The Duck of Minerva

The U.S. Versus China . . . Versus the Rest?

24 July 2018

This post comes from Steve Weber, Professor at the I-School and Department of Political Science and Director of the Center for Long-Term Cybersecurity at the University of California, Berkeley, and a co-director of the Bridging the Gap project.

It has become common in 2018 to hear that the United States and China are locking themselves into an Artificial Intelligence ‘arms race’. While global politics will certainly change in the machine learning era, the supposed ‘arms race’ between the US and China may turn out to be less interesting and relevant in this world than the relationships between the two machine learning superpowers and everyone else.

Which race will prove more relevant depends upon the long-term economic and security consequences of general purpose technologies, as well as the distinctive characteristics of the technologies that fall under the AI umbrella. (I prefer the term ‘machine learning’ because it carries fewer science-fiction connotations.) General purpose technologies are technologies that sweep across the economy and impact what is possible in many sectors, shaking up how companies and governments do what they do in the broadest sense. Steam locomotion is the obvious 19th century example. Machine learning is a 21st century general purpose technology because it can (and will) be applied in just about every economic production process you can imagine, from retail management to autonomous driving to drug discovery and beyond.

An even more important characteristic of machine learning as a technology is that it has strong first mover advantages and positive feedback loops. In simple terms, the better you are at machine learning at any given moment, the faster you are likely to improve relative to those ‘behind’ you. A firm that has excellent machine learning products (say, a great map application) will find that its products have greater success in the market. The more people who use the product, the more data are created for the firm to work with, which should lead to faster improvement in the underlying algorithms. In turn, that means the next iteration of the product will be even better. This positive feedback cycle can run on a very fast cadence, since data products can be updated far more frequently than any physical product (some are updated daily or even more frequently than that). All of this implies that the leader should speed away from competitors at an ever-accelerating pace. Michael Horowitz recently examined in the Texas National Security Review the potential military implications of such first-mover advantages in AI.

This simple model has a few limitations and caveats. The first and most obvious is that no positive feedback loop like this runs forever. But without a clear argument as to why, when, and how it would diminish or reverse, there’s justification for concern about natural monopolies, with real consequences. The logic could apply to firms or to countries, depending on how data and algorithms flow (or don’t) across corporate and national borders.

This positive feedback is not only a function of data flows. At least as important are the complementary growth effects that would further enable the loop, probably the most significant of which right now is human capital. Think about it this way: if the most sophisticated data products are being built in Firm A, then it becomes much easier for Firm A to attract the best data scientists and machine learning experts to work there. Not only can Firm A offer a great salary, it can offer the best data sets, the most inspirational scientific challenges, the largest potential distribution for a great product, and the opportunity to work with the other ‘best’ data scientists in the world. Other economic complements of a modern information technology eco-system would follow — including basic research, venture capital, legal and marketing services, and so on.

So the algorithm economy is almost the epitome of a ‘learn by doing’ system with positive spillovers and other cluster economy effects. And it is special in two more relevant ways. The first is the importance of what Paul Romer called ‘meta-ideas’, which are ideas about how to support the production and sale of other ideas. (Romer said, for example, that allowing people to live together in dense urban agglomerations has been one of the biggest meta-ideas of modern life.) In the context of machine learning, the meta-ideas that can keep the positive feedback loop going lie in finding the right answers to questions like: What is the best means of managing the intellectual property content of algorithms and software code? What are the most effective labor market institutions that can support the growth of algorithm-driven labor demand? The important point is that supportive meta-ideas are more likely to emerge in countries and firms that are already ahead in the data economy, because they are not going to be discovered by a great mind sitting in a quiet room. They are going to emerge through adaptive practice in markets, which means that meta-ideas as well are ‘learning by doing’ in character.

The second is the distinctive non-physical nature of machine learning products and what that means over time. The law of entropy doesn’t really apply to algorithms. Algorithms don’t wear out (like a pair of sneakers) or fatigue (like metal in an airplane) or rust (like a bridge) or deteriorate with age (like a human body). Data products have escaped the recalcitrance of the physical world; they tend to get better the more they are used. This has particularly interesting implications for any product that is a combination of physical and algorithmic elements, like an autonomous vehicle. The tires will suffer from aging—but the exact opposite is true for the software. The car gets better at what it does over time; and that increase in value is almost certainly going to benefit whoever supplies the parts that improve rather than those that wear out. Tires are commodities; autonomous vehicle algorithms are not.

These dynamics lies at the heart of how countries will pursue economic growth and security in parallel over the next decade. The big underlying question is how countries should think about data flows in the context of globalization and economic growth. Growth theory has long grappled with the consequences of cross-border flows of goods, services, ideas, and people. But the most significant growth in cross-border flows now comes in the form of data. And — just as they did with flows of goods and services — countries are starting to express concern about the nature of those data flows, including imbalances in ‘imports’ and ‘exports’, as well as the differences between ‘raw’ data (like raw materials) and high value data products.

The positive feedback dynamics of machine learning suggests that they are right to do so, in contrast to perspectives that see data flows as a rising tide that lifts all boats. Countries need a clear perspective here in order to make successful and adaptive policy choices on issues like data localization, subsidies for national data champions, foreign direct investment restrictions, and other controversial issues of the day.

But as challenging as these decisions are for the peer competitors US and China, they are even tougher problems for advanced industrial economies (like many European states) that are behind the leaders in machine learning. Can France or Italy reasonably expect to spawn companies that could compete with Google or Baidu? It’s hard to see how that would happen in the absence of historic subsidies, protection, and other national ‘industrial policy’ equivalents for the data era. Are they better off simply accepting a role as ‘consumer’ of high value-added data products? Could either country build state-of-the-art military equipment when the most important ingredients are the algorithms and not the steel?

And what about emerging economies and developing countries who are falling even further behind? One of the most striking personal observations I’ve had over the last decade through my economic development advisory work in the Middle East, in particular, is how few emerging economies have a plausible vision and strategy for growth that goes beyond the low-wage manufacturing ladder that worked for a few high-visibility cases in the mid to late 20th century (like Japan, Korea, China, and possibly Vietnam). It was possible at that time to climb from simple manufacturing to more complex and advanced manufacturing — if a lot of things went right. It is much harder for me to see how a country could move from low wage manufacturing to machine learning and algorithmic production. It is much easier to see in today’s global economy how even low-wage manufacturing is likely to be eviscerated by capital-intensive robotic and algorithmic production — some is already being ‘re-shored’ to the US, regardless of presidential trade politics and the like.

Put these ingredients together, and the international political economy landscape becomes much more interesting than a simple two-player superpower arms race. It becomes a place where the superpowers could plausibly see themselves accelerating away from everyone else for a decade or more. Convergence and catch-up dynamics recede into the background. The stories that matter become those about alliance politics, choosing sides, sharing and distributing advanced technologies among your allies . . . and perhaps attacking (in a digital manner, perhaps through the manipulation of data sets and adversarial machine learning techniques) the positive feedback algorithm loops of others. Right now is a good time for the academic international relations community to start building those stories.