Original article was published on Artificial Intelligence on Medium
Summary of Kai-Fu Lee’s “AI Superpowers”
Kai-Fu Lee is a celebrated computer scientist who had a long career at the leading technology companies of our time: Apple, Microsoft, and Google. He went on to become a venture capitalist, founding Sinovation Ventures, a fund focused on Internet startups in China. KFL was born in Taiwan to parents who originally emigrated from China. He himself emigrated as a boy to the US in 1973 (it was not unusual for Taiwan-based families to split up in the wake of the admission of China to the UN, replacing Taiwan, the Republic of China). He was an outstanding student and graduated from Columbia and Carnegie Mellon with a PhD in Computer Science.
In this book, KFL provides a unique perspective on the AI revolution, not only as a researcher and practitioner, but also as a business executive who has intimate knowledge of the two most important technology powers in the 21st century.
The History of AI
KFL recounts the interesting history of AI. He describes how by the 1980’s AI was divided into the rules-based approaches (aka “expert systems”) and neural networks. As it turned out, the rules-based approaches hit a dead end and lives on in specialized applications. He started in earnest on AI just as neural networks were in a resurgence. This resurgence was enabled by steady increases in computing power and data. Neural networks are modeled on the hierarchy of layers in the brain. The actual workings of neural nets are probably not be the same as brains, but they share some similarities, such as the ability to learn from examples, which are called “training data”. These supercharged neural nets resulted in an AI method called deep learning. KFL explains:
These algorithms use massive amounts of data … to make a decision that optimizes for a desired outcome. It [sic] does this by training itself to recognize deeply buried patterns and correlations connecting the many data points to the desired outcome. This pattern-finding process is easier when the data is labeled with that desired outcome — “cat” versus “no cat”; “clicked” versus “didn’t click”; “won game” versus “lost game.” It can then draw on its extensive knowledge of these correlations — many of which are invisible or irrelevant to human observers — to make better decisions than a human could. (pg. 16)
As an aside, at its core, deep learning is built on well-known mathematics that has been around since the 19th century: statistical inference, linear algebra, and numerical solutions of large sparse matrices. The difference is that the scale of the mathematical problems is far larger than anything attempted hitherto in other branches of engineering or science (with the exception of meteorology). Where a large, complex problem in the 1980’s might involve inverting sparse matrices with thousands of rows and columns, today deep learning requires inverting such matrices with millions of rows and columns! And doing so iteratively.
The Age of Implementation
Deep learning was invented in the US, UK, and Canada. But China didn’t wake up to it until 2016, when China’s Go champion was defeated by AlphaGo, an AI program. Since then, China has been on a crash program to catch up and lead the world. In next to no time, they’ve largely pulled abreast of the US. How did this happen so quickly even though the science was invented in the US and most of the important work was taking place in Western countries? The explanation is the actual science is fairly straightforward; it’s based on well-understood mathematics and computer science theories that any undergraduate student would understand.
What has changed is that we have now entered the age of implementation, where the theory and principles are well-known, it’s now a matter of application. Application does not require the highest calibre researchers, which disproportionately populate the best universities in the US, Europe and Canada. It does, however, require extremely capable PhD-level experts “to tweak complex mathematical algorithms, to manipulate massive amounts of data, to adapt neural networks to different problems”. (pg. 19)
Data is the Oil of the Digital Age
Furthermore, the algorithms require prodigious quantities of data and computing power, which China has in abundance. While American companies gather data about a person’s online behavior, Chinese companies are also gathering data about the real world, what people are doing, buying, eating, and so on. This data now powers the new age of AI.
Entrepreneurs in China
In addition, the startup culture of Chinese entrepreneurs is cutthroat and completely unsentimental about using any means fair or foul to best an adversary. These values were common in the Gilded Age era of American capitalism, but has since been tamed by regulatory, societal and other forces. Silicon Valley companies, for example, eschew business models that aren’t “highly scalable”; they look down on the inelegance of solving hard real world problems like delivery or owning depreciating assets. There is a bias towards building “platforms” where the hard work of creating content, managing customers, or engaging one-on-one with the real world is left to partners, contractors, or best of all, other platform users. In this way, AirBnb, Uber, Facebook and many other companies leave it to their users to do the heavy lifting, often for next to nothing. Few Silicon Valley companies try to emulate Amazon. And yet, Amazon is the model which many Chinese companies most closely follow. The advantage of doing the heavy lifting is as an organization you own the entire chain of value creation and all the data and insights derived.
