The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 27
The Economics of Artificial Intelligence Page 27

by Ajay Agrawal


  (Cowen 2011; Gordon 2016). Gordon (2016, 575) highlights this slowdown

  for the US economy. Between 1920 and 1970, total factor productivity grew

  at an annual average compound rate of 1.89 percent, falling to 0.57 per-

  cent between 1970 and 1994, then rebounding to 1.03 percent during the

  information technology boom between 1994 and 2004, before falling again

  to just 0.40 percent between 2004 and 2014. Even the maintenance of this

  lowered growth rate has only been possible due to exponential growth in

  the number of research workers (Jones 1995). Bloom et al. (2017) document

  that the total factor productivity in knowledge production itself has been

  falling both in the aggregate and in key specifi c knowledge domains such as

  transistors, health care, and agriculture.

  Economists have given a number of explanations for the disappointing

  growth performance. Cowen (2011) and Gordon (2016) point to a “fi shing

  out” or “low- hanging fruit” eff ect—good ideas are simply becoming harder

  to fi nd. Jones (2009) points to the headwind created by an increased “burden

  of knowledge.” As the technological frontier expands, it becomes harder for

  individual researchers to know enough to fi nd the combinations of knowl-

  edge that produce useful new ideas. This is refl ected in PhDs being awarded

  at older ages and a rise in team size as ever- more specialized researchers must

  combine their knowledge to produce breakthroughs (Agrawal, Goldfarb,

  and Teodoridis 2016). Other evidence points to the physical, social, and

  institutional constraints that limit access to knowledge, including the need

  to be physically close to the sources of knowledge (Jaff e, Trajtenberg, and

  Henderson 1993; Catalini 2017), the importance of social relationships in

  accessing knowledge (Mokyr 2002; Agrawal, Cockburn, and McHale 2006;

  Agrawal, Kapur, and McHale 2008), and the importance of institutions in

  facilitating—or limiting—access to knowledge (Furman and Stern 2011).

  Despite the evidence of a growth slowdown, one reason to be hopeful

  about the future is the recent explosion in data availability under the rubric

  of “big data” and computer- based advances in capabilities to discover and

  process those data. We can view these technologies in part as “meta tech-

  nologies”—technologies for the production of new knowledge. If part of the

  challenge is dealing with the combinatorial explosion in the potential ways

  that existing knowledge can be combined as the knowledge base grows, then

  meta technologies such as deep learning hold out the potential to partially

  overcome the challenges of fi shing out, the rising burden of knowledge, and

  the social and institutional constraints on knowledge access.

  Of course, meta technologies that aid in the discovery of new knowledge

  are nothing new. Mokyr (2002, 2017) gives numerous examples of how scien-

  tifi c instruments such as microscopes and x-ray crystallography signifi cantly

  Artifi cial Intelligence and Recombinant Growth 151

  aided the discovery process. Rosenberg (1998) provides an account of how

  technology- embodied chemical engineering altered the path of discovery in

  the petrochemical industry. Moreover, the use of artifi cial intelligence (AI)

  for discovery is itself not new and has underpinned fi elds such as chemin-

  formatics, bioinformatics, and particle physics for decades. However, recent

  breakthroughs in AI such as deep learning have given a new impetus to these

  fi elds.1 The convergence of graphical processing unit (GPU)- accelerated

  computing power, exponential growth in data availability buttressed in part

  by open data sources, and the rapid advance in AI- based prediction tech-

  nologies is leading to breakthroughs in solving many needle- in-a- haystack

  problems (chapter 3, this volume). If the curse of dimensionality is both

  the blessing and curse of discovery, advances in AI off er renewed hope of

  breaking the curse while helping to deliver on the blessing.

  Understanding how these technologies could aff ect future growth dynam-

  ics is likely to require an explicitly combinatorial framework. Weitzman’s

  (1998) pioneering development of a recombinant growth model has unfor-

  tunately not been well incorporated into the corpus of growth theory litera-

  ture. Our contribution in this chapter is thus twofold. First, we develop a

  relatively simple combinatorial- based knowledge production function that

  converges in the limit to the Romer/ Jones function. The model allows for

  the consideration of how existing knowledge is combined to produce new

  knowledge and also how researchers combine to form teams. Second, while

  this function can be incorporated into existing growth models, the specifi c

  combinatorial foundations mean that the model provides insights into how

  new metatechnologies such as artifi cial intelligence might matter for the path

  of future economic growth.

