some diagnosis of me and knows I need a different dose than the person one bed over. It does a whole bunch of things that a human being could do. What it can't do, at least right now, is make me feel better and show empathy. That ability remains a human, I think will remain, a human thing. But that skill--the question, those high paying nursing jobs--they might be gone in 20 years, most of them. Guest: Sure. So if those things are not complementary, then if we just have people who are paid empaths, who have no medical knowledge and don't need any, those aren't going to be highly paid jobs. There has to be a skill that they possess that is genuinely scarce. Russ: Well, I don't know about that. Being empathetic is pretty scarce. Doing it well--doing it well. Guest: Perhaps. Okay. Look, there's lots of people--okay, let's leave that alone. I do not foresee a time anywhere in the near or even relatively distant future where all the skilled activities are done by machinery and what's left for people to do is sit around and emote. I think there's a lot--in medicine there's a huge amount of skill in diagnosis and figuring out what someone's actual problem is. And it's more than just a chemical problem. It's a sort of set of complementary activities that allow that person to recover. Actually, my sister-in-law is a person who goes and helps elderly and infirm to sort of figure out a workable life. Russ: That's huge. Guest: She's a highly, highly trained nurse. But she's also a person who is a problem solver. I think there's lots and lots--I think in general people are complemented much more than they recognize by the things that make us more productive. In a variety of ways. My sister-in-law is complemented by her modern automobile that gets her reliably from one person's house to another. Even though, of course it means there's in theory fewer people like her needed in the course of a day to reach a given number of people. Russ: Which is a good thing, because it means it's cheaper and people can afford it; and there's a very elastic demand for those services. Guest: Exactly. It's difficult--I think the challenge is, though it's a challenge for our imagination, maybe, because we're being too optimistic, try to figure out: What will those new activities look like? So I give an example in the paper: at the turn of the 20th century something like 38% of all U.S. employment was in agriculture. A hundred years later, 2% of all employment was in agriculture. I think if you'd asked farmers at the turn of the 20th century, 'What do you think everyone will be doing a hundred years from now?' and especially if you told them, 'And by the way, only 2% of the people will be in farming'--they would have known--they knew at that time that farming was declining. In fact that was the genesis of the high school movement in the United States, to send everyone to high school, because they recognized the future was off the farm. But they wouldn't have been able to say, Oh, I think it will be software, health services, business services, entertainment, hotels, the movie industry, video games. That would have been impossible to predict. And similarly we find ourselves at a point where we've gotten a lot faster, a lot better at automating a lot of things that we thought were very readily automatable. I mean, just to give you one personal example, I remember some 25 years ago I was working as a temp at GTE[?] in their library, and I was working as a kind of librarian assistant. And a guy came up to me and he started telling me about this thing he was doing on his computer; and he was going to have the computer search, help you find articles that you needed. And I said, 'How are you going to do that?' And he said, 'We'll just read through the abstract.' And I said, 'But it doesn't understand language.' And he goes, 'Oh, no, no, no. It'll just recognize key words.' And I was like, 'Oh, yeah, good luck with that.' So I couldn't have been more wrong. But the point was, we have gotten good at things that we thought were exclusively human. We've learned how to automate them. So we're at a period where all of a sudden we see the frontier advancing very quickly, what we can automate, but we don't know what is going to replace it. And we sort of feel, is this time different? Is this the time in which all of a sudden we just run out of things for people to do? My own view is, no, we won't. On the other hand, it certainly is disruptive, and certainly for people who don't have some complementary skills, it's bad news. So if you took the workforce of the turn of the 20th century and brought them to the 21st century, many of them would be unemployable, because they would be innumerate and a substantial fraction would be illiterate as well. So you have to have skills that are complemented by a sort of modern set of demands, and many of those skills are a combination of our onboard equipment--i.e., allow you to work as a housecleaner--plus a set of scarcer skills that we want to combine that allow us to add more value. |
| 52:34 | Russ: Well, to take a recent episode that we have a lot of knowledge of, when outsourcing--that is, sending jobs overseas, sending tasks overseas--started to grow dramatically, a lot of really smart people said, 'This is different from the usual gains from trade, and it's going to have enormous impacts on U.S. wellbeing and workers' wellbeing.' And there were, as you point out, there were some workers who had a negative impact from it. But people grossly overestimated that trend. They extrapolated it way too quickly. They neglected the gains from keeping stuff close to home, that distance still matters. And I think the punditry and lots of well-trained economists overreacted, listening to them, overreacted to those fears. I think the issue here is, as you say, whether this is different. And I think what makes your--talk about this issue that your colleagues, Brynjolfsson and McAfee argue we should race with the machine. That is, as you've pointed out I think very elegantly in lots of analysis in our conversation and in the paper, there are a lot of ways that technology helps us, makes us more productive, and creates new opportunities that we