Perhaps we will be able to develop a culture in which everyone learns for its own sake and for the sake of having a richer inner life. But I think this not necessarily the most likely outcome of current trends. The danger of enfeeblement is real. Prof. Stuart Russell in conversation with Maciej Chojnowski

Maciej Chojnowski: Your new book is entitled Human Compatible: Artificial Intelligence and the Problem of Control. You are also the founder of the Center for Human-Compatible Artificial Intelligence at UC Berkeley. Who is responsible for making AI human compatible? Scientists? Engineers? Entrepreneurs? Politicians? Society as a whole?

Prof. Stuart Russell*: Scientists and engineers have to invent the basic technology and develop provably safe designs. Regulators and standards organizations have to define standards for AI systems, and (mostly) corporations have to build systems that comply with the standards.

Probably politicians will need to pass laws as appropriate, but they will usually be guided by experts and industry lobbyists on this. There is certainly a danger that industry will lobby against any constraints on the kinds of systems they can build.

Algorithmic bias and black box problem are one of the major challenges in today’s machine learning and deep learning. But what are the biggest challenges to human compatible AI? Will it require complex interdisciplinary approach to overcome these challenges?

The most serious challenges, in order of arrival, are: first, algorithms that manipulate human opinions and preferences and spread misinformation (already happened). Second, lethal autonomous weapons (on sale now). Third, major impacts on employment. And fourth, loss of control over increasingly intelligent machines.

Can you elaborate on that?

Fighting misinformation requires technology for detection, monitoring, watermarking, authentication, etc., plus institutions and structures to ensure that truth prevails, and perhaps laws too.

Lethal autonomous weapons require clear understanding of the problem by major governments and concerted diplomatic action to create a treaty.

Impact on employment requires a vision for how humans might have economically valued roles in a world where almost all of what we currently call work is done better and cheaper by machines. My guess is that these roles will involve mainly interpersonal services to enhance quality and richness of life; for this, we need to do a lot of scientific research to create the knowledge required for these roles to be effective and valuable, and education reforms needed to make that vision a reality.

At present, China explicitly recognizes that superintelligent AI represents an existential threat to humanity, whereas the US and EU do not

Finally, loss of control over intelligent machines requires basic AI research on provably safe and beneficial AI. Part of that involves learning preferences from human behavior (psychology), acting on behalf of multiple humans (moral philosophy, economics, political theory), etc.

Let’s assume that the EU and the USA have agreed on creating trustworthy, human compatible AI. How to make sure that the rest of the world will follow the same rules and values?

First, this kind of AI will just be better than the traditional kind (that optimizes a fixed objective that we supply). There would be little or no economic incentive to create and sell the old kind of AI because it just doesn’t work as well and carries a lot of risk too.

Second, most people agree that human-level AI can create such an abundance of wealth that there is no point keeping it; the technology should be shared so that all can benefit.

Third, the AI systems don’t need to have all the “same values.” What matters is the preferences that each person has about how the future should unfold. In principle, that’s eight billion preference models.

Fourth, we can institute standards similar to TCP: you simply cannot use the Internet without adhering to the TCP standard, so we need something similar for AI systems. We may still need some form of policing to make sure unsafe AI systems are detected and disabled, and that may require some international agreements. At present, China explicitly recognizes that superintelligent AI represents an existential threat to humanity, whereas the US and EU do not.

Youtube movie URL: https://youtu.be/p9kpG9AaXDs

Prof. Stuart Russell at Data Driven Paris.
Source: DataDrivenParis / YouTube

When a layman hears about achievements like the one of AlphaGo or AlphaZero, they may think that we are about to enter an era of AGI. Are we truly living in an age of sudden AI breakthroughs? And should we expect AGI coming soon?

Progress is certainly rapid, but we are several major conceptual breakthroughs away from human-level or superhuman AI. It’s unlikely these would all happen in a short time period. I’m generally more conservative that the great majority of AI researchers, who expect it to arrive in 30-40 years.

Incidentally, AlphaGo, although a very impressive system, does not constitute any sort of breakthrough. It combines three significant breakthroughs: reinforcement learning via self-play in games (late 1950s), decision-theoretic metareasoning to control search (late 1980s to early 2000s), and deep convolutional neural networks (mid- to late 1990s).

In your conversation with Martin Ford (in Architects of Intelligence) you said that today’s mindset of many governments about the future of work – „We need more data scientists” – isn’t perfectly right. We simply won’t need a billion data scientists. A few million will be enough. If so, in what disciplines or areas should we train or upskill people? Should we expect increased interest in behavioural skills?

Inevitably, most people will be engaged in supplying interpersonal services that can be provided—or which we prefer to be provided—only by humans. That is, if we can no longer supply routine physical labor and routine mental labor, we can still supply our humanity. We will need to become good at being human.

If we can no longer supply routine physical labor and routine mental labor, we can still supply our humanity

Current professions of this kind include psychotherapists, executive coaches, tutors, counselors, companions, and those who care for children and the elderly. Many more kinds will emerge as we develop effective methods—based on research—for improving individual quality of life in innumerable ways. This requires a major shift in our science base, which has focused almost entirely on the physical and biological world, as well as a reorientation of the humanities and social sciences.

When AI and automation change our economy people may have less incentives to learn because education won’t have an economic function it had before. Do you think that people in the era of basic income will be able to concentrate on the art of good living instead of working? Is radical transformation of our way of life possible?

Keynes seemed to believe in good living without work. In his 1930 article “Economic Possibilities for Our Grandchildren,” he wrote, “It will be those peoples, who can keep alive, and cultivate into a fuller perfection, the art of life itself and do not sell themselves for the means of life, who will be able to enjoy the abundance when it comes.”

Perhaps we will be able to develop a culture in which everyone learns for its own sake and for the sake of having a richer inner life and the opportunity to pursue a wide range of intellectual goals. But I think this is not necessarily the most likely outcome of current trends. The danger of enfeeblement is real. More than one hundred billion people have lived on Earth. They (we) have spent on the order of one trillion person-years learning and teaching, in order that our civilization may continue. Up to now, its only possibility for continuation has been through re-creation in the minds of new generations. (Paper is fine as a method of transmission, but paper does nothing until the knowledge recorded thereon reaches the next person’s mind.) That is now changing: increasingly, it is possible to place our knowledge into machines that, by themselves, can run our civilization for us.

Once the practical incentive to pass our civilization on to the next generation disappears, it will be very hard to reverse the process. One trillion years of cumulative learning would, in a real sense, be lost. We would become passengers in a cruise ship run by machines, on a cruise that goes on forever—exactly as envisaged in the film Wall-E.


*Prof. Stuart Russell is a computer scientist and a world-renowned expert on artificial intelligence. He is a Professor of Computer Science at the University of California, Berkeley and Adjunct Professor of Neurological Surgery at the University of California, San Francisco. He is the founder of the Center for Human-Compatible Artificial Intelligence at UC Berkeley. Along with Peter Norvig, he is the author of the most popular textbook on AI: „Artificial Intelligence: A Modern Approach”. He is on the Scientific Advisory Board for the Future of Life Institute and the Advisory Board of the Centre for the Study of Existential Risk. His research interests include i.a.: foundations: rationality and intelligence, the long-term future of AI, learning probability models, intelligent agent architecture. His latest book is „Human Compatible: Artificial Intelligence and the Problem of Control”.

Przeczytaj polską wersję tego tekstu TUTAJ

Russell at World Summit AI

Professor Stuart Russell is one of the keynote speakers at World Summit AI (WSAI) in Amsterdam (09–10 October 2019). Sztuczna Inteligencja is WSAI Media Partner.

Skip to content