by Ajay Agrawal
materials design and property prediction” (Goh, Hodas, and Vishu 2017).
Although the most publicized successes of deep learning have been in
areas such as image recognition, voice recognition, and natural language
processing, parallels to the way in which the new methods work on unstruc-
tured data are increasingly being identifi ed in many fi elds with similar data
challenges to produce research breakthroughs.7 While these new general
purpose research tools will not displace traditional mathematical models of
6. As described by LeCun, Bengio, and Hinton (2015, 436), “[c]onventional machine- learning techniques were limited in their ability to process natural data in their raw form. For decades, constructing a pattern- recognition or machine- learning system required careful engineering and considerable domain expertise to design a feature extractor that transformed the raw data (such as the pixel values of an image) into a suitable internal representation or feature vector from which the learning subsystem, often a classifi er, could detect or classify patterns in the input. . . . Representation learning is a set of methods that allows a machine to be fed with raw data and to automatically discover the representations needed for detection or classifi cation.”
7. A recent review of deep- learning applications in biomedicine usefully draws out these parallels: “With some imagination, parallels can be drawn between biological data and the types of data deep learning has shown the most success with—namely image and voice data.
A gene expression profi le, for instance, is essentially a ‘snapshot,’ or image, of what is going on in a given cell or tissue in the same way that patterns of pixilation are representative of the objects in a picture” (Mamoshina et al. 2016, 1445).
Artifi cial Intelligence and Recombinant Growth 169
cause and eff ect and careful experimental design, machine- learning methods
such as deep learning off er a promising new tool for discovery—including
hypothesis generation—where the complexity of the underlying phenomena
present obstacles to more traditional methods.8
5.5.2 Meta Ideas, Meta Technologies, and
General Purpose Technologies
We conceptualize AIs as general purpose meta technologies—that is,
general purpose technologies (GPTs) for the discovery of new knowledge.
Figure 5.6 summarises the relationship between Paul Romer’s broader idea
of meta ideas, meta technologies, and GPTs. Romer defi nes a meta idea as an
idea that supports the production and transmission of other ideas (see, e.g.,
Romer 2008). He points to such ideas as the patent, the agricultural exten-
sion station, and the peer- review system for research grants as examples
of meta ideas. We think of meta technologies as a subset of Romer’s meta
ideas (the area enclosed by the dashed lines in fi gure 5.6), where the idea for
how to discover new ideas is embedded in a technological form such as an
algorithm or measurement instrument.
Elhanan Helpman (1998, 3) argues that a “drastic innovation qualifi es
as a GPT if it has the potential for pervasive use in a wide range of sec-
tors in ways that drastically change their mode of operation.” He further
notes two important features necessary to qualify as a GPT: “generality
of purpose and innovational complementarities” (see also Bresnahan and
Trajtenberg 1995). Not all meta technologies are general purpose in this
sense. The set of general purpose meta technologies is given by the inter-
section of the two circles in fi gure 5.6. Cockburn, Henderson, and Stern
(chapter 4, this volume) give the example of functional MRI as an example
of a discovery tool that lacks the generality of purpose required for a GPT.
In contrast, the range of application of deep learning as a discovery tool
would appear to qualify it as a GPT. It is worth noting that some authors
discuss GPTs as technologies that more closely align with our idea of a meta
technology. Rosenberg (1998), for example, provides a fascinating examina-
tion of chemical engineering as an example of GPT. Writing of this branch
of engineering, he argues that a “discipline that provides the concepts and
methodologies to generate new or improved technologies over a wide range
of downstream economic activity may be thought of as an even purer, or
higher order, GPT” (Rosenberg 1998, 170).
8. A recent survey of the emerging use of machine learning in economics (including policy design) provides a pithy characterization of the power of the new methods: “The appeal of machine learning is that it manages to uncover generalizable patterns. In fact, the success of machine learning at intelligence tasks is largely due to its ability to discover complex structure that was not specifi ed in advance. It manages to fi t complex and very fl exible functional forms to the data without simply overfi tting; it fi nds functions that work well out of sample” (Mullainathan and Spiess 2017, 88).
170 Ajay Agrawal, John McHale, and Alexander Oettl
Fig. 5.6 Relationships between meta ideas, meta technologies, and general
purpose technologies
Our concentration on general purpose meta technologies (GPMTs) par-
allels Cockburn, Henderson, and Stern’s (chapter 4, this volume) idea of a
general purpose invention of a method of invention. This idea combines
the idea of a GPT with Zvi Griliches’ (1957) idea of the “invention of a
method of invention,” or IMI. Such an invention has the “potential for a
more infl uential impact than a single invention, but is also likely to be associ-
ated with a wide variation in the ability to adapt the new tool to particular
settings, resulting in a more heterogeneous pattern of diff usion over time”
(Cockburn, Henderson, and Stern, chapter 4, this volume). They see some
emerging AIs such as deep learning as candidates for such general purpose
IMIs and contrast these with AIs underpinning robotics that, while being
GPTs, do not have the characteristic features of an IMI.
