Latest / Trumponomics / How Artificial Intelligence is Taking Over the Economy

Transcript

How Artificial Intelligence is Taking Over the Economy · Trumponomics

• 0:00
  Speaker 1: From self driving cars to robot powered factories. Artificial intelligence
• 0:05
  Speaker 1: is taking over significant pieces of the global economy. This
• 0:09
  Speaker 1: is great for the businesses embracing AI, but there is
• 0:12
  Speaker 1: a downside. More robots in the workforce also means more
• 0:16
  Speaker 1: people losing their jobs to computers. So how bad will
• 0:20
  Speaker 1: the robot revolution be and how will it reshape the
• 0:23
  Speaker 1: global economy? Welcome to Benchmark. I'm Scott Landman and economics
• 0:37
  Speaker 1: editor with Bloomberg News in Washington. Returning as a guest
• 0:40
  Speaker 1: co host is my colleague Chris Content. He's a reporter
• 0:44
  Speaker 1: covering the Federal Reserve and u S economy also here
• 0:47
  Speaker 1: in d C. Chris, glad to have you back, Happy
• 0:50
  Speaker 1: to be here. So this week we're talking about the
• 0:52
  Speaker 1: rise of the robots and how they will impact the economy.
• 0:56
  Speaker 1: As someone who writes about the Federal Reserve, Chris, what
• 0:59
  Speaker 1: interest you and AI? Well, I can tell you, Scott,
• 1:02
  Speaker 1: that there is a great capacity actually of artificial intelligence
• 1:06
  Speaker 1: to transform one of the biggest jobs of central banks.
• 1:11
  Speaker 1: As you know, central banks are trying to regulate economies,
• 1:15
  Speaker 1: keep them in that Goldilock zone, and they do this
• 1:18
  Speaker 1: through short term interest rates. But that tool has a
• 1:22
  Speaker 1: lag to it. It takes a good six st eighteen
• 1:25
  Speaker 1: months before a move and interest rates has an impact
• 1:29
  Speaker 1: in the real economy. So these central bankers they've got
• 1:32
  Speaker 1: to forecast what what's the economy gonna be like a
• 1:36
  Speaker 1: year from now, eighteen months from now. But they're not
• 1:40
  Speaker 1: very good at forecasting the future when it comes to
• 1:43
  Speaker 1: things like inflation, unemployment, GDP growth. So some central banks
• 1:49
  Speaker 1: are starting to turn for a little help to artificial
• 1:53
  Speaker 1: intelligence because artificial intelligence can be used quite effectively at
• 1:58
  Speaker 1: spotting patterns in past events and then using that to
• 2:03
  Speaker 1: predict the future. So it's an area it's going to
• 2:05
  Speaker 1: take a while, but it's an area of great promise
• 2:08
  Speaker 1: for economics. It's not just economics, but it's business and
• 2:13
  Speaker 1: the broader economy as a whole. Forecasting prediction, it's it's
• 2:17
  Speaker 1: a really big part of what AI is meant to do.
• 2:21
  Speaker 1: And our guest is here to talk about that today.
• 2:24
  Speaker 1: His name is Joshua Gains and he's a professor at
• 2:27
  Speaker 1: the University of Toronto's Rotman School of Management. He's one
• 2:31
  Speaker 1: of the authors of a brand new book called Prediction Machines,
• 2:35
  Speaker 1: The Simple Economics of Artificial Intelligence, just published by Harvard
• 2:39
  Speaker 1: Business Review Press. Joshua, thanks for joining us. Good to
• 2:43
  Speaker 1: be here. Thanks. So we've already sort of alluded to this,
• 2:46
  Speaker 1: but why is the book called Prediction Machines. Well, it's
• 2:50
  Speaker 1: uh called that and not some more interesting title such as, um,
• 2:55
  Speaker 1: you know you're wonderful, great new robot and how your
• 2:59
  Speaker 1: life is going to get better, simply because the recent
• 3:02
  Speaker 1: developments in artificial intelligence are not about completely replacing human
• 3:07
  Speaker 1: intelligence per se, but actually all of being about one thing,
• 3:11
  Speaker 1: and that is prediction. That is taking a whole lot
• 3:14
  Speaker 1: of information that you do have and converting it into
• 3:18
  Speaker 1: information you do not have. That's very, very different from
• 3:21
  Speaker 1: something that does all of your choices for you and
• 3:24
  Speaker 1: things like that. Uh, it really really just does prediction.
