The Economics of Artificial Intelligence Read online

  Hatzius, Jan, and Kris Dawsey. 2015. “Doing the Sums on Productivity Paradox

  v2.0.” Goldman Sachs U.S. Economics Analyst, no. 15/ 30.

  Henderson, Rebecca. 1993. “Underinvestment and Incompetence as Responses to

  Radical Innovation: Evidence from the Photolithographic Industry.” RAND Jour-

  nal of Economics 24 (2): 248– 70.

  ———. 2006. “The Innovator’s Dilemma as a Problem of Organizational Compe-

  tence.” Journal of Product Innovation Management 23:5– 11.

  Holmes, Thomas J., David K. Levine, and James A. Schmitz. 2012. “Monopoly

  and the Incentive to Innovate When Adoption Involves Switchover Disruptions.”

  American Economic Journal: Microeconomics 4 (3): 1– 33.

  Hortaçsu, Ali, and Chad Syverson. 2015. “The Ongoing Evolution of US Retail:

  A Format Tug- of-War.” Journal of Economic Perspectives 29 (4): 89– 112.

  Jones, C. I., and P. M. Romer. 2010. “The New Kaldor Facts: Ideas, Institutions,

  Population, and Human Capital.” American Economic Journal: Macroeconomics

  2 (1): 224– 45.

  Jovanovic, Boyan, and Peter L. Rousseau. 2005. “General Purpose Technologies.”

  In Handbook of Economic Growth, vol. 1B, edited by Philippe Aghion and Steven

  N. Durlauf, 1181– 224. Amsterdam: Elsevier B.V.

  Kendrick, John W. 1961. Productivity Trends in the United States. National Bureau of Economic Research. Princeton, NJ: Princeton University Press.

  Levine, S., C. Finn, T. Darrell, and P. Abbeel. 2016. “End- to-End Training of Deep

  Visuomotor Policies.” Journal of Machine Learning Research 17 (39): 1– 40.

  Levitt, Steven D., John A. List, and Chad Syverson. 2013. “Toward an Understand-

  ing of Learning by Doing: Evidence from an Automobile Plant.” Journal of Po-

  litical Economy 121 (4): 643– 81.

  Levy, Frank. 2018. “Computers and Populism: Artifi cial Intelligence, Jobs, and Poli-

  tics in the Near Term.” Oxford Review of Economic Policy 34 (3): 393– 417.

  Liu, Y., A. Gupta, P. Abbeel, and S. Levine. 2017. “Imitation from Observation:

  Learning to Imitate Behaviors from Raw Video via Context Translation.” arXiv

  preprint arXiv:1707.03374. https:// arxiv .org / abs / 1707 .03374.

  Manyika, James, Michael Chui, Mehdi Miremadi, Jacques Bughin, Katy George,

  Paul Willmott, and Martin Dewhurst. 2017. “Harnessing Automation for a

  Future That Works.” McKinsey Global Institute, January. https:// www .mckinsey

  .com / global - themes / digital - disruption / harnessing - automation - for - a - future - that

  - works.

  McAfee, Andrew, and Erik Brynjolfsson. 2008. “Investing in the IT that Makes a

  Competitive Diff erence.” Harvard Business Review July:98.

  Milgrom, P., and J. Roberts. 1996. “The LeChatelier Principle.” American Economic

  Review 86 (1): 173– 79.

  Minsky, Marvin. 1967. Computation: Finite and Infi nite Machines. Upper Saddle

  River, NJ: Prentice- Hall.

  Mokyr, J. 2014. “Secular Stagnation? Not in Your Life.” Geneva Reports on the World Economy August:83– 89.

  Morris, David Z. 2016. “Today’s Cars Are Parked 95 Percent of the Time.” Fortune, Mar. 13.

  Nakamura, Leonard, and Rachel Soloveichik. 2015. “Capturing the Productivity

  Comment 57

  Impact of the ‘Free’ Apps and Other Online Media.” FRBP Working Paper no.