  The starting point for the model we develop is the Romer/ Jones knowl-

  edge production function. This function—a workhorse of modern growth

  theory—models the output of new ideas as a Cobb- Douglas function with

  the existing knowledge stock and labor resources devoted to knowledge

  production as inputs. Implicit in the Romer/ Jones formulation is that new

  knowledge production depends on access to the existing knowledge stock

  and the ability to combine distinct elements of that stock into valuable new

  ideas. The promise of AI as a meta technology for new idea production

  is that it facilitates the search over complex knowledge spaces, allowing

  for both improved access to relevant knowledge and improved capacity to

  predict the value of new combinations. It may be especially valuable where

  the complexity of the underlying biological or physical systems has stymied

  technological advance, notwithstanding the apparent promise of new fi elds

  such as biotechnology or nanotechnology. We thus develop an explicitly

  combinatorial- based knowledge production function. Separate parameters

  1. See, for example, the recent survey of the use of deep learning in computational chemistry by Garrett Goh, Nathan Hodas, and Abhinav Vishnu (2017).

  152 Ajay Agrawal, John McHale, and Alexander Oettl

  control the ease of knowledge access, the ability to search the complex space

  of potential combinations, and the ease of forming research teams to pool

  knowledge access. An attractive feature of our proposed function is that the

  Romer/ Jones function emerges as a limiting case. By explicitly delineating

  the knowledge access, combinatorial and collaboration aspects of knowl-

  edge production, we hope that the model can help elucidate how AI could

  improve the chances of solving needle- in-a- haystack- type challenges and

  thus infl uence the path of economic growth.

  Our chapter thus contributes to a recent but rapidly expanding literature

  on the eff ects of AI on economic growth. Much of the focus of this new

  literature is on how increased automation substitutes for labor in the produc-

  tion process. Building on the pioneering work of Zeira (1998), Acemoglu

  and Restrepo (2017) develop a model in which AI substitutes for workers in

  existing tasks
, but also creates new tasks for workers to do. Aghion, Jones,

  and Jones (chapter 9, this volume) show how automation can be consistent

  with relatively constant factor shares when the elasticity of substitution

  between goods is less than one. Central to their results is Baumol’s “cost

  disease,” which posits the ultimate constraint on growth to be from goods

  that are essential but hard to improve rather than goods whose production

  benefi ts from AI- driven technical change. In a similar vein, Nordhaus (2015)

  explores the conditions under which AI would lead to an “economic singu-

  larity” and examines the empirical evidence on the elasticity of substitution

  on both the demand and supply sides of the economy.

  Our focus is diff erent from these papers in that instead of emphasising the

  potential substitution of machines for workers in existing tasks, we empha-

  sise the importance of AI in overcoming a specifi c problem that impedes

  human researchers—fi nding useful combinations in complex discovery

  spaces. Our chapter is closest in spirit to Cockburn, Henderson, and Stern

  (chapter 4, this volume), which examines the implications of AI—and deep

  learning in particular—as a general purpose technology (GPT) for inven-

  tion. We provide a suggested formalization of this key idea. Nielsen (2012)

  usefully illuminates the myriad ways in which “big data” and associated

  technologies are changing the mechanisms of discovery in science. Nielsen

  emphasizes the increasing importance of “collective intelligence” in formal

  and informal networked teams, the growth of “data- driven intelligence”

  that can solve problems that challenge human intelligence, and the impor-

  tance of increased technology facilitating access to knowledge and data. We

  incorporate all of these elements into the model developed in this chapter.

  The rest of the chapter is organized as follows. In the next section, we

  outline some examples of how advances in artifi cial intelligence are chang-

  ing both knowledge access and the ability to combine knowledge in high-

  dimensional data across a number of domains. In section 5.3, we develop an

  explicitly combinatorial- based knowledge production function and embed

  it in the growth model of Jones (1995), which itself is a modifi cation of

  Artifi cial Intelligence and Recombinant Growth 153

  Romer (1990). In section 5.4, we extend the basic model to allow for knowl-

  edge production by teams. We discuss our results in section 5.5 and conclude

  in section 5.6 with some speculative thoughts on how an “economic singu-

  larity” might emerge.

  5.2

  How

  Artifi cial Intelligence Is Impacting the

  Production of Knowledge: Some Motivating Examples

  Breakthroughs in AI are already impacting the productivity of scientifi c

  research and technology development. It is useful to distinguish between

  such meta technologies that aid in the process of search (knowledge access)

  and discovery (combining existing knowledge to produce new knowledge).

  For search, we are interested in AIs that solve problems that meet two condi-

  tions: (a) potential knowledge relevant to the process of discovery is subject

  to an explosion of data that an individual researcher or team of researchers

  fi nds increasingly diffi

  cult to stay abreast of (the “burden of knowledge”);

  and (b) the AI predicts which pieces of knowledge will be most relevant to

  the researcher, typically through the input of search terms. For discovery,

  we also identify two conditions: (a) potentially combinable knowledge for

  the production of new knowledge is subject to combinatorial explosion,

  and (b) the AI predicts which combinations of existing knowledge will yield

  valuable new knowledge across a large number of domains. We now consider

  some specifi c examples of how AI- based search and discovery technologies

  may change the innovation process.