can't, as you say, imagine. The question is, other than that nice platitude, 'race with the machine,' which sounds nice, instead of racing against the machine like John Henry, what is it that actually mean? My suspicion is: There's no way to forecast that. Obviously those 19th century, turn of the 20th century workers, wouldn't be very useful now. We do need a better school system and we do need ways for people to be interactive with machines that only a small fraction right now can do. But having said that, I think there is grounds for optimism in just human creativity in coping with it. So, I'm not as worried as the pessimists, but I do think making our school system more flexible would help a lot. Guest: I agree. I don't mean to be Pollyanna-ish here. I find myself in an ironic position, in that I've been arguing for the last 15 years that computerization has had a very large effect on the labor market, and I've sort of been in some ways out in front of that argument. I'm now telling people not to panic. Because I think they are missing, or--they have [?] convinced: Computers are substituting for people. And they sort of have forgotten the second half of that: that that means they are complementing us in other ways. But absolutely, I think the human skill that is most complemented by automation is flexibility. And the thing that makes us flexible isn't just muscular flexibility. It's problem solving, mental acuity, the ability to apply fluid intelligence to apply to unforeseen situations, whether they are personal interactions or math proofs or scientific hypothesis formation or persuasion. And so the thing that we--when people say, 'Should my kids be spending all their time studying Java?' I'm like, 'No.' I think they should learn very strong math foundations. They should also learn how to write. To speak effectively, and to work in a team. And those skills are very broadly applicable. And one of the great advantages of people is that they are adaptable, and they are able to reinvent themselves, because they have the foundational skills that allow them to do that. And so in a time of change, being adaptable is a valuable fundamental skill. Russ: Yeah. I used to argue that that's why you should go to college and get a general set of skills. I now think college is overrated. But the parts of college that people ought to be focusing on are math, communication. It's not all STEM. STEM helps, but then there's communication; there's, as you say, problem solving; creativity. These are all things that can be improved through thinking about them a little bit, anyway. Guest: Yeah. Well at MIT, we're always telling people to learn how to write. Everybody can solve a differential equation, but they can't write a paragraph telling you how they did it. Russ: Yeah. Neither could Andrew Wiles. |
| 56:52 | Russ: I want to close with an example that I thought was a really interesting example, then I'll give you one more chance to finish up. Talk about the company that you identify, which is Kiva, because it's an interesting little case study of how technology and humans interact. And it's not to be confused with kiva.org, which is a very interesting philanthropy opportunity that we've talked about on the program before. But this is a warehousing, inventory company. Talk about it. Guest: That's right. So, Kiva, I think is a beautiful example, an iconic example, of environmental control--how do we deal with machines that are inflexible, and make the environment predictable so it doesn't demand flexibility. So the basic problem is companies like Amazon, which now owns Kiva, although it was actually an MIT-based startup, and many other large warehouse companies do a lot of direct consumer sales. They have these massive warehouses--a warehouse will be multi-million square feet; it's too expensive to air condition them. And you have stuff all over the place, a huge variety and number of objects. And so, traditionally, and it sounds sort of funny to say 'traditionally' when you are talking about Amazon--but historically, Amazon employed so-called pickers, basically people who were athletic young people who would run through the warehouses with little computers on their wrists that would tell them where to look for an object; and they would climb up over a shelf; they would grab, they would look at the object and make sure it was the one they wanted; they would grab it and they would bring it up to the front and it would be boxed up. Those people--it's a difficult job, involves running or walking 10-15 miles a day. It's hot and sweaty. And there's no robotic substitute for those people. There's no cost-effective machine that you could employ that would do that job. So what Kiva does--they're idea is, let's reorganize the warehouse in such a way that we vastly reduce the need for a human touch. In fact, Kiva talks about the idea of increasing the value of a human touch. The way this is done is, the warehouse is run basically by the Kiva software. And when goods come in the loading dock, the software is told: here are the objects that are coming in. There are then people who take those objects off the pallets and they put them onto shelves. But they are not ordinary shelves. They are shelves that are driven by little robotic drives. In fact the robots look like those old canister Hoover vacuum cleaners. They drive along the floor, they go under a set of shelves, and then they raise themselves up a few inches and then drive the shelves with them. That's all they do. It's not a very complicated robotic feat. The shelves come, and laser pointers controlled by the scheduling software tell the people, point and say, 'Take this object; put it on this shelf.' So now, the computer knows where the objects are, not because it can see them or recognize them, but because it's told you where to put them. Then the robots whisk the shelves away into the warehouse--or, I should say, not the robots organize them--the controlling software optimizes the warehouse according to the flow of goods. So it's going to put things that are used frequently near the front; things that are used infrequently near the back; things that are ordered together, they are going to put them