• 3:29
  Speaker 1: And when we talk about this prediction function and getting
• 3:33
  Speaker 1: all this information, can you bring us into the real
• 3:36
  Speaker 1: world here and what kinds of industries or jobs would
• 3:39
  Speaker 1: be most likely to or have the most room to
• 3:42
  Speaker 1: benefit from this kind of additional knowledge. Well, prediction is
• 3:47
  Speaker 1: something and from your business that you're in, you think
• 3:51
  Speaker 1: of it mainly about forecasting, for instance, economic variables and
• 3:55
  Speaker 1: things like that, and something of course I've worried about
• 3:58
  Speaker 1: as well. But what's really interesting about these new developments
• 4:02
  Speaker 1: is that they highlight problems that turn out to be
• 4:07
  Speaker 1: prediction problems. For instance, the whole issue of having a
• 4:11
  Speaker 1: digital image served up to you and knowing what's in
• 4:14
  Speaker 1: it is a prediction problem. Invariably, when you get an
• 4:17
  Speaker 1: image from the Internet, the label that's being attached to
• 4:20
  Speaker 1: it is the label that human would attach to it.
• 4:24
  Speaker 1: And so basically what Google are doing when you're searching
• 4:26
  Speaker 1: for a picture is predicting which pictures correspond to that label.
• 4:31
  Speaker 1: So that's a form of prediction. And it turns out
• 4:34
  Speaker 1: that that and language translation and understanding machines, understanding human speech,
• 4:40
  Speaker 1: and even self driving cars are all mainly a prediction problem,
• 4:46
  Speaker 1: and so that is where this new artificial intelligence is
• 4:50
  Speaker 1: being implemented. You wouldn't normally call that prediction problems. We
• 4:54
  Speaker 1: normally think about forecasting the weather or forecasting an earthquake
• 4:58
  Speaker 1: or something, but basically prediction is all around us. Joshua,
• 5:02
  Speaker 1: I think it's clear that this can have all sorts
• 5:04
  Speaker 1: of benefits in our lives and in the economy, but
• 5:08
  Speaker 1: I can also have some downsides, and particularly how how
• 5:11
  Speaker 1: these things are distributed throughout our society. Maybe a big
• 5:16
  Speaker 1: question mark first about the benefits. Where where do you
• 5:19
  Speaker 1: see that societally and also economically. So I think the
• 5:23
  Speaker 1: benefits come from wherever you think about where would it
• 5:28
  Speaker 1: be good to know things with greater certainty? And so
• 5:32
  Speaker 1: any kind of decisions that you're doing under uncertainty are
• 5:36
  Speaker 1: going to benefit from having better prediction so that you
• 5:40
  Speaker 1: can imagine being applied. For instance, you're trying to manage inventory.
• 5:45
  Speaker 1: If you've got better predictions regarding demand you're going to face,
• 5:48
  Speaker 1: you're going to be able to manage that inventory better.
• 5:50
  Speaker 1: You're going to be able to correctly adjust your behavior
• 5:53
  Speaker 1: to prevent shortfalls or worse, to prevent surpluses that end
• 5:59
  Speaker 1: up to over stocking your inventory. Those are the sorts
• 6:02
  Speaker 1: of places where prediction machines are going to work quite well. Uh.
• 6:07
  Speaker 1: And so basically anywhere where there's decision making being done
• 6:12
  Speaker 1: and there's uncertainty, there is room for better prediction, and
• 6:15
  Speaker 1: the machines might serve that up right, And that sounds,
• 6:18
  Speaker 1: of course like it will make our companies more efficient,
• 6:21
  Speaker 1: more productive, but also I think will disrupt how companies
• 6:26
  Speaker 1: operate and may disrupt people's lives. Do you think this
• 6:30
  Speaker 1: time is going to be any different. We've had lots
• 6:33
  Speaker 1: of technology disruption in our economy over the decades, in fact,
• 6:37
  Speaker 1: in centuries, um is the pace of change these days
• 6:41
  Speaker 1: going to be more disruptive and more problematic to adjust to.
• 6:46
  Speaker 1: I think there's a chance, as usual, whenever you've got
• 6:49
  Speaker 1: a very large and radical innovation occurring and being adopted,
• 6:55
  Speaker 1: there is the potential for disruptive change. But the hard
• 6:58
  Speaker 1: part is trying to predict exactly where that will be.
• 7:02
  Speaker 1: When we look back and we think about the rise
• 7:05
  Speaker 1: of the automobile, it's no surprise that we also point
• 7:08
  Speaker 1: to horses as being the disruptive workers in that disrupted
• 7:13
  Speaker 1: workers in that equation. When it comes to things like
• 7:17
  Speaker 1: better prediction, however, it's far subtler um. Some of the
• 7:21
  Speaker 1: discussion that's going around is saying, well, this is the
• 7:25
  Speaker 1: first Why is this time different? It's because it's the
• 7:27
  Speaker 1: first time you really have machines taking over cognitive tasks.
• 7:32
  Speaker 1: Well that's not true. We've had machines take over cognitive
• 7:35
  Speaker 1: task with computers, and you know, we seem to most
• 7:39
  Speaker 1: of us seem to be still be gainfully employed, and
• 7:42
  Speaker 1: the adjustments that took place have been worked out uh,
• 7:47
  Speaker 1: this time, you know, Okay, the computers are doing much
• 7:50
  Speaker 1: much more. They're doing much more in terms of thought. Yes,
• 7:53
  Speaker 1: But as we identify in the book, the one thing
• 7:56
  Speaker 1: that computers can't do is set goals. You still have
• 8:00
  Speaker 1: to have a human what we call human judgment, to
• 8:04
  Speaker 1: set the goals the trade offs. No prediction is going
• 8:07
  Speaker 1: to be perfect, so you have to work out what
• 8:10
  Speaker 1: how you're going to stomach errors and things like that.
• 8:13
  Speaker 1: Those are still roles for people to get into. It's
• 8:17
  Speaker 1: only where better prediction was like the final thing towards
• 8:21
  Speaker 1: getting full automation that you might see jobs actually replaced.
• 8:26
  Speaker 1: Speaking of not getting to full automation, one interesting example
• 8:31
  Speaker 1: you discussed a little bit in the book is that
• 8:34
  Speaker 1: of doctors and how their jobs might change under advances
• 8:39
  Speaker 1: in AI, in terms of they just won't have to
• 8:43
  Speaker 1: do the diagnoses as much anymore. The computer is going
• 8:46
  Speaker 1: to do it for them, and their jobs are going
• 8:48
  Speaker 1: to substantially change. Can you talk about that a bit well,
• 8:52
  Speaker 1: basically in a conference a couple of years ago. Jeff Hinton,
• 8:55
  Speaker 1: who is one of the pioneers of the new development
• 9:00
  Speaker 1: it's in AI. He now it works part time at Google.
• 9:03
  Speaker 1: He basically said to the conference, well, I think we
• 9:06
  Speaker 1: should stop training radiologists. Now, what he meant by that
• 9:10
  Speaker 1: was his view of what a radiologist does is look
• 9:14
  Speaker 1: at images and then decide, you know, is there a
• 9:18
  Speaker 1: problem or not that requires further treatment. Well, obviously, prediction
• 9:23
  Speaker 1: machines have the capability and have been demonstrated in some
• 9:26
  Speaker 1: settings to be far superior to people looking at those
• 9:29
  Speaker 1: pictures and identifying exactly what's going on with a far
• 9:34
  Speaker 1: greater degree of accuracy. So if that's your view of
• 9:36
  Speaker 1: what a radiologist does, well it sounds like curtains for them.
• 9:41
  Speaker 1: But actually radiologists have been dealing with these sorts of
• 9:44
  Speaker 1: issues for fifty years and they're quite aware of them.
• 9:47
  Speaker 1: That has technology improves their jobs change. What happens in
• 9:52
  Speaker 1: terms of radiologists is it's not some a simple choice.