  5.2.1 Search

  Meta produces AI- based search technologies for identifying relevant

  scientifi c papers and tracking the evolution of scientifi c ideas. The company

  was acquired by the Chan- Zuckerberg Foundation, which intends to make

  it available free of charge to researchers. This AI- based search technology

  meets our two conditions for a meta technology for knowledge access: (a) the

  stock of scientifi c papers is subject to exponential growth at an estimated

  8– 9 percent per year (Bornmann and Mutz 2015), and (b) the AI- based

  search technology helps scientists identify relevant papers, thereby reduc-

  ing the “burden of knowledge” associated with the exponential growth of

  published output.

  BenchSci is an AI- based search technology for the more specifi c task of

  identifying eff ective compounds used in drug discovery (notably antibod-

  ies that act as reagents in scientifi c experiments). It again meets our two

  conditions: (a) reports on compound effi

  cacy are scattered through mil-

  lions of scientifi c papers with little standardization in how these reports are

  provided, and (b) an AI extracts compound- effi

  cacy information, allow-

  ing scientists to more eff ectively identify appropriate compounds to use in

  experiments.

  154 Ajay Agrawal, John McHale, and Alexander Oettl

  5.2.2 Discovery

  Atomwise is a deep learning- based AI for the discovery of drug molecules

  (compounds) that have the potential to yield safe and eff ective new drugs.

  This AI meets our two conditions for a meta technology for discovery: (a) the

  number of potential compounds is subject to combinatorial explosion, and

  (b) the AI predicts how basic chemical features combine into more intricate

  features to identify potential compounds for more detailed investigation.

  Deep Genomics is a deep learning- based AI that predicts what happens

  in a cell when DNA is altered by natural or therapeutic genetic variation.

  It again meets our two conditions: (a) genotype- phenotype variations are

  subject to combinatorial explosion, and (b) the AI “bridges the genotype-

  phenotype divide” by predicting the results of complex biological processes

  that relate variations in the genotype to observable characteristics of an

  organism, thus helping to identify potentially valuable therapeutic interven-

  tions for further testing.

  5.3 A Combinatorial- Based Knowledge Production Function

  Figure 5.1 provides an overview of our modeling approach and how it

  relates to the classic Romer/ Jones knowledge production function. The solid

  lines capture the essential character of the Romer/ Jones function. Research-

  ers use existing knowledge—the standing- on- shoulders eff ect—to produce

  Fig. 5.1 Romer/ Jones and combinatorial- based knowledge production functions

  Artifi cial Intelligence and Recombinant Growth 155

  new knowledge. The new knowledge then becomes part of the knowledge

  base from which subsequent discoveries are made. The dashed lines capture

  our approach. The existing knowledge base determines the potential new

  combinations that are possible, the majority of which are likely to have no

  value. The discovery of valuable new know
ledge is made by searching among

  the massive number of potential combinations. This discovery process is

  aided by meta technologies such as deep learning that allow researchers to

  identify valuable combinations in spaces where existing knowledge interacts

  in often highly complex ways. As with the Romer/ Jones function, the new

  knowledge adds to the knowledge base—and thus the potential combina-

  tions of that knowledge base—which subsequent researchers have to work

  with. A feature of our new knowledge production function will be that the

  Romer/ Jones function emerges as a limiting case both with and without

  team production of new knowledge. In this section, we fi rst develop the new

  function without team production of new knowledge; in the next section,

  we extend the function to allow for team production.

  The total stock of knowledge in the world is denoted as A, which we

  assume initially is measured discretely. An individual researcher has access

  to an amount of knowledge, A (also assumed to be an integer), so that the

  share of the stock of knowledge available to an individual researcher is A– 1.2

  We assume that 0 < < 1. This implies that the share of total knowledge

  accessible to an individual researcher is falling with the total stock of knowl-

  edge. This is a manifestation in the model of the “burden of knowledge”

  eff ect identifi ed by Jones (2009)—it becomes more diffi

  cult to access all the

  available knowledge as the total stock of knowledge grows. The knowledge

  access parameter, , is assumed to capture not only what a researcher knows

  at a point in time, but also their ability to fi nd existing knowledge should they

  require it. The value of the parameter will thus be aff ected by the extent to

  which knowledge is available in codifi ed form and can be found as needed

  by researchers. The combination of digital repositories of knowledge and

  search technologies that can predict what knowledge will be most relevant

  to the researcher given the search terms they input—think of the ubiquitous

  Google as well as more specialized search technologies such as Meta and

  BenchSci—should increase the value of .

  2. Paul Romer emphasized the importance of distinguishing between ideas (a nonrival good) and human capital (a rival good). “Ideas are . . . the critical input in the production of more valuable human and non- human capital. But human capital is also the most important input in the production of new ideas. . . . Because human capital and ideas are so closely related as inputs and outputs, it is tempting to aggregate them into a single type of good. . . . It is important, nevertheless, to distinguish ideas and human capital because they have diff erent fundamental attributes as economic goods, with diff erent implications for economic theory”

 

‹ Prev