together. Then as orders come in, the robot drives again go. Let's say I order from Amazon a book, a box of diapers, and a video game. And the robots then go collect the shelves that contain those objects. And they line up for another human picker. And that person stands there, and the shelves drive up to the picker. A laser pointer on the ceiling points to the object that the picker is supposed to pick. The picker takes it off the shelf, puts it in a box, tapes[?] these three objects together, sticks a label on it and sends it off. And then the robots whisk the shelves back to the warehouse. So there are only two points in the system where people are involved in physically handling objects: when they are put on the shelf and when they are taken off the shelf. All the rest of the kind of transportation and sorting of objects is done by the robots. But the reason they are able to do this is because the need for all that dexterity, flexibility, visual recognition has all been pared down to these two points of contact: the loading and the unloading. Everything else is transportation. And so--and if you talk even, basically people say, when are you going to put hands, these robotic hands, that will do what the workers do. And you say, 'That's not cost effective.' Making sensory robotic arms that will deal with all these robotic objects and all their inconsistencies is really, really expensive. That's what people are for. What we do is we take almost all the rest of the routine activity and we mechanize that. And so, what you see there is, one, this environment control occurring, where you've reshaped the environment to make it very consistent and predictable. And, two, this sort of redivision of labor according to comparative advantage. The comparative advantage of machinery is, you know, moving heavy objects in a hot, confined space. The comparative advantage of humans is recognizing objects and handling their inconsistencies in a delicate way. And that is the division of labor that you see occurring in that warehouse. |
| 1:02:10 | Russ: Yeah. It's an extraordinary story. I think--I read recently, talking about something like driverless cars delivering items, and some workers, combining with Uber, say; and some workers saying, 'You'll always need somebody to walk up to the door.' I'm thinking, 'No, they won't.' They'll put a box on the street; and the arm of the drone or the driverless car will push the item into the box and it will know what the shape of the item is, so whether it will fit in the box or which door to use. And we're going to make progress on all those things. Guest: Sure. Russ: So, I guess to some extent, the optimist and for the human enterprise--and by the way, I'm not worried about, maybe I should be, about machines getting so smart that they'll run--that they'll take over. Guest: I'm not worried about that, either. Russ: Which is a whole 'nother issue. But I think the complementarity--the ultimate types of complementarity--we haven't imagined yet. And [?] it would be like to live in a world where stuff is incredibly low-cost, and there's lots of leisure. And therefore we are not going to use that leisure to surf the Internet or go hiking all day or go into the national parks. We're going to use that leisure in ways you and I can't imagine. Just like those farmers back in 1900--or at least that my last [?] where I'm at now. Why don't you close us out and talk about your optimism or pessimism for the future and what policies you'd like to see to help us get to the more optimistic view? Guest: Sure. So, on the leisure side--it is the case we consume more leisure than we used to. Although, as Bob Gordon points out in his recent working papers, the increase in leisure was much, much more concentrated in the first half of the 20th century than in the second half. Russ: That's because we take leisure on the job. That's my theory. And we know that. We know that people spend less time working when they are at work. For most occupations. Not the pickers--that's another contrast. But go ahead. Sorry. Guest: Right. But I think also it won't be just more leisure. It will also we will be doing other types of work that we didn't imagine, that just didn't exist. So, we will be spending our time doing activities that you wouldn't have thought of as--so, whether those are making movies or writing software--these are all things that we do know about but that we might not have anticipated them. I think the case for optimism, which I don't want to make too strongly, is that productivity ultimately benefits--it accrues to humans. The machines don't get rich when they become more productive. We do. So the problem it creates primarily is not one of lack of wealth. It's one of income distribution. And that's the one we face so much of--it's because this is very disequalizing, it's not that there isn't wealth being created; it's that the wealth is not--effectively diminishment in labor scarcity among some skill groups means that their claim on the wealth is very limited. Right? So that the concern-- Russ: Well, their shares might fall. But the question is whether their overall well-being is rising. Still, there might be a cultural issue of gaps. But it's a different issue if you say they are going to be emiserated--they are going to be literally worse off. Guest: Sure. Russ: If they have lower wages-- Guest: Well, that's a [?] Russ: If they have higher real wages--as long as their real wages rise, as long as--that barber, a barber today has a much bigger command over goods and services than a barber in 1850, even though the technology is almost exactly the same. Guest: Sure. Absolutely. On the other hand, in the last 30 years, if we look at wages of people with high school and lower education, they've arguably fallen in real terms. Russ: Absolutely. Guest: So, I think the challenges are significant, because of the distributional consequences, not because of the prospects of, you know, lack of progress. In fact the progress is--that's what we're worrying about at some level. And I also think the challenge is that our system, your[?] capitalist system is organized around labor scarcity. Your most valuable asset is your scarce