5.5.3 Beyond AI: Potential Uses of the
New Knowledge Production Function
Although the primary motivation for this chapter is to explore how break-
throughs in AI could aff ect the path of economic growth, the knowledge
production function we develop is potentially of broader applicability. By
deriving the Romer/ Jones knowledge production function as the limiting
case of a more general function, our analysis may also contribute to pro-
viding candidate microfoundations for that function.9 The key conceptual
9. In developing and applying the Romer/ Jones knowledge production function, growth theorists have understood its potential combinatorial underpinnings and the limits of the Cobb-Douglas form. Charles Jones (2005) observes in his review chapter on “Growth and Ideas” for the Handbook of Economic Growth: “While we have made much progress in understanding economic growth in a world where ideas are important, there remain many open, interesting research questions. The fi rst is ‘What is the shape of the idea production function?’ How do
Artifi cial Intelligence and Recombinant Growth 171
change is to model discovery as operating on the space of potential combi-
nations (rather than directly on the knowledge base itself ). As in Weitzman
(1998), our production function focuses attention explicitly on how new
knowledge is discovered by combining existing knowledge, which is left
implicit in the Romer/ Jones formulation. While this shift in emphasis is
&
nbsp; motivated by the particular way in which deep learning can aid discovery—
allowing researchers to uncover otherwise hard- to-fi nd valuable combina-
tions in highly complex spaces—the view of discovery as the innovative
combination of what is already known has broader applicability. The more
general function also has the advantage of providing a richer parameter
space for mapping how meta technologies or policies could aff ect knowledge
discovery. The parameter captures how access to knowledge at the indi-
vidual researcher level determines the potential for new combinations to be
made given the inherited knowledge base. The parameter captures how
the available potential combinations (given the access to knowledge) map
to new discoveries. Finally, the parameter captures the ease of forming
research teams and ultimately the average team size. To the extent that the
capacity to bring the knowledge of individual researchers together through
research teams directly aff ects the possible combinations, the ease of team
formation can have an important eff ect on how the existing knowledge base
is utilized for new knowledge discovery.
We hope this more general function will be of use in other contexts.
In a recent commentary celebrating the twenty- fi fth anniversary of the
publication of Romer (1990), Joshua Gans (2015) observes that the Romer
growth model has not been as infl uential on the design of growth policy as
might have been expected despite its enormous infl uence on the subsequent
growth theory literature. The reason he identifi es is that it abstracts away
“some of the richness of the microeconomy that give rise to new ideas and
also their dissemination” (Gans 2015). By expanding the parameter space,
our function allows for the inclusion of more of this richness, including the
role that meta technologies such as deep learning can play in knowledge
access and knowledge discovery, but potentially other policy and insti-
tutional factors that aff ect knowledge access, discovery rates, and team
formation as well.
ideas get produced? The combinatorial calculations of Romer (1993) and Weitzman (1998) are fascinating and suggestive. The current research practice of modelling the idea production function as a stable Cobb- Douglas combination of research and the existing stock of ideas is elegant, but at this point we have little reason to believe it is correct. One insight that illustrates the incompleteness of our knowledge is that there is no reason why research productivity should be a smooth monotonic function of the stock of ideas. One can easily imagine that some ideas lead to domino- like unravelling of phenomena that were previously mysterious . . . Indeed, perhaps decoding of the human genome or the continued boom in information technology will lead to a large upward shift in the production function for ideas. On the other hand, one can equally imagine situations where research productivity unexpectedly stagnates, if not forever then at least for a long time” (Jones 2005, 1107).
172 Ajay Agrawal, John McHale, and Alexander Oettl
5.6 Concluding Thoughts: A Coming Singularity?
We developed this chapter upon a number of prior ideas. First, the pro-
duction of new knowledge is central to sustaining economic growth (Romer
1990, 1993). Second, the production of new ideas is fundamentally a combi-
natorial process (Weitzman 1998). Third, given this combinatorial process,
technologies that predict what combinations of existing knowledge will yield
useful new knowledge hold out the promise of improving growth prospects.
Fourth, breakthroughs in AI represent a potential step change in the ability
of algorithms to predict what knowledge is potentially useful to researchers
and also to predict what combinations of existing knowledge will yield use-
ful new discoveries (LeCun, Bengio, and Hinton 2015).
In a provocative recent paper, William Nordhaus (2015) explored the pos-
sibilities for a coming “economic singularity,” which he defi nes as “[t]he
idea . . . that rapid growth in computation and artifi cial intelligence will
cross some boundary or singularity after which economic growth will accel-
erate sharply as an ever- accelerating pace of improvements cascade through
the economy.” Central to Nordhaus’ analysis is that rapid technological
advance is occurring in a relatively small part of the economy (see also
Aghion, Jones, and Jones 2018). To generate more broadly based rapid
growth, the products of the new economy need to substitute for products
on either the demand- or supply- side of the economy. His review of the evi-
dence—including, critically, the relevant elasticities of substitution—leads
him to conclude that a singularity through this route is highly unlikely.
