Interview with Stuart Russell: "AI Alignment and the Future of Humanity"

Dr. Stuart Russell 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.

Dr. Russell’s research spans many areas of artificial intelligence, including machine learning, probabilistic reasoning, and philosophical foundations. Recently, his work has focused on ensuring that advanced AI is developed safely.

If we succeed, we’re going to need to…know how to control these systems. Otherwise, we’ll have a catastrophe.

Jump to a question:

• How would you respond to those who argue that any generally intelligent system should, by definition, be able to understand human goals?

• Could you describe the unintended consequences that have resulted from recommendation algorithms?

• Do problems with recommendation algorithms portend broader challenges of control and misspecified objectives in AI?

• How would you respond to those who argue that any generally intelligent system should, by definition, be able to understand human goals?

• Is the problem of misspecified objectives the same as the so-called “value alignment” problem?

• What is wrong with the popular view that advanced AI will spontaneously develop a will to power (as distinct from fears about instrumental goals)?

• Could you describe your principles of “Provably Beneficial AI?”

• Why isn't it as simple as finding and satisfying the "human utility function?"

• Would a morality that far surpasses our own even be recognizable to us as moral?

• Do you think we’ll encounter conflicts between what we learn from AI and our intuitions about justice, autonomy, etc.?

• How can we make it easier for AI to learn human values?

• What role can psychologists play in AI Alignment?

• Could you explain what you’ve called the “Dr. Evil” problem?

• Could you explain what you’ve called the “Somalia” problem?

• How should autonomous cars trade off between the safety of drivers and pedestrians?

• Do even narrow AI systems need to be “provably beneficial” in the sense you’ve described?

• How would you respond to those who argue that these concerns are too far off (like “overpopulation on Mars”) or low in probability to work on?

• How should young people prepare themselves for the changing job landscape of the future?