The Economics of Artificial Intelligence
Page 78
competitive goods and services.
Many experts are weighing in on how to counter the “Chinese threat”
and, more generally, how to enrich local economies through cluster poli-
cies that support sustained competitive advantage in AI- based market
segments. Geoff Hinton and collaborators have convinced Canadian gov-
ernments to develop a major AI institute that would “graduate the most
machine- learning PhDs and master’s students globally” and “become the
engine for an AI supercluster that drives the economy of Toronto, Ontario,
and Canada.”5 Hinton also emphasizes the importance of access to data.
“Why? Because for a machine to ‘think’ intelligently, it must be trained with
lots of data.”
While there are potential benefi ts from Hinton’s initiative, it raises two
important points that loom large in our thinking. First, economists who
specialize in clusters are deeply skeptical about the effi
cacy of cluster policies
(e.g., Duranton 2011). Such policies have failed more often than not, and the
theoretical justifi cation for cluster policies is highly sensitive to assumptions
about knowledge diff usion. For example, will Hinton’s PhDs stay in Canada
and will the knowledge they generate be commercialized in Canada? Second,
a host of behind- the- border regulations on privacy, data localization, tech-
nology standards, and industrial policy will aff ect the ability of Canadian
fi rms to access data relative to their competitors in larger markets such as the
5. Globe and Mail, Jan. 7, 2017.
468 Avi Goldfarb and Daniel Trefl er
United States, Europe, and China. What is the current state of these domes-
tic data regulations, how do they eff ect trade patterns, do they serve a public
interest, are they being used as disguised protection to generate comparative
advantage, and should they be covered by international trade agreements (as
some would have been in the TPP e-commerce chapter)?
The following sections help answer these questions and move us toward
better policies for promoting AI and preventing both corporate welfare and
welfare- reducing disguised protection.
19.2 The Technological Backdrop: Scale, Scope,
Firm Size, and Knowledge Diff usion
The Oxford English Dictionary defi nes AI as “the theory and develop-
ment of computer systems able to perform tasks normally requiring human
intelligence.” This has meant diff erent things at diff erent times. In the 1960s
and 1970s, computer scientists approached this using rules, if- then state-
ments, and symbolic logic. It worked well for factory robots and for playing
chess. By the 1980s, it became clear that symbolic logic could not deal with
the complexities of nonartifi cial settings, and AI research slowed substan-
tially. Various approaches continued to be supported in a small number
of locations, including by the Canadian Institute for Advanced Studies
(CIFAR).
The recent resurgence in AI research is driven by one such approach: the
insight that computers can “learn” from example. This approach is often
called “machine learning” and is a fi eld of computational statistics. The
algorithm that has received the most attention is back propagation in neural
networks, most notably through “deep learning,” but there is a large suite
of relevant technologies including deep learning, reinforcement learning,
and so forth. Because the current excitement about AI is driven by machine
learning, we focus on this particular set of algorithms here.
For our purposes, we need to zero in on those aspects of AI technology
that are central to thinking about the economics of AI. We identify four
aspects: economies of scale associated with data, economies of scale associ-
ated with an AI research team, economies of scope in the use of the team
for multiple applications, and knowledge externalities.
19.2.1 Economies of Scale from Data
Statistical predictions improve with the quantity and quality of data.
Recall from statistics 101 that the quality of prediction increases with N (or, more precisely with root N ). All else being equal, this means that companies
that have more observations will generate more accurate predictions. It is in
this sense that economies of scale matter. Still, because predictions increase
in root N, then, while scale matters, there are decreasing returns to scale in
terms of the accuracy of prediction.
Artifi cial Intelligence and International Trade 469
It is subtler than this, however. Google and Microsoft both operate search
engines. Google has claimed their search engine has higher market share
because it has better quality.6 Microsoft has claimed the higher quality is a
direct consequence of scale. By having more data, Google can predict what
people want in their search results more accurately. Google responds that
Microsoft has billions of search results. While Google has more data, surely
the law of large numbers applies before one billion results. And so, more data
does not give a meaningful advantage. Microsoft’s response is the essence
of where economies of scale bind. While they have billions of searches,
many search queries are extremely rare. Microsoft may only see two or three,
and so Google can predict those rare queries much better. If people choose
search engines based on quality diff erences in rare searches, then Google’s
better data will lead to a substantial increase in market share. Having a larger
share gives Google more data, which in turn improves quality and supports
an even larger share.