• 9:56
  Speaker 1: You get a prediction of something and you know exactly
• 9:58
  Speaker 1: what to do. You're right. If that was the case,
• 10:01
  Speaker 1: then anyone could just look it up in a manual
• 10:03
  Speaker 1: and you wouldn't need a radiologist. But invariably there are
• 10:06
  Speaker 1: other factors, other criteria, and in particular the personal dimension
• 10:11
  Speaker 1: of the patient situations that will also impact on the
• 10:15
  Speaker 1: what the treatment decision actually is. So prediction is an
• 10:19
  Speaker 1: important input into that, but there are other factors going on,
• 10:23
  Speaker 1: and we're a long way off being able to automate
• 10:26
  Speaker 1: all of that. And thus far, the evidence is that
• 10:31
  Speaker 1: you make the radiologists actually better at their jobs by
• 10:34
  Speaker 1: having these prediction machines. They are able to act with
• 10:37
  Speaker 1: more certainty and therefore able to come up with more
• 10:41
  Speaker 1: confident recommendations. This allows them to save some time and
• 10:45
  Speaker 1: save some other things, and develop other skills, and may
• 10:48
  Speaker 1: change the allocation of tasks between radiologists and other medical
• 10:53
  Speaker 1: practitioners quite a bit. So I don't necessarily see it
• 10:57
  Speaker 1: is obvious that those jobs are going to be disrupted
• 11:00
  Speaker 1: as quickly as some of the engineers do. What kind
• 11:04
  Speaker 1: of jobs, Josh, would you say are the least vulnerable
• 11:08
  Speaker 1: to being replaced or even disrupted by this kind of technology.
• 11:13
  Speaker 1: Well that's a really interesting question. You know. We tend to,
• 11:17
  Speaker 1: when confronted with this thing, point to all the important
• 11:20
  Speaker 1: things that we do and how it can't be replaced
• 11:24
  Speaker 1: by a machine. For instance, some of the jobs that
• 11:28
  Speaker 1: people talk about as being hard to be replaced by
• 11:31
  Speaker 1: machines are ones that require emotional input. Interestingly, Danny Khneman
• 11:36
  Speaker 1: at a conference here just last year. He's the Nobel
• 11:41
  Speaker 1: Prize winning economist who has been responsible for behavior or
• 11:44
  Speaker 1: economics and thinking about decision making and judgment in the
• 11:49
  Speaker 1: context of psychology, and his answer was equivocal. He sees
• 11:54
  Speaker 1: humans as ultimately very flawed and doesn't see any reason
• 11:57
  Speaker 1: why machines wouldn't take over. His view was, do you
• 12:01
  Speaker 1: really want your care to be managed by disinterested children
• 12:06
  Speaker 1: when you've become elderly, or would you rather have a
• 12:09
  Speaker 1: robot who's been trained to work out exactly what your
• 12:12
  Speaker 1: needs are. So it's very hard to work out what
• 12:15
  Speaker 1: is safe and what isn't. All we can say right
• 12:18
  Speaker 1: now is that the tools in their current instantiation and
• 12:21
  Speaker 1: what we're going to see probably over the next five
• 12:23
  Speaker 1: years and maybe ten years, is all about that prediction function,
• 12:27
  Speaker 1: which leaves a lot of room open for people. Speaking
• 12:30
  Speaker 1: of jobs that are likely safe from AI disruption, I
• 12:33
  Speaker 1: think Chris will appreciate being in Washington, d C. Federal government, Congress, etcetera.
• 12:40
  Speaker 1: Those will probably be pretty safe from disruption for a
• 12:43
  Speaker 1: long time. So if we ever get some disruption that sector,
• 12:46
  Speaker 1: that would be something definitely to watch. But josh I
• 12:49
  Speaker 1: wanted to turn to another issue that you discuss when
• 12:53
  Speaker 1: when you're talking about economic impact in the book. Uh,
• 12:56
  Speaker 1: it's really intriguing that you talk about inequality and how
• 13:00
  Speaker 1: that would evolve if AI were to take hold in
• 13:06
  Speaker 1: the economy. Why could that get worse with the rise
• 13:10
  Speaker 1: of AI. Well, it's you know, as usual, it's difficult
• 13:14
  Speaker 1: to forecast these things. You never should trust economis on
• 13:18
  Speaker 1: on those sorts of big trend forecasts. But the concern
• 13:21
  Speaker 1: that we have is that certain skills become more valuable
• 13:26
  Speaker 1: valuable and other skills become less so. And one of
• 13:30
  Speaker 1: the problems is if you wanted to be the sort
• 13:33
  Speaker 1: of person who could take advantage of AI, it's invariably
• 13:37
  Speaker 1: not going to be a skill that is like routine. Necessarily,
• 13:42
  Speaker 1: once you know, once you have the sort of job
• 13:44
  Speaker 1: where you can look at a prediction and you know
• 13:46
  Speaker 1: exactly what to do, your place in that is devalued.