Chinese companies are often derided as copycats and for stealing intellectual property. But this interpretation partly reflects a Western cultural bias, since copying in art, for example, is considered perfectly normal in East Asian cultures, with successive versions of art or architecture enjoying the same status as the “original”. In business, the prevalence of copying a successful product or business model is so pervasive and impossible to stop that firms don’t bother to deter competitors from doing so. Instead, they seek to innovate faster, build deeper moats around their businesses, and to tie in capabilities faster than their competitors. In this way, they extend and defend their products. Thus, competitors that copy just some aspects of a successful product cannot be successful because they do not attain the entire ecosystem or range of supporting products and services. In this way, Chinese companies think very much like the largest and most successful American companies, e.g., Apple and Amazon. These firms seek to create successful products, but also ecosystems and interconnected networks of products that make any given product in their portfolio much stronger than it would be on its own merits.
KFL asserts that the Chinese “willingness to get one’s hands dirty in the real world separates Chinese technology companies from their Silicon Valley peers. American startups like to stick to what they know: building clean digital platforms that facilitate information exchanges. Those platforms can be used by vendors who do the legwork, but the tech companies tend to stay distant and aloof from these logistical details.” (pg. 54) As examples, he cites Dianping/Meituan and Didi/Cuxing, which pioneered in food delivery and ride hailing by investing enough to build out and own the infrastructure in their verticals.
Finally, China’s national government actively supports AI and aims to be world’s leading AI country by 2030. As an example, in September 2014, China announced that it would pursue “mass entrepreneurship and mass innovation”, something that sounded platitudinous and Maoist in tone. What it signaled for government officials at the four levels of government in China is that new goals and measures were now in place and their success as bureaucrats would be measured against those goals. Almost overnight 6600 startup incubators around the country were launched! It was state capitalism in the raw! This redirection jump started a wave of startups and entrepreneurial innovation across the nation.
Putting together China’s strengths in implementation, data, entrepreneurship, government support, and owing a business vertical end-to-end, KFL believes China will match and then overtake the US in AI deployment, yielding tremendous productivity gains. “AI deployment will add $15.7 trillion to global GDP by 2030. China is predicted to take home $7 trillion of that total, nearly double North America’s $3.7 trillion in gains. As the economic balance of power tilts in China’s favor, so too will political influence and “soft power,” the country’s cultural and ideological footprint around the globe.” (pg. 23)
In the following graphic, KFL defines the 4 areas of AI and handicaps the two leaders; he sees the US maintaining its lead in Business AI:
KFL points out that the AI he refers to and that will generate vast opportunities for some companies, is actually quite narrow: it encompasses a few important areas of application: Internet, Business, Perception, and Autonomous. These AIs are not the same as nor are they likely to lead to what researchers call Artificial General Intelligence — the kind of intelligence a human has. AGI is the ability to learn without vast amounts of training data and to apply the sinews of intelligence from one area of expertise to another almost seamlessly, just like humans do. In his view and that of many other computer scientists, there is no credible timeline for AGI; it could happen in 20 years or 50. That said, fundamental research work is continuing in great American universities; if AGI happens in the next 20 years, it will likely emerge from a US institution. But, the AI that is scheduled to change the world in the next 10 years is not AGI; it’s the four branches mentioned above.
Most people are familiar with Internet AI: it powers Google, YouTube, Byte Dance, Amazon and many other websites, offering users recommendations, targeting content to users, and so on. In the next five years, KFL sees China with an advantage. The impact of this form is AI is largely in the digital, online world.
Businesses and other institutions have been collecting, curating and organizing data for decades. Banking, medicine and other data intensive applications will benefit from Business AI. Here, the US has a strong and likely insurmountable lead. It’s also the area where the most immediate profits can be made.
Traditional, offline businesses will be ripe for the revolution coming in Perception AI, which will start with capabilities like vision and speech recognition, but eventually include all kinds of smart devices to monitor the world around us, for us. AI and the devices it runs on will be embedded throughout our environment so that we won’t really think of AI as a separate thing, just like electricity has permeated most of our environment. Here, partly because of the strength of China’s devices industry and its vast amounts of real life data, it will emerge the strong leader.
Self-driving cars are the example that comes to most people’s minds when considering Autonomous AI. But it will encompass people-free factories, seamless retail checkout, people-free warehouses, and so on. In this branch of AI, KFL sees a toss-up; while China has its vast repositories of data and implementation skills, Silicon Valley and the leading technology companies in the US have a significant head start and no illusions about the importance of global competition.
