AI Superpowers by Kai-Fu Lee

11 Mar 2019

Deep learning will eliminate many jobs that were once considered safe from automation. White collar, blue collar… every profession will feel the impact before the year 2039.

Machines powered by artificial intelligence have excelled at games for decades. IBM’s Deep Blue defeated chess champion Garry Kasparov in 1996. Alphabet’s AlphaGo defeated world champion Go player, Ke Jie, in 2017.

Game-focused AI is highly specialized and mostly ineffective in in other areas. What about general purpose AI, software that can mimic human intelligence and make human-like decisions? What about applications beyond games?

Dr. Kai-Fu Lee examines these questions in his book, AI Superpowers: China, Silicon Valley, and the New World Order. Lee draws on his background as a pioneering AI researcher and successful entrepreneur to explain why certain subsets of AI produce results, and how readers of the book might benefit.

Rules, Neural Nets, and Deep Learning

In their quest to mimic human intelligence, AI researchers in the 1980s pursued two main paths.

• Rule-based systems. Also known as symbolic systems or expert systems. Long chains of if-then-else conditions. In order to build a rules-based system, the developer must have access to subject matter experts with intimate knowledge of the problem domain. And then they must work arduously to predict every possible stimulus and response. Broad, unpredictable domains are more difficult to handle this way.

• Neural networks. Rather than try to think of every possible scenario in advance, the developer exposes the neural net to a wide variety of scenarios that force the system to “learn” the desired behavior over time. For example, when teaching a neural net to recognize a picture of a cat, the developer might show the system a large number of cat pictures (labeled “cat”) and a large number of pictures without cats (labeled “not-cat”). After enough iterations, the neural net learns how to classify the photos correctly.

Neural networks fell out of favor in the 1990s because computer hardware was not powerful enough to make them work well. However, the technology has recently re-emerged under a new moniker, deep learning networks.

The most advanced deep learning networks learn from each other. For example, AlphaGo Zero, a newer version of the system that beat the human Go champion, learned to play the game by competing against other instances of itself. As with humans, the toughtest competitors emerge from the fiercest battles.

Deep Learning Re-Emerges

Of all the AI subsets, Lee believes that deep learning will generate the largest ROI in the next few years. His belief is based on decades of experience and these observations: Today, our computers are powerful enough and our algorithms are good enough.

Advancements in computer vision and speech recognition have given our machines new ways to collect massive amounts of data. Cameras are cheap and ubiquitous. Speech recognition is so inexpensive and so good that Amazon/Google can give it away for (almost) free and turn a profit when customers buy goods through the voice interfaces. And let us not forget the data. Deep learning needs massive amounts of data in order to work. Data collected by so many cameras and microphones is more valuable than petroleum.

Deep Learning and Human Careers

Deep learning will wreak havoc with careers. Human jobs that require pattern recognition will be replaced by deep learning tools. For example:

• Driving. When Lee was in graduate school, most researchers believed that we would consider a computer “intelligent” if it could drive a car. Over time, our scientists reframed the car driving problem as a challenge involving pattern recognition and appropriate responses. And while better algorithms were being developed, the price-to-performance ratio of computer hardware continued to improve. For safety reasons, today’s self-driving cars must have a human back-up driver. But full autonomous driving is within our grasp.

• Law. As recently as a few decades ago, the newest lawyers on a legal team might be assigned the task of discovery, sifting through mountains of paper documents looking for strings of words relevant to the case. This pattern recognition problem is better handled by a deep learning system that reads flawlessly without getting tired. A character from Shakespeare advised his companions to “kill all the lawyers.” Deep learning would be a happier alternative for most members of the bar.

• Medicine. Physicians who practice radiology spend years in medical school developing their skills. The experienced eyes of a radiologist can tell the difference between a life-threatening growth and something benign. Of course, this is another example of a pattern recognition problem. A computer vision algorithm that can accurately identify faces in a crowd can also scan x-ray images for health issues. Further, the algorithm can draw on every piece of medical information stored in any shared database on the planet, in an instant. And the algorithm never gets tired.

Bottom line: Deep learning will eliminate (or drastically alter) many professions that were once considered safe from any sort of mechanical automation. Lee predicts that 40% of jobs worldwide will be displaceable by deep learning and related technologies by 2039.

Deep Learning in China

As the founder of a VC firm operating in China, Lee offers some interesting thoughts about how Chinese companies will perform in an era of deep learning.

Deep learning thrives when given two key resources: Processing power and large amounts of data. The processing power is here, today. But data can be tougher to get, especially in the United States where privacy concerns are part of the culture.

But in China, the data required to train deep learning algorithms is plentiful and largely untethered by privacy laws. Algorithms designed to learn from Chinese consumer behavior have plenty of data - a big advantage. Lee shares examples of companies, based in China, that were once considered imitators of their US-counterparts. But in most cases, the Chinese company has grown larger than the US company it formerly emulated, thanks to deep learning and a mountain of data.

People Displaced by Deep Learning? What’s Next?

Revisiting the concern about displaced careers: Since we know that the wave of deep learning is coming, what should we do?

Some people will argue that we’ve been through changes like this before. Whenever humans develop new, labor-saving technology, new jobs emerge that we never would have dreamed of before. How many readers of this blog work in careers that didn’t exist twenty years ago?

Lee argues that the deep learning wave will be different in two ways:

• Speed. Never in human history have we displaced 40% of our jobs, worldwide, in the span of a few decades.

• Range. When farmers were displaced by farm automation, things worked out because that was during the industrial revolution, which opened factory opportunities for the former farm workers. One segment reduced in size as a different segment emerged. But this time, deep learning will touch every economic level, blue collar and white collar, college educated and trade-school grad.

Even deeper, so many of us, especially in the USA, define ourselves by our careers. What will it mean to be human when career-based definitions have evaporated?

Lee observes that the shift may give us an opportunity to connect with a uniquely human gift in new and positive ways. Lee believes that we will re-connect with our ability to love.

Lee shares more about the path that led to this way of thinking in the book.

About Kai-Fu Lee

For deeper insight into AI Superpowers, it might be helpful to know more about the author.

Dr. Kai-Fu Lee is a PhD computer scientist and founder of Sinovation Ventures, a $2 billion technology investment firm with offices in Beijing, Shanghai, Shenzhen, Seattle, and Silicon Valley.

Lee was born in China. He was sent to the USA at the age of eleven because his mother wanted him to be educated beyond the rote memorization curriculum that was available in the town of his birth. To maintain his Chinese writing skills, Lee’s mother required him to write to her weekly in Chinese. And she would return each of his letters to him with corrections. Motherly encouragement paid off. Today, Lee is fluent in Chinese and English and he flows easily between the cultures.

Lee holds degrees in Computer Science from Columbia University (A.B.) and Carnegie Mellon (PhD). He pioneered speech recognition research at Apple. Later, he served in technical leadership positions at SGI, Microsoft, and Google. He founded Sinovation Ventures in 2009.

Conclusion

AI Superpowers offers insights that go beyond technology. The author considers economic, social, and emotional concerns that are rarely discussed in a business book. Anyone interested in AI, business, or the interaction between Chinese and American industries will find the book valuable.