The source of economies of scale here is therefore in the form of direct
network externalities. More customers generate more data, which in turn
generates more customers. This is diff erent from the literature on two- sided
markets and indirect network externalities. The network externalities re-
semble the phone network, rather than externalities between buyers and
sellers on a marketplace like Ebay. This is signifi cant in a trade context
because the trade literature has emphasized two- sided matching, for ex-
ample, in Rauch (1999) and McLaren (2000). This is also diff erent from all
of the trade and market structure literature, which emphasize economies of
scale that are driven by fi xed costs, so trade theory does not currently have
models that are applicable to the AI technology environment.
The direct network externalities environment leads to a core aspect of
competition in AI: competition for data. The companies that have the best
data make better predictions. This creates a positive feedback loop so that
they can collect even more data. In other words, the importance of data leads
to strong economies of scale.
19.2.2 Economies of Scale from the
Overhead of Developing AI Capabilities
Another source of economies of scale in AI involves the fi xed cost of
building an AI capability within a fi rm. The main cost is in personnel. Much
of the software is open source, and in many cases hardware can be purchased
as a utility through cloud services. The uses of AI need to be big enough to
justify the substan
tial cost of building a team of AI specialists. World lead-
ers in AI command very high pay, often in the millions or tens of millions.
6. There is a chicken and egg problem, whether good algorithms drive market share or
whether market share drives hiring that leads to better algorithms. For one point of view, see https:// www .cnet .com/ news/ googles- varian- search- scale- is- bogus/.
470 Avi Goldfarb and Daniel Trefl er
Top academic researchers have been hired to join Google (Hinton), Apple
(Salakhutdinov), Facebook (LeCunn), and Uber (Urtasun). So far, there has
been a meaningful diff erence between employing the elite researchers and
others in terms of the capabilities of the AI being developed.
19.3.3 Economies of Scope
Perhaps more than economies of scale, the fi xed cost of building an AI
capacity generates economies of scope. It is only worth having an AI team
within a company if there are a variety of applications for them to work
on. Many of the currently leading AI fi rms are multiproduct fi rms. For ex-
ample, Google parent Alphabet runs a search engine (Google), an online
video service (YouTube), a mobile device operating system (Android), an
autonomous vehicle division (Waymo), and a variety of other businesses.
In most cases, the economies of scope happen on the supply side through
AI talent, better hardware, and better software.
Another important source of economies of scope is the sharing of data
across applications. For example, the data from Google’s search engine
might be valuable in helping determine the eff ectiveness of YouTube adver-
tising, or its mapping services might be needed for developing autonomous
vehicles. The sharing of data is a key source of international friction on
disguised protection behind the border. Diff erences in privacy policies mean
that it is easier to share data across applications in some countries compared
to others. For example, when Ebay owned PayPal, it faced diff erent restric-
tions for using the PayPal data in Canada compared to the United States.
We will return to this subject later.
This contrasts with the main emphasis in the trade literature on economies
of scope, which emphasizes the demand side. Economies of scope in AI do
not seem to be about demand externalities in brand perception or in sales
channels. Instead, they appear to be driven by economies of scope in innova-
tion. A wider variety of potential applications generates greater incentives
to invest in an AI research team, and it generates more benefi ts to each
particular AI project due to the potential to share data across applications.
19.3.4 Knowledge
Externalities
There is a tension in discussing knowledge diff usion in the AI sphere.
On the one hand, the spectacular scientifi c advances are often taught at
universities and published in peer- reviewed journals, providing businesses
and government personnel with quick and easy access to frontier research.
Further, there is the migration of personnel across regions and countries as
the above examples of Robin Li and Qi Lu show. This suggests that knowl-
edge externalities are global in scope.
On the other hand, AI expertise has also tended to agglomerate in several
narrowly defi ned regions globally. As with other information technologies,
much of the expertise is in Silicon Valley. Berlin, Seattle, London, Boston,
Artifi cial Intelligence and International Trade 471
Shanghai, and to some extent Toronto and Montreal can all claim to be hubs
of AI innovation. This suggests that AI involves a lot of tacit knowledge
that is not easily codifi ed and transferred to others.