  15– 25, Federal Reserve Bank of Philadelphia.

  Nordhaus, W. D. 2015. “Are We Approaching an Economic Singularity? Informa-

  tion Technology and the Future of Economic Growth.” NBER Working Paper

  no. 21547, Cambridge, MA.

  Organisation for Economic Co- operation and Development (OECD). 2015. The

  Future of Productivity. https:// www .oecd .org / eco / growth / OECD-2015-The- future

  - of-productivity- book .pdf.

  Orlikowski, W. J. 1996. “Improvising Organizational Transformation over Time: A

  Situated Change Perspective .” Information Systems Research 7 (1): 63– 92.

  Pratt, Gill A. 2015. “Is a Cambrian Explosion Coming for Robotics?” Journal of

  Economic Perspectives 29 (3): 51– 60.

  Saon, G., G. Kurata, T. Sercu, K. Audhkhasi, S. Thomas, D. Dimitriadis, X. Cui,

  et al. 2017. “English Conversational Telephone Speech Recognition by Humans

  and Machines.” arXiv preprint arXiv:1703.02136. https:// arxiv .org / abs / 1703

  .02136.

  Smith, Noah. 2015. “The Internet’s Hidden Wealth.” Bloomberg View, June 6. http://

  www .bloombergview .com / articles / 2015– 06– 10/ wealth- created- by- the- internet

  - may- not- appear- in-gdp.

  Solow, Robert M. 1957. “Technical Change and the Aggregate Production Func-

  tion.” Review of Economics and Statistics 39 (3): 312– 20.

  ———. 1987. “We’d Better Watch Out.” New York Times Book Review, July 12, 36.

  Song, Jae, David J. Price, Fatih Guvenen, Nicholas Bloom, and Till von Wachter.

  2015. “Firming Up Inequality.” NBER Working Paper no. 21199, Cambridge, MA.

  Stiglitz, Joseph E. 2014. “Unemployment and Innovation.” NBER Working Paper

  no. 20670, Cambridge, MA.

  Syverson, Chad. 2013. “Will History Repeat Itself ? Comments on ‘Is the Informa-

  tion Technology Revolution Over?’ ” International Productivity Monitor 25:37– 40.

  ———. 2017. “Challenges to Mismeasurement Explanations for the US Productiv-

  ity Slowdown.” Journal of Economic Perspectives 31 (2): 165– 86.

  Yang, Shinkyu, and Erik Brynjolfsson. 2001. “Intangible Assets and Growth Ac-

  counting: Evidence from Computer Investments.” Unpublished manuscript, Mas-

  sachusetts Institute of Technology.

  Comment Rebecca Henderson

  “Artifi cial Intelligence and the Modern Productivity Paradox” is a fabulous

  chapter. It is beautifully written, extremely interesting, and goes right to the

  heart of a centrally important question, namely, what eff ects will AI have on

  economic growth? The authors make two central claims. The fi rst is that AI

  Rebecca Henderson is the John and Natty McArthur University Professor at Harvard Uni-

  versity, where she has a joint appointment at the Harvard Business School in the General Management and Strategy units, and a research associate of the National Bureau of Economic Research.

  For acknowledgments, sources of research support, and disclosure of the author’s material fi nancial relationships, if any, please see http:// www .nber .org / chapters / c14020 .ack.

  58 Rebecca Henderson

  is a general purpose technology, or GPT, and as such is likely to have a dra-

  matic impact on productivity and economic growth. The second is that the

  reason we do not yet see it in the productivity statistics is because—like all

  GPTs—this is a technology that will take time to diff use across the economy.

  More specifi cally, the authors argue that AI will take time to diff use

  because its adoption will require mastering “adjustment costs, organiza-

  tional changes, and new skills.” They suggest that just as we did not see IT

  in the productivity statistics until fi rms had made the organizational changes

  and hired the human capital necessary to master it, so the adoption of AI

  will require not only the diff usion of the technology itself but also the de-

  velopment of the organizational and human assets that will be required to

  exploit its full potential.

