Jason Furman's Reviews > AI Superpowers: China, Silicon Valley, and the New World Order

AI Superpowers several short books in one: a fantastic book about China’s recent internet innovation, a very good book about the current and near future of AI, a sometimes moving but mixed personal reflection memoir and a terrible set of predictions about the economics of AI. It also doesn’t ask some of the most fundamental questions one would want in a book like this.

Leading Chinese AI researcher and internet innovator (founder of Microsoft Research China, Google China, and now a leading venture firm Sinovention Ventures) Kai-Fu Lee sets out to compare the prospects of AI development in China and the United States. Many would give the United States the edge because of our greater degrees of creativity and cutting edge research whereas China, to date, has done more to copy, adapt and apply. Lee argues that this misses what is actually going on in AI: “the casual observer—or even expert analyst—would be forgiven for believing that we are consistently breaking fundamentally new ground in artificial intelligence research. I believe this impression is misleading. Many of these new milestones are, rather, merely the application of the past decade’s breakthroughs—primarily deep learning but also complementary technologies like reinforcement learning and transfer learning—to new problems... Much of the difficult but abstract work of AI research has been done, and it’s now time for entrepreneurs to roll up their sleeves and get down to the dirty work of turning algorithms into sustainable businesses.”

Lee argues that the success of this “dirty work” will require four inputs: (1) data, (2) entrepreneurs, (3) AI scientists, and (4) a policy-friendly environment. Taking these four in turn:

DATA: Lee argues that China is collecting much more data and willing to use it, including data that intersects between the real and online world. This seems overwhelmingly true and was not exactly a novel insight. The only open question is how important will data be in the future. My guess is very and Lee’s statement “Given much more data, an algorithm designed by a handful of mid-level AI engineers usually outperforms one designed by a world-class deep-learning researcher.” will remain true. But I don’t think we’re sure, as AlphaGo Zero, for example, was better than the original AlphaGo and used no data. Lee is thinking within the current deep learning paradigm and may be understating the importance of what comes next.

ENTREPRENEURS: This was the most fascinating and novel part of the book. Lee provides a capsule history of the internet in China, the copycat businesses, the failure of American businesses in China, and the emergence of much more innovative firms. At its heart is the argument that recent Chinese entrepreneurship arguing that it is more successful than the American variant because it is hungrier, willing to work harder, less distracted by any purpose other than moneymaking, and more willing to copy instead of innovate. A lot of this seems true but I think devalues the successes of Silicon Valley and underestimates the degree to which Chinese entrepreneurs may become more like that in the future. And perhaps for obvious and forgivable reasons, but some of this discussion misses the role of the Chinese policy—for example the book portrays the exit of Google from China as an unfortunate decision by the company not the result of censorship policies in China.

AI SCIENTISTS: Lee argues there are more and more of them in China, they’re excellent, but not as innovative as Geoffrey Hinton or Yann Le Cunn. But he argues that the innovations are not needed anymore. As discussed above, this is interesting, plausible, but may undervalue the paradigm changing innovations we can’t foresee.

AI POLICY: Lee argues that China has made a concerted policy effort to encourage AI while the United States has effectively ignored it. This is true. The question is how consequential it is. Lee gives the example of China using subsidies to create an AI/innovation cluster outside of Beijing to mimic the dense networks in Silicon Valley. But by analyzing only one success case Lee fails to understand what worked about the policy. Governments around the world have done similar cluster funding with nothing like China’s success, so this hardly seems necessary (Silicon Valley did without) or sufficient. Moreover, I would think future Chinese policy is a big risk to AI if continued clamp downs discourage innovation of certain types in China.

Lee’s analysis leads him to a make predictions about four areas of AI applications and how relative US-China capabilities today will shift over the next five to ten years (e.g., from 50-50 in internet applications to 60-40 in favor of China, 10-90 in autonomous applications today in favor of the US to 50-50, with similar trends in business applications and perception AI).

Lee’s predictions about China’s development in the future and the relative balance of capabilities seems broadly reasonable, subject to the caveats above. But he never addresses the fundamental question: is this zero sum or positive sum? The “Superpowers” of the title would make one think it was zero sum, that there is a first mover advantage and whoever, for example, figures out self-driving cars will own that industry permanently. In reality, in much of AI I think “positive sum” is likely to be the better model as innovations are disseminated, copied or reverse engineered. The entire relative capabilities frame makes much less sense in this positive sum world.

All of the above was the meat of the book and the reason it was much more interesting than even the summaries I read. I would recommend most readers just stop there, about 60% of the way through, and I wish I did too. Instead what follows is an absurd semi-apocalyptic argument that AI will take all our jobs, cause massive inequality, and lead to a huge economic gap between the US and China and the rest of the world.

The AI jobs discussion is economically rudimentary and Lee misses some major points (e.g., he acknowledges AI will create new types of jobs but seems to miss that richer people will want more of old types of jobs, eg will eat out more and thus more restaurant servers). The GDP discussion is similarly rudimentary (I would predict US and Australia GDP closer 50 years from now than US and China GDP, he seems to not understand how these innovations can be used by countries that didn’t make them and the notion of spillovers and convergence). More importantly, there is a huge tension between Lee’s skepticism about the imminence of general AI, his belief that we will not make future paradigm innovations but instead will continue to tinker and apply, and then his claim that existing modeling attempts understate the pace of AI progress.

Finally, Lee has a sometimes moving account of his own battle with cancer, how it changed his priorities towards more balance with life, and why this tells him that future AI policy should center not around UBI but subsidies for volunteering, caring, and other human interaction. This was sometimes moving, a little thought provoking, but also had the usual shortcoming of a highly successful person who can afford to make these types of statements without giving up any of that success.

Overall, I thought I knew the main points of this book from the reviews and that it would not add much. The first 60% of the book did indeed add a lot. Unfortunately, if you paid attention to the last 40% it would also subtract a lot so maybe just skip or skim that part.