Human Compatible — Stuart Russell

Summary in 15 minutes. For decades AI researchers have tried to make machines intelligent. Russell asks a simple, uncomfortable question: what if we actually succeed? AI can go wrong in two ways: people can use it for harm on purpose, and well-meant machines can chase the wrong goal. If the root problem is a fixed objective, what is the fix? Russell’s answer: machines that are uncertain about what we want. The AI world has changed a lot since the book appeared. What changed, and what can ordinary users do?

Key ideas of "Human Compatible: Artificial Intelligence and the Problem of Control"

If We Succeed

For decades AI researchers have tried to make machines intelligent. Russell asks a simple, uncomfortable question: what if we actually succeed?

1. The Last Invention?: The question AI researchers avoided
2. What Intelligence Means: The “standard model” and its hidden trap
3. Where AI Is Heading: The near future, and the breakthroughs still missing

The Risks

AI can go wrong in two ways: people can use it for harm on purpose, and well-meant machines can chase the wrong goal.

1. Misuse by People: Surveillance, persuasion, weapons and jobs
2. The King Midas Problem: You get exactly what you asked for
3. “Nothing to Worry About”: Arguments that dismiss AI risk, and Russell’s replies

A New Foundation for AI

If the root problem is a fixed objective, what is the fix? Russell’s answer: machines that are uncertain about what we want.

1. Three Principles for Beneficial Machines: Machines that pursue our objectives, not their own
2. Learning from Human Behavior: Inverse reinforcement learning and assistance games
3. The Off-Switch: A machine that lets itself be switched off
4. Complications: Us: Many people, nasty preferences, mistakes and changing minds

Today and Tomorrow

The AI world has changed a lot since the book appeared. What changed, and what can ordinary users do?

1. What Has Changed Since 2019: Our note: the AI world after the book
2. 7-Day Mindful AI Plan: One task a day. Tick it off when done.
3. One-Page Cheat Sheet: The whole book at a glance. Save it.
4. Final Word: Staying Human: The danger of over-reliance, and Russell’s hope

From the free first chapter

Stuart Russell is a professor of computer science at the University of California, Berkeley. The textbook he wrote with Peter Norvig, Artificial Intelligence: A Modern Approach, is used in universities around the world. This warning comes from deep inside the field.

Russell noticed that almost everyone in AI was working on how to make machines smarter, and almost no one was asking what happens if we get there. Superhuman AI, he argues, could be the biggest event in human history. Handled badly, it could also be the last.

Read the first chapter free · Price: ৳250 · or every book with a membership