  This is a fascinating idea
. One of the reasons I like the chapter so much

  is that takes seriously an idea that economists long resisted—namely, that

  things as nebulous as “culture” and “organizational capabilities” might be

  (a) very important, (b) expensive, and (c) hard to change. Twenty- fi ve years

  ago, when I submitted a paper to the RAND Journal of Economics that

  suggested that incumbents were fundamentally disadvantaged compared to

  entrants because they were constrained by old ways of acting and perceiving,

  I got a letter from the editor that began “Dear Rebecca, you have written

  a paper suggesting that the moon is made of green cheese, and that econo-

  mists have too little considered the motions of cheesy planetoids”

  I like to think that few editors would respond that way today. Thanks

  to a wave of new work in organizational economics and the pioneering

  empirical research of scholars like Nick Bloom, John van Reenen, Raff aella

  Sadun, and the authors themselves, we now have good reason to believe that

  managerial processes and organizational structures have very real eff ects

  on performance and that they take a signifi cant time to change. One of the

  most exciting things about this chapter is that it takes these ideas suffi

  ciently

  seriously to suggest that the current slowdown in productivity is largely a

  function of organizational inertia—that a central macroeconomic outcome

  is a function of a phenomenon that thirty years ago was barely on the radar.

  That’s exciting. Is it true? And if it is, what are its implications?

  My guess is that the deployment of AI will indeed be gated by the need to

  change organizational structures and processes. But I think that the authors

  may be underestimating the implications of this dynamic in important ways.

  Take the case of accounting. A few months ago, I happened to meet the

  chief strategy offi

  cer for one of the world’s largest accounting fi rms. He

  told me that his fi rm is the largest hirer of college graduates in the world—

  which may or may not be true, but which he certainly believed—and that

  his fi rm was planning to reduce the number of college graduates they hire

  by 75 percent over the next four to fi ve years—largely because it is increas-

  ingly clear that AI is going to be able to take over much of the auditing work

  currently performed by humans. This shift will certainly be mediated by

  Comment 59

  every accounting fi rm’s ability to integrate AI into their procedures and to

  persuade their customers that it is worth paying for—examples of exactly

  the kinds of barriers that this chapter suggests are so important—but in

  principle it should dramatically increase the productivity of accounting ser-

  vices, exactly the eff ects that Erik and his coauthors are hoping for.

  But I am worried about all the college graduates the accounting fi rms are

  not going to hire. More broadly, as AI begins to diff use across the economy

  it seems likely that a lot of people will get pushed into new positions and a

  lot of people will be laid off . And just as changing organizational processes

  takes time, so it’s going to take time to remake the social context in ways

  that will make it possible to handle these dislocations. Without these kinds

  of investments—one can imagine they might be in education, in relocation

  assistance, and the like—there is a real risk of a public backlash against AI

  that could dramatically reduce its diff usion rate.

  For example, the authors are excited about the benefi ts that the wide-

  spread diff usion of autonomous vehicles are likely to bring. Productivity

  seems likely to skyrocket, while with luck tens of thousands of people will

  no longer perish in car crashes every year. But “driving” is one of the larg-

  est occupations there is. What will happen when millions of people begin to

  be laid off ? I’m with the authors in believing that the diff usion of AI could

  be an enormous source of innovation and growth. But I can see challenges

  in the transition at the societal level, as well as at the organizational level.

  And there will also be challenges if too large a share of the economic gains

  from the initial deployment of the technology goes to the owners of capital

  rather than to the rest of society.

  Which is to say that I am a little more pessimistic than Erik and his co-

  authors as to the speed at which AI will diff use—and this is even before I

  start talking about the issues that Scott, Iain, and I touch on in our own

  chapter, namely, that we are likely to have signifi cant underinvestment in AI

  relative to the social option, coupled with a fair amount of dissipative racing.