labor. And if actually labor does become less scarce, we [?] large groups of workers, or people, how do we organize society in a way that people still have purpose and structure that gives meaning to what they do? Russ: You want to talk about policies? Guest: Policies. So, well, as you and I have discussed at length during this conversation, things that improve the primary and secondary education and the efficiency of post-secondary education, college education, I think are really, really good. As are things that improve the conduits into the skilled vocations. The other thing I think is actually extremely important is policies that make work more attractive for less educated workers. We see huge declines in labor force participation among demographic groups that have falling real wage levels. I think programs like the EITC (Earned Income Tax Credit) are a wonderful thing and I think they ought to be expanded to include males without dependent children. Russ: You are talking about the Earned Income Tax Credit. Guest: That's correct. And right now you can get up to approximately $5000 if you are a woman with dependent children. But our biggest employment, in the annual wage subsidy, our biggest employment problems are actually among kind of young and middle-aged men who are not married; and although they may have children, those children may not be dependents in the standard sense in that they can't claim them on their tax forms. They can, like, get about $400 a year. So the Earned Income Tax Credit is a policy that has been effective in increasing employment, increasing participation. And I think it's a good investment and it ought to be made more broadly accessible. So that's a policy I would strongly advocate. Russ: But ultimately a lot of this is going to come from our culture changing, about how we see what university life is, we see high school life is; and I think that's going to be the most interesting thing to watch: that no one controls what will emerge in ways that we can't expect. Guest: We're always surprised. And, as economists we don't really believe in central planning. So we should hope that the innovation will be something we didn't anticipate and couldn't have dictated. |
READER COMMENTS
Glenn Mercer
Oct 6 2014 at 11:54am
Fascinating podcast, thanks.
On the Google car example (and to make a futile attempt to head off those who will accuse me of being a Luddite, let me state that I am a big fan of autonomous vehicles, and see their arrival as virtually inevitable) let me add color: on the one hand auto collisions kill roughly 33,000 or more people in the USA every year, which is roughly 33,000 too many, and is a number which must be reduced. On the other hand, given that we drive some 3 trillion miles annually, that is about one fatality for every 90 million miles. Implying that, even though many drivers have not seen the inside of a licensing bureau or driver’s school in decades, humans’ tacit skills here are pretty good! Will autonomous vehicles be able to do a better job eventually? No doubt. I am only pointing out how high the safety bar has already been set, by humans working together with ever-more-advanced vehicles. And this is a moving target: as ABS, ESC, lane-departure-warning systems, advanced braking, active cruise control, drowsiness alarms, and all the other technologies out there gradually penetrate the fleet, by the time fully autonomous vehicles arrive they may have to be beating 1 death per 150,000,000 miles. Not that they won’t or can’t! (Though I do dread the first time an autonomous vehicle kills someone, and we have to read the inevitable headline “FrankenCar Kills Family of Four!” Somehow death-by-robot is more unsettling than death-by-human, at least for some people.)
rhhardin
Oct 6 2014 at 4:02pm
Stanley Cavell in _The Claim of Reason_ has the ultimate what is a chair illustration. The whole book ties all this to Wittgenstein, who without this explication is generally not understood.
One conclusion is that we don’t know what we’re doing when we imagine machine intelligence. It’s a bunch of pictures and partial stories we tell ourselves, the analysis of which would make the problem go away, at least the philosophical problem.
Cavell on chair starts here here.
MarkOS
Oct 6 2014 at 6:03pm
Cards on the table, I came to Econtalk in order to avoid an echo chamber of my generally socialist opinions. Over the last three years I have grown to respect Russ for his interview technique, broad range of guests and un-ashamed championing of high quality education for all.
Russ often equates economics and history, this is the second of recent podcasts that I feel could do with a historical perspective. I am not qualified to do so but what the hey this the internet!
Recently we had a discussion about what is the problem with a tiny elite owning the majority of the wealth?
Errr – America might still be run by the Tribal Nations if it wasn’t for a hereditary elite across europe using their money and power to throw people off land, pass laws outlawing generations old modes of income generation, and generally abusing the poor. Or that hand in glove with the land owners an equally small, and wealthy, religious elite controlled the gates to paradise. Thus providing america with a bounty of honest, hard working inventive types.
This week we have the fact that ‘not everyone can be a doctor’
In europe the term scholar was for many years used in the phrase ‘A Gentleman and a Scholar’ the fact being that only the landed gentry had the time, opportunity and money to pursue academic avenues. That this education took place in latin and greek whilst what little common education that existed was in the vernacular further excluded the common folk from prestige employment. In those days only the children of the elite could become a doctor or lawyer. Which is what made them elite professions, not due to them requiring an elite skill set.
In the 1700’s It took at least seven years to become a master mason whilst it was possible to get your classics degree from most universities in two years, a pre-requisite for studying medicine. “I might only be guessing how to treat you but I am doing so in grammatically perfect Latin.”
A master mason was also an architect who used geometry and physics on a daily basis in a way that required a highly refined working knowledge of those fields that was both theoretical and highly practical.
Fast forward to today, your guest allows that these very same prestige jobs are only affected by tech in pushing up their wages. Whilst the modern equivalent of mason is finding it hard to attract trainees. I wonder whose children are training for these “top” professions?
America, a nation built by carpenters whose invention and use of the framing square was a mathematical wonder. In the home of human treasures, from Ben Franklin to Bill Gates, that have lifted the yolk from the poor across the whole world (and for the moment ignoring your civil rights record). Your Law schools are full and your technical colleges are empty.
Everyone knows that when injured it is time and first aid that save more lives than operations by elite surgeons. Henry Ford has probably saved many millions of lives around the world by enabling people to access aid quickly. Yet you still appear to give more kudos to those professions chosen centuries ago by bewigged european inbreds.
Imagine what might happen if you chose to choose your own prestige professions.
You lot ought to have a revolution!
Amren Miller
Oct 6 2014 at 9:25pm
I was confused at first, I googled polyanis paradox and was shown to some articles about the limits of markets. I thought econtalk was finally getting a little edgy! Maybe someday…..I still love the podcast though.
Russ Roberts
Oct 6 2014 at 10:38pm
Amren Miller,
It appears that David Autor is speaking of a different paradox stemming from Polanyi’s work than is sometimes described. On the issue of edginess and discussing the limits of markets—if you did not find the recent episodes with Piketty and Nussbaum edgy then I fear you may be destined to always be disappointed with the edginess of EconTalk…
Doug Tree
Oct 7 2014 at 2:47am
—
On the issue of edginess and discussing the limits of markets—if you did not find the recent episodes with Piketty and Nussbaum edgy then I fear you may be destined to always be disappointed with the edginess of EconTalk…
—
see also: Jeffrey Sachs
Perhaps there should be an ‘edgy’ category for sorting the podcasts!
SaveyourSelf
Oct 7 2014 at 11:21am
David Autor was a great interview. He said:
46 “I do not foresee a time anywhere in the near or even relatively distant future where all the skilled activities are done by machinery and what’s left for people to do is sit around and emote.”
46 “So we’re at a period where all of a sudden we see the frontier advancing very quickly, what we can automate, but we don’t know what is going to replace it.” 1:08 “We’re always surprised.”
46 “And we sort of feel, is this time different? Is this the time in which all of a sudden we just run out of things for people to do?”
1:04 “On the other hand over the last 30 years when we look at the wages of people with high school or lower education, they arguably have fallen in real terms.”
1:06 “The other thing that I think is extremely important is policies that make work more attractive for less educated workers. We see huge declines in labor force participation among demographic groups that have falling real wage levels. I think programs like the EITC are wonderful things and I think they should be expanded…”
“Our biggest employment problems are actually among young and middle aged men who are not married, although they may have children, those children may not be dependents in the standard sense and they can’t claim them on their tax forms. The earned income tax credit has been a policy that has been effective at increasing employment and increasing participation and I think it’s a good investment.”
Martin Brock
Oct 7 2014 at 3:13pm
This is now my second favorite EconTalk episode after Gordon on Ants. It was fascinating all around. I don’t usually leave kudos comments, but this one merits an exception.
Mort Dubois
Oct 7 2014 at 9:33pm
Re driverless cars: I think that the people trying to design driverless cars face not only the difficulty of helping a machine to understand the physical environment, but also the social environment on the road. All of the existing cars have drivers, drivers are human, humans are incredibly good at quickly switching between following rules and breaking them depending on which choice leads to immediate advantage. We all know that the act of driving in traffic is a continuous negotiation with the drivers surrounding you. Sure, there are rules, but most of us are willing to break many of them at almost any time. How do you program a driverless car to understand this? I don’t see how it can be done. The consequence will be that the first driverless cars on the road will have very slow average speeds, as they will lose every negotiation for momentary advantage on the road. It’s hard to imagine how it could be otherwise – would you program a driverless car to ignore a jaywalking pedestrian or someone cutting into their lane? Even if driverless, rule-following vehicles became a majority on the road, they would still be vulnerable to aggressive behavior from a human driven car. A human guided vehicle can simply force their way into traffic, because the driverless car must be programed to avoid an accident if possible.
Maybe there’s a way around this, possibly if all vehicles switched to driverless guidance at the same time. That won’t happen. It will be interesting to see how whether there is a technological fix to this problem. I suspect that it doesn’t exist.
David Thalheimer
Oct 8 2014 at 7:43pm
To Mort Dubois:
I agree with you that driverless cars will be at the mercy of aggressive human drivers, which will make people less likely to want to use them, especially when in a hurry.
However, there are possible regulatory and technological solutions to this. First, a driverless car is already replete with sensors, so it would not be difficult at all to program them to document the aggressive and illegal activity of other drivers. Think of them as potential mobile automated traffic cops that can take photos, video, speed measurements and make all kinds of other data available to police authorities should they be legally permitted to use it. We currently have apps that have some traffic-related reporting features, but nothing on the scale that automated vehicles would be able to deliver. They might even be networked to assist police in locating escaping suspects.
Furthermore, it would also be possible to set aside lanes, such as existing carpool lanes, for autonomous vehicles only. Once in their own lane (or road) they would be able to travel at higher speeds more safely than human operated cars.
So, I think we will be surprised at just how revolutionary these machines will be. For a laugh, here is an interesting related link:
https://earthcrazy.wordpress.com/2014/09/11/my-cars-a-bitch/
amdisappoint
Oct 9 2014 at 1:09am
[Comment removed for supplying false email address and for crude language. Email the webmaster@econlib.org to request editing your comment. We’d be happy to publish your comment. A valid email address is nevertheless required to post comments on EconLog and EconTalk.–Econlib Ed.]
Avram
Oct 9 2014 at 12:30pm
The explanation given to the task of Machine Learning is correct but only to a single type of machine learning. The example given, Google lab learning to identify cats, is however not of the type described. There is supervised and un-supervised learning. Supervised is the most intuitive, the learning algorithm is given some “positive” and “negative” example (e.g. pictures that either do or don’t contain a human face) and is also told which is which beforehand (each picture is annotated). This however is not what was done by Google.
Unsupervised learning is the task of inferring from unlabeled data. This is what Google did. They build a large Artificial Neural Network (ANN) and fed it unlabeled images extracted from YouTube. A large part of the paper talks about how to know what the network “Learned” (has strong response to). Interesting, YouTube has lots of people and cats. http://arxiv.org/pdf/1112.6209v5.pdf
Research has shown humans actually learn much more in an unsupervised manner. Correcting a kid’s grammar (reinforcement learning) is less important than simply exposing the kid to correct grammar.
Perhaps Andrew Y. Ng would be a good guest for econtalk. Besides being an authority on Machine learning and ANN he’s also the Chairman and Co-founder of Coursera. His Mooc has been viewed by over 100,000 students (according to coursera).
If we’re on the topic of being displaced by robots I would recommend having Andrew Davison on the show. He leads the Imperial College robotics lab (Mostly computer vision) and is one of many very serious scientists that believe in super human AI (singularity). A discussion with him about what jobs may become absolute to humans will be very different.
He’s also posted on occasion about “Positive Money” and reform of all current monetary systems. Maybe a worthy topic as well
rhhardin
Oct 9 2014 at 5:50pm
Kids learn language by learning to disassemble and reassemble cliches, rather than learning words.
To motivate that they need a reason to do it, some immersion in the real world.
The detailed reassembly rules get eventually refined by taught grammar, but in the meantime you get
Child: My teacher holded the baby rabbits and we patted them.
Adult: Did you say your teacher held the baby rabbits?
Child: Yes.
Adult: What did you say she did?
Child: She holded the baby rabbits and we patted them.
Adult: Did you say she held them tightly?
Child: No, she holded them loosely.
The essentials first.
Shawn Barnhart
Oct 10 2014 at 7:53am
I enjoy these Econtalk episodes about the role of technology and changing human labor. Just learning about Polanyi’s Paradox and tacit knowledge was worth the listen and very thought-provoking.
In these discussions there’s a lot of concern about the role of technology in eliminating jobs and it often seems to be couched in terms of “is this a bad thing or a good thing” as if we should choose between the new technology and the old technology and the jobs associated with it.
Rendering classes of employment and the employees redundant with technology is bad not because blacksmithing or file-clerking is somehow transcendentally valuable to human culture but because our economy is so completely merciless to these displaced workers — lifetimes spent developing skills and careers are now worthless and our economy treats these people as worthless, pushing them from some kind of prosperity into marginal existences if not poverty.
In some ways, automation seems like a zero-sum game where the increased value from technology for complementary jobs (ie, the accountant who can now manage more information due to a computer) comes from displacing the worker whose work is automated (the book keeper who assisted the accountant).
I think there needs to be more focus and discussion on mitigating the impacts on displaced labor rather than just saying its bad. The general opinion seems to be there will be more displacement and not figuring out how to mitigate it in some way will have significant economic and political impacts. Mass un- and underemployment of displaced workers doesn’t help a consumer-centric economy. Politically, mass unemployment has always been a fertile political ground for despotic political systems.
SaveyourSelf
Oct 10 2014 at 4:55pm
@Shawn Barnhart
“Rendering classes of employment and the employees redundant with technology is bad not because blacksmithing or file-clerking is somehow transcendentally valuable to human culture but because our economy is so completely merciless to these displaced workers — lifetimes spent developing skills and careers are now worthless and our economy treats these people as worthless, pushing them from some kind of prosperity into marginal existences if not poverty.”