In fact, the traditional discussion of knowledge externalities takes on a
more nuanced hue in the context of AI. Can these researchers communicate
long distance? Do they have to be together? How important are agglomera-
tion forces in AI? As of 2017, AI expertise remains surprisingly rooted in the
locations of the universities that invented the technologies. Google’s Deep-
Mind is in London because that is where the lead researcher lived. Then the
fi rst expansion of DeepMind outside the United Kingdom was to Edmon-
ton, Alberta, because Richard Sutton, a key inventor of reinforcement learn-
ing, lives in Edmonton. Uber opened an AI offi
ce in Toronto because it
wanted to hire Raquel Urtasun, a University of Toronto professor.
Generally, there are a small number of main AI research departments:
Stanford, Carnegie Mellon University, the University of Toronto, and several
others. Their location is often surprisingly disconnected from headquarters,
and so companies open offi
ces where the talent is rather than forcing the
talent to move to where the company is.
As we shall see, the exact nature of knowledge externalities is terribly
important for understanding whether cluster and other policies are likely to
succeed. The nature of these externalities also has some unexpected implica-
tions such as the implications of noncompete clauses (Saxenian 1994) and
the asymmetries in access to knowledge created by asymmetries in who can
speak English versus who can speak Chinese versus who can speak both.
19.3 Trade Theory and the Case for Industrial
and Strategic Trade Policies
There are many voices in the industrialized world arguing for industrial
policies and strategic trade policies to promote rising living standards. Many
of these voices point to the achievements of China as an example of what
is possible. Much of what is claimed for China, and what was once claimed
for Japan, is of dubious merit. China has redirected vast resources from
the rural poor and urban savers toward state- owned enterprises that have
massively underperformed. Those fi rms continue to be major players in the
economy and a major drag on economic growth (Brandt and Zhu 2000). It is
thus signifi cant that China’s greatest commercial successes in AI have come
from private companies. So if we are to make the case for industrial and
strategic trade policies, we cannot blithely appeal to Chinese state- directed
successes. Rather, we must understand the characteristics of industries that
increase the likelihood that government policy interventions will be suc-
cessful.
To this end, we start with a vanilla- specifi c factors model of international
trade (Mussa 1974; Mayer 1974) in which the case for departures from free
472 Avi Goldfarb and Daniel Trefl er
trade is weak. We then add on additional elements and examine which of
these is important for policy success. The fi rst conclusion is that scale and
knowledge externalities are critical. The second is that these two elements
alone are not enough: their precise form also matters.
19.3.1 Scientists, Heterogeneous Scientists, and Superstar Scientists
Many factors enter into the location decisions of AI fi rms including access
to local talent, local fi nancing/ management, and local markets. In this sec-
tion, we focus on the role of university- related talent. Among the partici-
pants of this conference are three head researchers at top AI companies:
Geoff
rey Hinton (University of Toronto and Google), Russ Salakhutdi-
nov (Carnegie Mellon University and Apple), and Yann LeCun (New York
University and Facebook). Each joined his company while retaining his
academic position, and each continues to live near his university rather than
near corporate headquarters. These three examples are not exceptional, as
indicated by the above examples of DeepMind and Richard Sutton, and
Raquel Urtasun and Uber.
Scientists. We begin with the simplest model of trade that allows for two
types of employees, scientists, and production workers. There are two indus-
tries, search engines and clothing. Production workers are employed in both
industries and move between them so that their wages are equalized across
industries. Scientists are “specifi c” to the search engine industry in that they
are very good at AI algorithms and useless at sewing. We also assume that
scientists and workers cannot migrate internationally. Then it is immediately
obvious that the more scientists a country has, the larger will be both the size
and service exports of the search engine industry.
We start with this benchmark model because, in this setting, without scale
or externalities there is no scope for market failure and hence there is no
simple case for any trade policy other than free trade. For example, consider
a policy of restricting imports of search engine services, as China has done
with Google. This restriction helps Chinese scientists but can hurt Chinese
production workers and consumers (Ruffi
n and Jones 1977).
There are several departures from this benchmark model that lead to
welfare- enhancing export subsidies and other departures from free trade.
As we shall see, the two most important are economies of scale and knowl-
edge creation. However, we start instead with profi ts because profi ts are at
the core of arguments supporting strategic trade policies (Krugman 1986).
Since there are no profi ts in the specifi c factors model, we introduce profi ts
by introducing scientists of heterogeneous quality.
Heterogeneous Scientists. Consider an industry in which fi rms provide
a search engine and generate advertising revenue. There is a continuum of
scientists distinguished by their “quality” q. A fi rm is distinguished by the