  2

  The Technological Elements

  of Artifi cial Intelligence

  Matt Taddy

  2.1 Introduction

  We have seen in the past decade a sharp increase in the extent that compa-

  nies use data to optimize their businesses. Variously called the “Big Data” or

  “Data Science” revolution, this has been characterized by massive amounts

  of data, including unstructured and nontraditional data like text and images,

  and the use of fast and fl exible machine learning (ML) algorithms in anal-

  ysis. With recent improvements in deep neural networks (DNNs) and related

  methods, application of high- performance ML algorithms has become

  more automatic and robust to diff erent data scenarios. That has led to the

  rapid rise of an artifi cial intelligence (AI) that works by combining many ML

  algorithms together—each targeting a straightforward prediction task—to

  solve complex problems.

  In this chapter, we will defi ne a framework for thinking about the ingre-

  dients of this new ML- driven AI. Having an understanding of the pieces

  that make up these systems and how they fi t together is important for those

  who will be building businesses around this technology. Those studying the

  economics of AI can use these defi nitions to remove ambiguity from the

  conversation on AI’s projected productivity impacts and data requirements.

  Finally, this framework should help clarify the role for AI in the practice

  of modern business analytics1 and economic measurement.

  This article was written while Matt Taddy was professor of econometrics and statistics at the University of Chicago Booth School of Business and a principal researcher at Microsoft Research New England. He is currently at Amazon.com.

  For acknowledgments, sources of research support, and disclosure of the author’s material fi nancial relationships, if any, please see http:// www .nber .org/ chapters/ c14021.ack.

  1. This material has been adapted from a chapter in Business Data Science, forthcoming from McGraw-Hill.

  61

  62 Matt Taddy

  2.2 What Is AI?

  In fi gure 2.1, we show a breakdown of AI into three major and essential

  pieces. A full end- to-end AI solution—at Microsoft, we call this a System

  of Intelligence—is able to ingest human- level knowledge (e.g., via machine

  reading and computer vision) and use this information to automate and

  accelerate tasks that were previously only performed by humans. It is neces-

  sary here to have a well- defi ned task structure to engineer against, and in a
/>
  business setting this structure is provided by business and economic domain

  expertise. You need a massive bank of data to get the system up and running,

  and a strategy to continue generating data so that the system can respond

  and learn. And fi nally, you need machine- learning routines that can detect

  patterns in and make predictions from the unstructured data. This section

  will work through each of these pillars, and in later sections we dive in detail

  into deep learning models, their optimization, and data generation.

  Notice that we are explicitly separating ML from AI here. This is impor-

  tant: these are diff erent but often confused technologies. Machine learn-

  ing can do fantastic things, but it is basically limited to predicting a future

  that looks mostly like the past. These are tools for pattern recognition. In

  contrast, an AI system is able to solve complex problems that have been

  previously reserved for humans. It does this by breaking these problems

  into a bunch of simple prediction tasks, each of which can be attacked by

  a “dumb” ML algorithm. Artifi cial intelligence uses instances of machine

  learning as components of the larger system. These ML instances need to

  be organized within a structure defi ned by domain knowledge, and they

  need to be fed data that helps them complete their allotted prediction tasks.

  This is not to down- weight the importance of ML in AI. In contrast to

  earlier attempts at AI, the current instance of AI is ML driven. Machine-

  learning algorithms are implanted in every aspect of AI, and below we

  describe the evolution of ML toward status as a general purpose technology.

  This evolution is the main driver behind the current rise of AI. However,

  ML algorithms are building blocks of AI within a larger context.

  To make these ideas concrete, consider an example AI system from the

  Microsoft- owned company Maluuba that was designed to play (and win!)

  the video game Ms. Pac- Man on Atari (van Seijen et al. 2017).The system

  Fig. 2.1 AI systems are self- training structures of ML predictors that automate

  and accelerate human tasks

  The Technological Elements of Artifi cial Intelligence 63

  is illustrated in fi gure 2.2. The player moves Ms. Pac- Man on this game

  “board,” gaining rewards for eating pellets while making sure to avoid get-