Review of Human Compatible. Human Compatible: AI and the Problem of…

Human Compatible: AI and the Problem of Control is the 2019 book by Professor Stuart Russell.

Chapter 1 asks the question, what will our future look like if we succeed in creating human level or superhuman level AI? Russell thinks the invention of superintelligent AI will be the biggest event in humanity.

Chapter 2 Intelligence in Humans and Machines begins with an explanation of the Nash Equilibrium relating to two player games, and tragedy of the commons. Then it goes through history beginning with Turing’s 1936 concept of universality and 1950 concept of the imitation game. It discusses chess programs and tells the story of AlphaGo and AlphaZero. Russell also explains Gödel’s completeness theorem and mentions Shakey the robot.

Russell is of the opinion that deep learning is not enough to get us all the way to human level AI. The chapter ends with image classification mishaps such as labelling Jacky Alcine and her friend as gorillas in 2015, and Google’s subsequent removal of its ability to identify real gorillas.

Chapter 3 is How Might AI Develop in the Future and this begins with the story of Deep Blue defeating Gary Kasparov in 1997. When it starts looking part the basic assistants Alexa and Siri to imagine what a more powerful assistant would look like, we see that in this area this book, written in 2019, has already become very outdated. These super assistant programs mostly exist today in the form of ChatGPT and the like.

A few areas where these programs yet don’t match up to Russell’s descriptions:

• Keeping track of the user’s body — this is done by the iWatch, but the iWatch is limited in other AI areas
• Understanding of user’s bank account — well you COULD tell ChatGPT about this, but at your own risk. Russell talks a bit about privacy issues.
• Teaching Sting Theory in Laotian, probably not your concern as you are reading this in English.

Russell also survey’s the latest robotic technology at the time of his writing, which were BRETT and Spot. These robot technologies have evolved significantly over the past 6 years. Russell is still correct that the robot butler is still some way off, but today’s Figure 03 is getting closer to this dream.

Rusell then moves to the question of when we will have SuperIntelligent AI, making the point that there’s no clear point because machines already exceed humans in some areas.

Then he moves on to discussing the conceptual breakthroughs needed to achieve SuperIntelligent AI. Some of this also feels outdated as many of the breakthroughs have already happened. He mentions the NELL project (Never-Ending Language Learning) as an example of a program with a lot of meaningless beliefs, nonsense such as “Nepal is a country also known as United States”. Language processing has since moved on a long way.

Russell spoke about the topic of reaching Artificial General Intelligence in 2015 as part of AAAI webinar which you can find here:

Also in Chapter 3, Russell says Goodman’s Fact, Fiction and Forecast is one of the most underrated books on machine learning.

Here are some possible ways that we could benefit from future AI:

• Extending human life
• Raising the average standard of human life
• Personal tutoring for every child
• Everything as a service

Chapter 4 is Misuses of AI and this begins with the subject of surveillance. Russell warns “in regimes where this technology is adopted for the purposes of control, it will be as if every citizen had their own personal Stasi operative watching over them twenty-four hours a day.”

Surveillance technology can be used for blackmailing purposes, especially if someone is caught doing something they don’t want to be made public. The first well known blackmail bot was Delilah.

AI can manipulate humans in ways including coercing them to buy products that they don’t need by feeding them tailored messages. AI can use feedback to quickly master deceptive techniques.

Deepfakes are another problem Russell mentions, although this is another areas where things have come a long way since 2019.

On the subject of reality vs fake world, Russell argues that mental security should be a human right, and that includes the ability to tell fact from fiction.

The W3C Credible Group aims to “shift the Web toward more trustworthy content without increasing censorship or social division.” and the Credibility Coalition has a similar mission.

Russell then moves on to discuss issues around Lethal Autonomous Weapons Systems. He calls them “scalable weapons of mass destruction” which could be “used to wipe out an entire ethnic group”.

The next topic is AI taking over people’s jobs. Russell uses a paintbrushing example to show that as technology improves, first the number of human jobs increases to support the industry but once it reaches a very high level of automation, there is very little need for human involvement.

The Great Decoupling states the productivity is rising but employment is not. Since 1973 wages stagnated while productivity roughly doubled. CGP Grey documented the employment problem we are facing in his 2014 video Humans Need Not Apply. This chapter also covers algorthmic bias.

Chapter 5 is Overly Intelligent AI, or as Russell calls it, the Gorilla problem. He gives a historical account of the many warnings of the danger of AI takeover, starting in Richard Thornton’s warning in 1847. The earliest mainstream warning was Samuel Butler’s Erewhon. This book was read by Alan Turing and he went on BBC Radio to make a similar warnign in 1951.

Norbert Wiener pioneered many new ideas about machine learning in his book God and Golem, and he warned “human incompetence has shielded us from the full destructive impact of human folly.”

Russell goes on to explain that machines may actively prevent us from turning them off, like in the plot of 2001.

Russell concludes the chapter by reviewing our not-so-great options for dealing with this threat. Hans Moravec, whose writings have influenced Richard Sutton, wrote “The immensities of cyberspace will be teeming with unhuman superminds, engaged in affairs that are to human concerns as ours are to those of bacteria.”

Chapter 6 is The Not-So-Great AI Debate. It begins with the various forms of denialism. We are getting into the meat of the book now. Russell presents and refutes several of the denialist arguments which have been made.

Russell criticises the report AI and Life in 2030 for claiming super intelligent AI is probably impossible, arguing that it demonstrates denialism and tribalism. He also dismantles the “It’s too soon to worry” argument and takes on the “We’re the experts” arguments.

The next set of arguments that Russell lays out is deflection; acknowledging their are risks, but arguing that we should do nothing. These arguments are subdivided into “you can’t control research”, “whataboutery” and “silence”.

Also discussed is the problem of tribalism, and the various Can’t We Just… arguments. An interesting resource for AI researchers is Thinking Inside the Box: Controlling and Using an Oracle AI.

Chapter 7 is AI: A Different Approach

Chapter 8 is Provably Beneficial AI

Chapter 9 is Complications: Us

Chapter 10 is Problem Solved?

Appendices are Searching For Solutions, Knowledge and Logic, Uncertainty and Probability, Learning From Experience.

I will be updating this book review as I progress through this book.

What I can definitely say is I would love for there to be a 2nd edition of this book. The last 6 years have seen tremendous change, and I think Russell could now improve on this work significantly. However this is not to say this book is completely outdated, mostly it is still relevant and important, just certain sections feel like they are coming from a previous era.

Full version of Lex Fridman interview: