The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 67
The Economics of Artificial Intelligence Page 67

by Ajay Agrawal

able cost (renting time on the data center). An organization can purchase

  virtually any amount of cloud services, so even small companies can start at

  a minimal level and be charged based on usage. Cloud computing is much

  more cost eff ective than owning your own data center, since compute and

  data resources can be purchased on an as- needed basis. Needless to say,

  most tech start-ups today use a cloud provider for their hardware, software,

  and networking needs.

  Cloud providers also off er various machine- learning services such as

  voice recognition, image recognition, translation, and so on. These systems

  are already trained by the vendor and can be put to immediate use by the

  customer. It is no longer necessary for each company to develop its own

  software for these tasks.

  Competition among the cloud providers is intense. Highly detailed and

  specifi c image recognition capabilities are off ered at a cost of a tenth- of-a-

  cent per image or less, with volume discounts on top of that price.

  A user may also have idiosyncratic data relevant to its own business like

  the point- of-sale data mentioned above. The cloud provider also provides

  up- to-date, highly optimized hardware and software than implements

  popular machine- learning algorithms. This allows the use immediate access

  to high- powered tools . . . providing that they have the expertise to use them.

  If the hardware, software, and expertise are available, all that is needed is

  the labeled data. There are a variety of ways to acquire such data.

  • As By- Product of Operations. Think of a chain of restaurants where

  some perform better than others, and management may be interested in

  factors that are associated with performance. Much of the data in the

  Kaggle competitions mentioned above are generated as a byproduct of

  day- to-day operations.

  • Web Scraping. This is a commonly used way to extract data from web-

  sites. There is a legal debate about what exactly is permitted with respect

  to both the collection of data and how it is used. The debate is too com-

  plex to discuss here, but the Wikipedia entry on Web scraping is good.

  An alternative is to use data that others have scraped. For example, the

  Common Crawl database contains petabytes of data compiled over

  eight years of Web crawling.

  • Off ering a Service. When Google started its work on voice recognition,

  it had no expertise and no data. It hired the expertise and they came up

  404 Hal Varian

  with the idea of a voice- input telephone directory as a way to acquire

  data. Users would say “Joe’s Pizza, University Avenue, Palo Alto” and

  the system would respond with a phone number. The digitized question

  and the resulting user choices were uploaded to the cloud and machine

  learning was used to evaluate the relationship between Google’s answer

  and the user action—for example, to call the suggested number. The

  ML training used data from millions of individual number requests and

  learned rapidly. ReCAPTCHA applies a similar model where humans

  label images to prove they are human and not a simple bot.

  • Hiring Humans to Label Data. Mechanical Turk and other systems can

  be used to pay people to label data (see Hutson 2017).

  • Buying Data from Provider. There are many providers of various sorts

  of data such as mail lists, credit scores, and so on.

  • Sharing Data. It may be mutually advantageous to parties to share

  data. This is common among academic researchers. The Open Images

  Data set contains about nine million labeled images contributed by

  universities and research labs. Sharing may be mandated for a variety

  reasons, such as concerns for public safety. Examples are black boxes

  from airplanes or medical data on epidemics.

  • Data from Governments. There are vast amounts of data available

  from governments, universities, research labs, and nongovernmental

  agencies.

  • Data from Cloud Providers. Many cloud providers also provide public

  data repositories. See, for example, Google Public Data sets, Google

  Patents Public Data set, or AWS Public Data sets.

  • Computer- Generated Data. The Alpha Go 0 system mentioned earlier

  generated its own data by playing Go games against itself. Machine-

  vision algorithms can be trained using “synthetic images,” which are

  actual images that have been shifted, rotated, and scaled in various

  ways.

  16.2.3 Important Characteristics of Data

  Information science uses the concept of a “data pyramid” to depict the

  relationship between data, information, and knowledge. Some system has

  to collect the raw data, and subsequently organize and analyze that data

  in order to turn it into information—something such as a textual docu-

  ment image that can be understood by humans. Think of the pixels in an

  image being turned into human- readable labels. In the past this was done

  by humans; in the future more and more of this will be done by machines.

  (See fi gure 16.1.)

  This insights from the information can then turned into knowledge, which

  generally is embodied in humans. We can think of data being stored in bits,

  information stored in documents, and knowledge stored in humans. There

  are well- developed markets and regulatory environments for information

  Artifi cial Intelligence, Economics, and Industrial Organization 405

  Fig. 16.1 The information pyramid

  (books, articles, web pages, music, videos) and for knowledge (labor markets,

  consultants). Markets for data—in the sense of unorganized collections of

  bits—are not as developed. Perhaps this is because raw data is often heavily

  context dependent and is not very useful until it is turned into information.

  Data Ownership and Data Access

  It is said that “data is the new oil.” Certainly, they are alike in one respect:

  both need to be refi ned in order to be useful. But there is an important dis-

  tinction: oil is a private good and consumption of oil is rival: if one person consumes oil, there is less available for someone else to consume. But data

  is nonrival: one person’s use of data does not reduce or diminish another

  person’s use.

  So instead of focusing on data “ownership”—a concept appropriate for

  private goods—we really should think about data access. Data is rarely

  “sold” in the same way private goods are sold, rather it is licensed for specifi c

  uses. Currently there is a policy debate in Europe about “who should own

  autonomous vehicle data?” A better question is to ask “who should have

  access to autonomous vehicle data and what can they do with it?” This for-

  mulation emphasizes that many parties can simultaneously access autono-

  mous vehicle data. In fact, from the viewpoint of safety it seems very likely

  that multiple parties should be allowed to access autonomous vehicle data.

  There could easily be several data collection points in a car: the engine, the

  navigation system, mobile phones in rider’s pockets, and so on. Requiring

  exclusivity without a good reason for doing so would unnecessarily limit

  what can be done with the data.

  Ross An
derson’s description of what happens when there is an aircraft

  406 Hal Varian

  crash makes an important point illustrating why it may be important to

  allow several parties to access data.

  When an aircraft crashes, it is front page news. Teams of investigators

  rush to the scene, and the subsequent enquiries are conducted by experts

  from organisations with a wide range of interests—the carrier, the insurer,

  the manufacturer, the airline pilots’ union, and the local aviation author-

  ity. Their fi ndings are examined by journalists and politicians, discussed

  in pilots’ messes, and passed on by fl ying instructors. In short, the fl y-

  ing community has a strong and institutionalised learning mechanism.

  (Anderson 1993)

  Should we not want the same sort of learning mechanism for autonomous

  vehicles? Some sorts of information can be protected by copyright. But in

  the United States, raw data such as a telephone directory is not protected

  by copyright. (See Wikipedia entry on the legal case Feist Publications, Inc

  v. Rural Telephone Service Co.)

  Despite this, data providers may compile some data and off er to license on

  certain terms to other parties. For example, there are several data companies

  that merge US census data with other sorts of geographic data and off er

  to license this data. These transactions may prohibit resale or relicensing.

  Even though there is no protectable intellectual property, the terms of the

  contract form a private contract that can be enforced by courts, as with any

  other private contract.

  Decreasing Marginal Returns

  Finally, it is important to understand that data typically exhibits decreas-

  ing returns to scale like any other factor of production. The same general

  principle applies for machine learning. Figure 16.2 shows how the accuracy

  of the Stanford dog breed classifi cation behaves as the amount of training

  data increases. As one would expect, accuracy improves as the number of

  training images increases, but it does so at a decreasing rate.

  Figure 16.3 shows how the error rate in the ImageNet competition has

  declined over the last several years. An important fact about this competition

  is that the number of training and test observations has been fi xed during

  this period. This means that the improved performance of the winning sys-

  tems cannot depend on sample size since it has been constant. Other factors

  such as improved algorithms, improved hardware, and improved expertise

  have been much more important than the number of observations in the

  training data.

  16.3 Structure of ML- Using Industries

  As with any new technology, the advent of machine learning raises several

  economic questions.

  Fig. 16.2 Machine- learning adoption by economic sector

  Source: http:// vision.stanford .edu/ aditya86/ ImageNetDogs/.

  Fig. 16.3 Imagenet image recognition

  Source: Eckersley and Nasser (2017).

  408 Hal Varian

  Fig. 16.4 Number of AI- related technologies adopted at scale or in a core part of

  the business

  Source: McKinsey (2017).

  • Which fi rms and industries will successfully adopt machine learning?

  • Will we see heterogeneity in the timing of adoption and the ability to

  use ML eff ectively?

  • Can later adopters imitate early adopters?

  • What is the role of patents, copyright, and trade secrets?

  • What is the role of geography in adoption patterns?

  • Is there a large competitive advantage for early, successful adopters?

  Bughin and Hazan (2017) recently conducted a survey of 3,000 “AI

  Aware” C- level executives about adoption readiness. Of these executives,

  20 percent are “serious adopters,” 40 percent are “experimenting,” and

  28 percent feel their fi rms “lack the technical capabilities” to implement ML.

  McKinsey identifi es key enablers of adoption to be leadership, technical

  ability, and data access. Figure 16.4 breaks down how ML adoption varies

  across economic sectors. Not surprisingly, sectors such as telecom, tech, and

  energy are ahead of less tech- savvy sectors such as construction and travel.

  16.3.1 Machine Learning and Vertical Integration

  A key question for industrial organization is how machine- learning tools

  and data can be combined to create value. Will this happen within or across

  corporate boundaries? Will ML users develop their own ML capabilities or

  purchase ML solutions from vendors? This is the classic make versus buy

  Artifi cial Intelligence, Economics, and Industrial Organization 409

  question that is the key to understanding much of real- world industrial

  organization.

  As mentioned earlier, cloud vendors provide integrated hardware and

  software environments for data manipulation and analysis. They also off er

  access to public and private databases, provide labeling services, consulting,

  and other related services that enable one- stop shopping for data manipula-

  tion and analysis. Special- purpose hardware provided by cloud providers

  such as GPUs and TPUs have become key technologies for diff erentiating

  provider services.

  As usual there is a tension between standardization and diff erentiation.

  Cloud providers are competing intensely to provide standardized environ-

  ments that can be easily maintained. At the same time, they want to provide

  services that diff erentiate their off erings from competitors.

  Data manipulation and machine learning are natural areas to compete

  with respect to product speed and performance.

  16.3.2 Firm Size and Boundaries

  Will ML increase or decrease minimum effi

  cient scale? The answer de-

  pends on the relationship between fi xed costs and variable costs. If fi rms

  have to spend signifi cant amounts to develop customized solutions to their

  problems, we might expect that fi xed costs are signifi cant and fi rm size must

  be large to amortize those costs. On the other hand, if fi rms can buy off -

  the- shelf services from cloud vendors, we would expect that fi xed costs and

  minimum effi

  cient scale to be small.

  Suppose, for example, that an oil change service would like to greet return-

  ing customers by name. They can accomplish this using a database that joins

  license plate numbers with customer names and service history. It would be

  prohibitively expensive for a small provider to write the software to enable

  this, so only the large chains could provide such services. On the other hand,

  a third party might develop a smartphone app that could provide this ser-

  vice for a nominal cost. This service might allow minimum effi

  cient scale to

  decrease. The same considerations apply for other small service providers

  such as restaurants, dry cleaners, or convenience stores.

  Nowadays new start-ups are able to outsource a variety of business pro-

  cesses since there are a several providers of business services. Just as fast-

  food providers could perfect a model with a single establishment and then

  go national, business service companies can build systems once and replicate

  them globally.

&n
bsp; Here is a list of how a start-up might outsource a dozen business pro-

  cesses.

  • Fund your project on Kickstarter.

  • Cloud cloud computing and network from Google, Amazon, or Micro-

  Soft.

  • Use open- source software like Linux, Python, Tensorfl ow, and so forth.

  410 Hal Varian

  • Manage your software using GitHub.

  • Become a micromultinational and hire programmers from abroad.

  • Set up a Kaggle competition for machine learning.

  • Use Skype, Hangouts, Google Docs, and so forth for team communi-

  cation.

  • Use Nolo for legal documents (company, patents, NDAs).

  • Use QuickBooks for accounting.

  • Use AdWords, Bing, or Facebook for marketing.

  • Use ZenDesk for user support.

  This is only a partial list. Most start-ups in Silicon Valley and SOMA

  avail themselves of several of these business- process services. By choos-

  ing standardizing business processes, the start-ups can focus on their core

  competency and purchases services as necessary as they scale. One would

  expect to see more entry and more innovation as a result of the availability

  of these business- process services.

  16.3.3 Pricing

  The availability of cloud computing and machine learning off ers lots of

  opportunities to adjust prices based on customer characteristics. Auctions

  and other novel pricing mechanisms can be implemented easily. The fact

  that prices can be so easily adjusted implies that various forms of diff erential

  pricing can be implemented. However, it must be remembered that custom-

  ers are not helpless; they can also avail themselves of enhanced search capa-

  bilities. For example, airlines can adopt strategies that tie purchase price to

  departure date. But services can be created that reverse- engineer the airline

  algorithms and advise consumers about when to purchase (see, e.g., Etzioni

  et al. (2003). See Acquisti and Varian (2005) for a theoretical model of how

  consumers might respond to attempts to base prices on consumer history

  and how the consumers can respond to such attempts.

  16.3.4 Price Diff erentiation

  Traditionally, price diff erentiation has been classifi ed into three categories:

  1. First degree (personalized),

  2. second degree (versioning: same price menu for all consumers, but

 

‹ Prev