“In some ways, automation seems like a zero-sum game”
“Mass un- and underemployment of displaced workers doesn’t help a consumer-centric economy.”
Tony Mercurio
Oct 10 2014 at 4:58pm
Structural un- and underemployment seems to be a growing problem. Autor indicated that one reason for displacement is a lack of specific training early in life. While the division of labor has been indispensable to growth, I’m not sure where it leaves the obsolete worker of today, or in my case, the generalist. For either, not having the proper training, I think, is creating an employment mismatch between willing employers and prospective employees. With rates of change ever increasing, can market equilibria be sustained long enough to avoid self-destruction?
Yes, I think so… but I do not think that there is any gold at the end of the rainbow. It is not imaginable that technology could propel us to a point of satisfactory universal prosperity. Remember that necessity is the mother of invention. If technology allowed the world as a whole eventually to enjoy an adequate standard of living, then because of its abundance the value of money would decline and relative measures would still reign supreme. Who would clean toilets for a living if the world were rich? Or perhaps the toilet cleaners would be the rich. How much money would it take for an investment banker to choose custodial arts over his own field? …
I know I’m all over the place here, I just have a lot of thoughts on these issues. One thing I know for sure is that the law of supply and demand (for any type of resource) will continue to be as predictable as the rising of the sun.
Shawn Barnhart
Oct 10 2014 at 5:35pm
I appreciate what you’re saying, but I think it tends to trivialize the reduction in standards of living. This isn’t about someone merely skipping a vacation or not buying a new car, for many people these “transient reductions” end up being catastrophic — loss of homes, cars, divorces, illness brought on by stress.
I think it also trivializes the challenges in regaining economic status through new jobs skills. Gaining skills of equivalent value can be expensive and time-consuming, something made worse by the loss of economic status. Even when the skills can be developed they’re not of meaningful value without years of experience. That’s more than just uncomfortable, especially to older workers who don’t have 10 or 20 years to get back to where they were and face a lot of age discrimination in the workplace.
While I think your strategy is sound, the reality is most people aren’t going to do it, aren’t able to do it or don’t do it right — what happens when what seems like a marginally profitable sideline or hobby turns out to be not something that scales into a salable skill? We’re kind of back to the merciless economy just discarding the surplus.
It’s kind of magical thinking to rely on “Free-Competitive-Informed-Just Markets” because they don’t exist. I have to work with the political economy I have. I write my congressman and try to vote intelligently, but changing the fundamental nature of the political economy is beyond the ability of most people, myself included.
Charlie
Oct 10 2014 at 11:39pm
I have a biased, self-interested view in this particular topic…
I’ve been an engineer for 34 years. For all that time we’ve had a “critical” shortage of engineers being educated in our system – now referred to as “STEM.” The need for more STEM graduates is promulgated in this episode, as well.
I have two points, not well supported here, to suggest there is no special shortage of STEM education.
First – the price paid. Like the episode that refutes “bees dying off” because the pollination services are not seeing excessive price growth, a review of engineer’s wages shows nominal or below nominal increases and below nominal increases with experience. The finances are saying that there is no shortage.
Second – A review of graduates 5-10 years after they get their degree places STEM at less than 50% working in a field that uses the STEM degree they earned. Based on employment, there’s not a sufficient shortage to encourage people to work in or stay in the field.
I am greatly impressed with the podcast, the style, and the topics. This has modified my personal philosophy and perspectives more than any other source in the past couple decades.
Nevertheless, the assertion that we need more STEM education to prepare ourselves for the automated future isn’t playing out in the economy around us.
SaveyourSelf
Oct 12 2014 at 6:11am
@ Shawn Barnhart
“It’s kind of magical thinking to rely on “Free-Competitive-Informed-Just Markets” because they don’t exist.”
“I have to work with the political economy I have. I write my congressman and try to vote intelligently, but changing the fundamental nature of the political economy is beyond the ability of most people, myself included.”
Changing the world for the better is a lot more work—work that begins with just trying to understand the “fundamental nature of the political economy” and its relationship to the larger economy. Manipulating it comes easily once accurate understanding exists. At least, that is my hope.
Floccina
Oct 13 2014 at 7:46pm
Like MarkOS, I am not convinced that being an MD is so complex that only a few can handle it. I would like to see medical care given to engineers do design systems that make medicine simpler and fool proof to practice.
bogwood
Oct 15 2014 at 6:02pm
3 thoughts from the eco-energy perspective.
Costs need to be evaluated on full-cycle basis so something like Kiva needs to be amortized over its full life including repairs and replacement. It looks like promising technologies like off-shore wind and Tesla cars will not make the grade. Remember Amazon is still not profitable(PE 781). Tesla is on government life support.
Do you have support the workers anyway, somewhat analogous to supporting the base load electric even with copious renewables?
What is the embedded energy in say the driverless car? The whole system? Electric trikes are less glamorous but more practical. The total vehicle weight has to be near the weight of the passenger or less. Is the tech reversible? Going back to horses is no longer possible, do not hollow out?
Ben
Oct 15 2014 at 7:30pm
[Comment removed for supplying false email address. Email the webmaster@econlib.org to request restoring your comment. We’d be happy to publish your comment. A valid email address is nevertheless required to post comments on EconLog and EconTalk.–Econlib Ed.]
bogwood
Oct 15 2014 at 8:23pm
Time to move on but I am haunted by the brief comments concerning agriculture. Extremely important,but more complicated than implied. Due to interest rate manipulation and generations of government interference no one knows the real cost(price) of a bushel of corn. Cheap? Expensive? For whom? Externalities are undervalued,loss of topsoil and the extinction of the monarch butterfly(displaced milkweed).
Grain reserves are not that high given two consecutive years of perfect weather. Corn ethanol is a stain on the career of a Nobel Prize winner at the Department of Energy. Farmers on average are losing money at the spot price so not cheap in the long run. There will be offsetting loses not obvious on the futures market but spread through out society.
But a terrific model of what technology can do, for better or worse.
Ron Crossland
Oct 21 2014 at 9:53am
Roberts and Autor both confess to the lack of understanding of what’s happening to pressures on the middle class and the diminishing wage rate. The Economist recently posted analysis that wages in the US have decreased by 4% over the past decade.
Some of this decline is likely a result of intelligent machines displacing certain types of jobs. Some of the people who lost those jobs are not capable of finding replacement jobs at similar rates of pay. Moreover, some of the enhanced job positions that require higher level skills, are permanently outside the reach of those displaced.
By definition 50% of the world’s population’s IQ is under 100. Roughly a billion people cannot read or write beyond very basic skill levels. STEM competitiveness sounds good for advanced countries, but I consult with organizations who have rafts of workers whose ability to acquire higher level skills have plateaued, both of the cognitive intelligence and emotional intelligence variety. It may feel dour or even politically incorrect to roughly equate general intelligence with job growth, but the requirements for future complex work outpace the ability of large groups of human beings. Generally speaking our mental abilities are not infinitely flexible. And enhanced education processes cannot overcome all intellectual deficits.
Foxconn, the giant Chinese electronics manufacturer, is planning to double it’s robotics over the near term. It already has one of the largest seasonal workforce dynamics in the world. It is seeking to control the costs of this seasonality through robotics. What’s telling about this example is that Foxconn is not seeking lower wage workers in nearby countries as avidly as it is seeking robotic solutions. How many of Foxconn’s displaced workers or workers in lower wage regions are likely to replace their jobs in health care, hair styling, or teaching?
I’m not a pessimist about this topic, but the displacement of farm work to machines was first about muscle not brains. It is increasing about brains, now that most of the muscle displacement has occurred. It is different this time and most arguments I hear about remaining optimistic discount this difference too much.
If you read Brynjolfsson and McAfee – their optimism is welcomed but not overly convincing. It’s easy to look at the current limits of robotics, as discussed in this interview. What’s not so easy is to look at the other ways computer intelligence is used. Which is more economically advantageous to Google today – it’s next new hire, or it’s current complement of robotic algorithms that tirelessly work 24/7 to collect, store, collate, and analyze data? And what happens when that algorithm is bolstered by just a 20% increase in computer AI ability?
You don’t need a robot that is a full replica of a human being to replace a human being’s ability to perform work.
Robin Hanson
Oct 22 2014 at 5:10pm
Some of my colleagues just pointed me to this episode, saying they got the impression that you guys thought you were disagreeing with me.
So I want to make it clear that I agree with everything I heard in the section near when my name was mentioned. I was associated with the idea that we will eventually understand the brain because “it is just chemistry”, and then Russ asked if understanding would come “via a chemical analysis”, and David said he didn’t think that would work, that you’d have to think about the brain in terms of higher level abstractions such as “information” and “symbolic processing”. And I completely agree with that. (I was an AI researcher for nine years before returning to get my Ph.D. in social science at Caltech.)
I do expect it will likely be possible to make substantial economic use of only understanding brains at low levels of organization, via the construction of “brain emulations.” I’ve been writing a book on the economic implications of such a technology, which may appear sometime in the next century.
Jonas Koblin
Mar 21 2016 at 7:07am
We created a video explaining the theory of the Future of Work by David Autor. Mr. Autor loved it and promoted it on his twitter.
Here’s the link: https://www.youtube.com/watch?v=thFlKOqHJog
Comments are closed.