• 13:50
  Speaker 1: On the other hand, there are some situations where it's
• 13:53
  Speaker 1: going to take a bit of art to understand what
• 13:55
  Speaker 1: the prediction really is. We don't tend to think of
• 13:59
  Speaker 1: machine learn in computer sciences art, but there are so
• 14:02
  Speaker 1: many variables that the engineers have to control and think
• 14:06
  Speaker 1: about that. It does tend to have that quality to it,
• 14:10
  Speaker 1: and similarly, how we use those tools also has a
• 14:13
  Speaker 1: bit of an artistic flare to it. And by that
• 14:16
  Speaker 1: I mean that it's not a dents sure why someone
• 14:20
  Speaker 1: who's more productive is more productive, but somehow they get
• 14:24
  Speaker 1: it more and so the people who are able to
• 14:27
  Speaker 1: do that are probably going to fare better. The one
• 14:30
  Speaker 1: concern we have is when I'm a person who's using
• 14:33
  Speaker 1: an AI tool, I can use it at a much
• 14:36
  Speaker 1: larger scale and make decisions for many more across a
• 14:41
  Speaker 1: greater domain. And in that regard, I'm sort of made
• 14:45
  Speaker 1: a superhuman. But you can only have so many superhumans,
• 14:48
  Speaker 1: and that's where we might worry about inequality. Josh, talk
• 14:52
  Speaker 1: to us a little bit also about this idea explored
• 14:55
  Speaker 1: in your book on how artificial intelligence might actually contribute
• 14:59
  Speaker 1: to the concentration of certain industries in our economy, even
• 15:04
  Speaker 1: aid the creation of monopolies. Well, I think this is
• 15:09
  Speaker 1: something that has occurred with any digital platforms, especially ones
• 15:13
  Speaker 1: that work better with larger scale. So the current AI
• 15:17
  Speaker 1: runs on machine learning, and I have to emphasize the
• 15:20
  Speaker 1: learning part. You get better by putting your machines out
• 15:25
  Speaker 1: in the field and continuously adjusting them with new data
• 15:30
  Speaker 1: and new learning. In so, what that means is that
• 15:34
  Speaker 1: a company such as Google that has a lot more
• 15:37
  Speaker 1: people using its search engine is able to use AI
• 15:40
  Speaker 1: to improve that those search engines at a greater rate
• 15:43
  Speaker 1: than competitors such as Being and Duct Duct Go, and
• 15:47
  Speaker 1: so in that regard, you could have the preservation of
• 15:52
  Speaker 1: a dominance or the emergence of dominance. And similarly this
• 15:55
  Speaker 1: might hold for companies like Facebook, and it might hold
• 15:59
  Speaker 1: to a degree for companies like Apple and Amazon as well.
• 16:03
  Speaker 1: They just have a lot more activity and so their
• 16:05
  Speaker 1: potential for to to use AI to learn at a
• 16:09
  Speaker 1: faster rate might be there. But as with all of
• 16:13
  Speaker 1: these things, is sometimes these firms do get into a rut,
• 16:16
  Speaker 1: and sometimes people find better ways of learning and doing
• 16:20
  Speaker 1: things that might initially perform worse but have a better trajectory.
