Human Compatible: Artificial Intelligence and the Problem of Control

I've just finished reading this book (actually, a cheaper advance uncorrected proof) by Stuart Russell, which represents a continuation of my interest in AI as reflected in the books I discussed in 2017 and 2018 by Max Tegmark and Luke Dormehl. Russell is a prominent AI researcher at UC Berkeley, so his perspective is a little different from that of Tegmark, who is a physicist, and Dormehl, who is a journalist. I found Russell more focused on the important issues at hand, but some of the chapters weren't of much interest to me and I consequently spent little time on them. I got the sense that academic specialization has impeded Russell's perspective a little (it usually does), though he did manage to convey some of the urgency more than Tegmark did in his book and much more than Dormehl did in his.

As the title suggests, the ascent of AI poses new problems for humans that are best addressed now rather than later. Russell uses the analogy that if we knew that an asteroid was going to strike the earth in a few years, we wouldn't wait for its arrival to start planning. The main question is how to prevent AI from becoming dangerous to mankind as technology approaches Artificial General Intelligence (AGI), or, more loosely, superintelligence. I am not a programmer and don't care about programming, so I didn't pay much attention to those chapters. However, Russell also traces the history of AI and compares it to developments that have taken place in other scientific fields. He mentions a story about the physicists Ernest Rutherford and Leo Slizard. In 1933, Slizard read an article by Rutherford stating that although it was known that matter contains a great deal of energy, there was no practical way to extract it. The same day, Slizard thought of the idea of a nuclear chain reaction, which led directly to the first man-made, self-sustaining nuclear chain reaction in 1942. Russell thinks that although the path to AGI is not immediately clear, a new idea that makes it possible could suddenly emerge at any time. So much AI research is occurring at the moment that it would not be surprising at all to have AGI appear within a few decades. Russell believes that the main pieces for its development are already in place. The necessary hardware exists and the use of probabilistic, self-learning algorithms is currently producing useful results.

Most of the problems related to AGI that are addressed in this book concern control. Sections are devoted to finding ways of incorporating human values into algorithms. Russell casts a wide net, which I think is far wider than necessary. For example, he considered ideas from economics and philosophy that I don't think are relevant. I liked the fact that he disagreed with Stephen Pinker, who thinks that AI doesn't pose a risk. The main risk is that AGI will be far more intelligent than any human ever could be and that if it has goals that do not appropriately value human well-being, the consequences could be disastrous. The problem consists mainly of incorporating the right human-friendly behavior into the algorithms and enforcing such procedures in national and international agreements. From a formal standpoint, the agreements would be similar to current international agreements on potentially dangerous gene-editing techniques such as CRISPR.

Although Russell covers his bases well and discusses what life might be like if a safe AGI arrives, that is not the main theme of the book. I think that by focusing on the risk to human existence, he doesn't devote enough space to changes that might occur in favorable circumstances. Even while stating that this would be the most significant event in human history, he doesn't speculate much on how we would react to it. He mentions Ray Kurzweil's idea of merging human brains with robotic bodies and expresses some skepticism, as I do. He also touches on some of the social changes that would be likely to occur. Perhaps everyone will spend most of their time in advanced virtual realities. More immediately, it would be a dramatic shock if AGI, in short order, stopped climate change, created a source of unlimited clean energy and replaced most human jobs along with all governments. While on the surface these would represent positive changes, they would also be traumatic. New ways would have to be found for people to occupy their time. Up to now, people have been able to distinguish themselves by working harder, having better ideas or being more productive than their peers, and those human-level talents would pale in comparison to AGI. The very idea of "genius" would instantly become obsolete, because all of the most important ideas that were originated by humans could perhaps be discovered and improved upon in seconds by AGI. How would people increase their social status under such circumstances? How would human interactions be affected by the introduction of realistic androids? Russell repeatedly refers to the gorilla analogy, in which gorillas were once a dominant primate species but became endangered when humans came along: the same may apply to us when AGI arrives.

I am also a little surprised that Russell has little to say about the potential abuse of AGI by humans. It is to be expected that individuals will attempt to control it in order to achieve their personal objectives rather than to support the welfare of mankind. What would happen, for example, if Vladimir Putin took control of an AGI developed in Russia? This would probably be just as disastrous an outcome as the rogue AGI to which Russell devotes so much of the book.