However, our chapter’s analysis suggests an alternative route to an eco-
nomic singularity—a broad- based alteration in the economy’s knowledge
production function. Given the centrality of new knowledge to sustained
growth at the technological frontier, it seems likely that if an economic sin-
gularity were to arise, it would be because of some signifi cant change to the
knowledge production function aff ecting a number of domains outside of
information technology itself. In a world where new knowledge is the result
of combining existing knowledge, AI technologies that help ease needle-
in-the- haystack discovery challenges could aff ect growth prospects, at least
along the transition path to the steady state. It does not take an impossible
leap of imagination to see how new meta technologies such as AI could
alter—perhaps modestly, perhaps dramatically—the knowledge production
function in a way that changes the prospects for economic growth.
References
Acemoglu, Daron, and Pascual Restrepo. 2017. “The Race between Machine and
Man: Implications of Technology for Growth, Factor Shares and Employment.”
NBER Working Paper no. 22252, Cambridge, MA..
Artifi cial Intelligence and Recombinant Growth 173
Agrawal, A., I. Cockburn, and J. McHale. 2006. “Gone But Not Forgotten: Knowl-
edge Flows, Labor Mobility, and Enduring Social Relationships.” Journal of Eco-
nomic Geography 6 (5): 571– 91.
Agrawal, Ajay, and Avi Goldfarb. 2008. “Restructuring Research: Communication
Costs and the Democratization of University Innovation.” American Economic
Review 98 (4): 1578– 90.
Agrawal, A., Avi Goldfarb, and Florenta Teodordis. 2016. “Understanding the
Changing Structure of Scientifi c Inquiry.” American Economic Journal: Applied
Economics 8 (1): 100– 28.
Agrawal, A., D. Kapur, and J. McHale. 2008. “How to Spatial and Social Proxim-
ity Infl uence Knowledge Flows: Evidence from Patent Data.” Journal of Urban
Economics 64: 258– 69.
Angermueller, Christof, Tanel Pärnamaa, Leopold Parts, and Oliver Stegle. 2016.
“Deep Learning for Computational Biology.” Molecular Systems Biology 12 (878):
1– 16.
Arthur, Brian W. 2009. The Nature of Technology: What it Is and How it Evolves.
London: Penguin Books.
Bloom, Nicholas, Charles Jones, John Van Reenen, and Michael Webb. 2017. “Are
Ideas Getting Harder to Find?” Working Paper, Stanford University.
Bornmann, Lutz, and Rüdiger Mutz. 2015. “Growth Rates of Modern Science
: A
Bibliometric Analysis Based on the Number of Publications and Cited Refer-
ences.” Journal of the Association for Information Science and Technology 66 (11): 2215– 22.
Bresnahan, Timothy, and Manuel Trajtenberg. 1995. “General Purpose Technolo-
gies ‘Engines of Growth’?” Journal of Econometrics 65:83– 108.
Brynjolfsson, Erik, and Andrew McAfee. 2014. The Second Machine Age: Work Pro-
gress and Prosperity in a Time Of Brilliant Technologies. New York: W. W. Norton.
Catalini, Christian. 2017. “Microgeography and the Direction of Inventive Activity.”
Management Science 64 (9). https:// doi .org/ 10.1287/ mnsc.2017.2798.
Cowen, Tyler. 2011. The Great Stagnation: How America Ate All the Low- Hanging
Fruit of Modern History, Got Sick, and Will (Eventually) Feel Better. New York: Dutton, Penguin Group.
Fleming, Lee, and Olav Sorenson. 2004. “Science as a Map in Technological Search.”
Strategic Management Journal 25:909– 28.
Furman, Jeff rey, and Scott Stern. 2011. “Climbing atop the Shoulders of Giants:
The Impact of Institutions on Cumulative Research.” American Economic Review
101:1933– 63.
Gans, Joshua. 2015. “The Romer Model Turns 25.” Digitopoly Blog. Accessed
Aug. 21, 2017. https:// digitopoly .org/ 2015/ 10/ 03/ the- romer- model- turns- 25/.
Goh, Garrett, Nathan Hodas, and Abhinav Vishnu. 2017. “Deep Learning for Com-
putational Chemistry.” Journal of Computational Chemistry 38 (16): 1291– 307.
Gordon, Robert. 2016. The Rise and Fall of American Growth: The U.S. Standard of
Living Since the Civil War. Princeton, NJ: Princeton University Press.
Griliches, Zvi. 1957. “Hybrid Corn: An Exploration in the Economics of Technical
Change.” Econometrica 25 (4): 501– 22.
Helpman, Elhanan. 1998. “Introduction.” In General Purpose Technologies and Eco-
nomic Growth, edited by Elhanan Helpman. Cambridge, MA: MIT Press.
Jaff e, Adam, Manuel Trajtenberg, and Rebecca Henderson. 1993. “Geographic
Localization of Knowledge Spillovers as Evidenced by Patent Citations.” Quar-
terly Journal of Economics 108 (3): 577– 98.
Jones, Benjamin. 2009. “The Burden of Knowledge and the ‘Death of the Renais-