• 16:24
  Speaker 1: So it's not a given that we're going to have
• 16:27
  Speaker 1: the current monopolies, be the future monopolies or anything like that,
• 16:31
  Speaker 1: or the current large companies, but we might see new
• 16:34
  Speaker 1: ones develop on the basis of new tools. Every other time,
• 16:37
  Speaker 1: we've had a large technological revolution that has occurred, I'd
• 16:40
  Speaker 1: expect it to occur this time too. Just taking a
• 16:44
  Speaker 1: broader view of competition, you also get into which country
• 16:50
  Speaker 1: might be the dominant force in AI. If there's a
• 16:52
  Speaker 1: dominant force, the US has already has a clear lead,
• 16:57
  Speaker 1: and yet there's a lot of activity going on in China.
• 17:02
  Speaker 1: How do you see China ascending and competing against the
• 17:06
  Speaker 1: United States in AI. Well, this is where being able
• 17:10
  Speaker 1: to have access to the right sort of data comes
• 17:12
  Speaker 1: to play, not just data but also talent. So let's
• 17:15
  Speaker 1: talk about data first. The United States is sort of
• 17:18
  Speaker 1: a middle ground in terms of privacy regulation. Europe is
• 17:22
  Speaker 1: far more stringent, but China there's none at all, and
• 17:26
  Speaker 1: so Chinese firms have the ability to appropriate and use
• 17:30
  Speaker 1: consumer data to develop AI at a much greater rate
• 17:34
  Speaker 1: than you would be able to do inside the United States.
• 17:37
  Speaker 1: So there's a benefit. Secondly, there's the issue of capabilities.
• 17:41
  Speaker 1: At the moment, AI resources are thin on the ground.
• 17:44
  Speaker 1: It's hard to hire engineers they command six or seven
• 17:48
  Speaker 1: figure sums. But the United States at the moment is
• 17:51
  Speaker 1: cutting itself off from the global pool of talent in
• 17:55
  Speaker 1: machine learning, and this is something other countries aren't doing.
• 17:58
  Speaker 1: Not only they not cutting them selves off, They're also
• 18:01
  Speaker 1: providing resources. China is spending several billion dollars on a
• 18:05
  Speaker 1: technology cluster in this area. Russia are doing the same,
• 18:10
  Speaker 1: and even countries like Canada, the government's actively supporting the
• 18:14
  Speaker 1: development of news superclusters in the space. That's going to
• 18:18
  Speaker 1: help attract talent around the world, because talent wants to
• 18:22
  Speaker 1: be able to work, and I think that's another area
• 18:25
  Speaker 1: where the United States faces some risks. Let's end this
• 18:29
  Speaker 1: interview with the existential question that you address in your
• 18:33
  Speaker 1: book or try to address. We've all seen our share
• 18:35
  Speaker 1: of science fiction movies. The Rise of the Robots Terminator
• 18:39
  Speaker 1: to looms very large in my mind for me, josh
• 18:43
  Speaker 1: is this the end of the world as we know it? Well,
• 18:45
  Speaker 1: as we say in the book, not enough time yet
• 18:49
  Speaker 1: to tell. But you've got enough time to read our
• 18:51
  Speaker 1: book and be on the right side of it. All right, Well,
• 18:54
  Speaker 1: let's send it there. Joshua Against from the University of Toronto,
• 18:57
  Speaker 1: author of Prediction Machines, thank you very much for running
• 19:00
  Speaker 1: us on Benchmark. Thank you. Benchmark will be back next week.
• 19:06
  Speaker 1: Until then, you can find us on the Bloomberg terminal
• 19:09
  Speaker 1: Bloomberg dot com. Our Bloomberg app and podcast destinations such
• 19:13
  Speaker 1: as Apple Podcast, Spotify, or wherever you listen. We'd love
• 19:17
  Speaker 1: it if you took the time to post a review
• 19:19
  Speaker 1: of the show so more listeners can find us. You
• 19:22
  Speaker 1: can also check us out on Twitter, follow me at
• 19:25
  Speaker 1: at scott Landman, Chris You're at Chris Ja Conden, and
• 19:30
  Speaker 1: our guest Josh Gannes is at Josh gannes g. A
• 19:33
  Speaker 1: n S Benchmark is produced by Toper Foreheaz. The head
• 19:37
  Speaker 1: of Bloomberg Podcasts is Francesca Levy. Thanks for listening. To
• 19:40
  Speaker 1: see you next time.