The Problems Created by AI
KFL identifies two main problems created by AI. First, the concentration of AI in the US and China will relegate other countries even those in the developed, rich world to second tier status (i.e., the Level 4 countries in Rosling parlance¹), as they miss out on much of the wealth creation from the new technologies. While they can benefit from the breakthroughs as consumers of the breakthroughs, they will lack the ability to steer the technology and will become vassals of the AI giants in China and the US. Their firms will be driven out of business or colonized. Worse still, the developing world will not be able to climb onto this ladder as earlier countries did by leveraging their low cost labor; the value of that labor will be automated out of existence.
Second, within the US and China (and also in other countries), AI will replace what has been the mainstay of work and professions for most of the middle class. It will also sever the bridges by which the poor or new immigrants move up economically, thus impeding social mobility. This eventuality portends a societal crisis.
General Purpose Technologies
The author admits that technological progress has always led to greater wealth and more job, even as much work and many professions were eliminated. However, general purpose technologies (GPT) are qualitatively different. Among GPTs are technologies like the steam engine and electrification. AI is the third. What characterizes GPTs is that they remove skill and craftsmanship from tasks, simplifying manufacturing, and reducing production to repetitive tasks performed by relatively lower skilled workers. This trend has increased output, while holding down wages. The AI revolution will do all that and even more, so that workers will be almost entirely unnecessary. The ones that remain will find their labor worth less and less. In fact, by 2030, most of the $15.7 trillion gains in productivity will go to China and the US and within those countries to a small group of people.
In the following graphic, KFL plots the types of professions and jobs that will be at greater (and lesser) risk through the proliferation of AI as a GPT.
A key insight KFL makes is that when automation of existing work and workflows principally performed by humans is considered, the impact of AI is modest, perhaps eliminating 10–20% of activity. But he states that as AI is adopted, entire workflows designed around humans as workers and processors will be reinvented, eliminating not just a few tasks in an existing workflow, but the entire workflow itself. He calls these “ground-up replacements” rather than “one-to-one replacements”. As an example, he cites ByteDance, which is a technology that has eliminated editorial curation entirely, not just assisting editors in a pre-existing workflow. He summarizes the impact: “Within ten to twenty years, … we will be technically capable of automating 40 to 50 percent of jobs in the United States. For employees who are not outright replaced, increasing automation of their workload will continue to cut into their value-add for the company, reducing their bargaining power on wages and potentially leading to layoffs in the long term. We’ll see a larger pool of unemployed workers competing for an even smaller pool of jobs, driving down wages and forcing many into part-time or “gig economy” work that lacks benefits.” (pg. 147)
As his comparison of the strengths of China and the US and his description of a dystopian future for most people reach a crescendo, KFL pivots suddenly to reveal that in September 2013, he was diagnosed with stage IV lymphoma. He describes with great pathos his comeuppance, laying bare the admitted shallowness of how he lived life to date, and how he had neglected the things that would ultimately matter the most: his family and their love of each other. With feeling and honesty, he talks about his mother, wife, children, sister and all they have done for him, expecting nothing in return — and how he committed only enough time and energy to maintain, but not invest in the relationships. “Like so many people forced to suddenly face their own mortality, I was filled with fear for my future and with a deep, soul-aching regret over the way I had lived my life.” (pg. 157)
Eventually, he recovers; the cancer goes into remission. His diagnosis was typical of many medical diagnoses, based on simple rules that doctors can readily apply, but which often ignored confounding, extenuating, and mitigating factors. Indeed he uses this opportunity to show how AI could have done a better job than most clinicians because an algorithm can learn and apply far more rules about a disease and its attributes to arrive at a more complete and subtle diagnosis and prognosis. And, it turned out the diagnosis of the advanced stage of his cancer was a first order analysis; subsequently, the identification of extenuating conditions (which he played no small role in) made his recovery much more likely.
In a skillful turn, he uses this life-changing experience to motivate the denouement — to arrive at the solution to the problem of AI.
KFL returns to his VC roots to propose similar mechanism to address the problems wrought by AI: service-focused social impact investing, funded by investors and government who measure returns differently than traditional capitalists and are content with linearly growing rather than exponential returns. The outcome would be investments meant to generate jobs that improve human happiness and satisfaction; for example, to create work for people who assist the elderly, especially with tasks that cannot be automated. Often, these jobs are done by volunteers or others who are not financially compensated (or if so, very poorly). In this way, humans will do what they are best at: loving others, showing empathy, talking to people, and so on, while AI and machines do the boring, dangerous, repetitious, complex jobs.
Finally, KFL points out that even as we look at AI as a competition between nations (his book after all seems to frame it that way), the threat isn’t so much to nations but to people everywhere, even in China and the US. So, the way forward is greater cooperation at the national level — even as individual companies compete with each other in the global marketplace.