The safest AI might be one that doesn't know what we want | Stuart Russell

For 50,000 of generations, humans have passed civilization down to their children. We could be the first generation to hand its future to AIs instead. Should we? Or could the best AI be the kind that chooses to step back and leave some things to us?

In this episode we sit down with Stuart Russell, professor of computer science at UC Berkeley and co-author of the world's standard textbook on AI. He’s also a leading proponent of provably beneficial AI: systems that are safe by design because their only goal is to further human interests.

We talk about: How AI has changed over the past 50 years, from simple game-playing programs to today's large language models Why handing an AI a fixed objective becomes dangerous once it is more capable than us, and what a safer approach could look like How an AI might learn what we really want, even when we don't fully know ourselves How the race toward more powerful AI can still be steered somewhere safer What happens to human purpose as AI becomes more and more capable

Chapters: 0:00 Cold open 0:39 How AI has changed over the past 50 years, from chess to large language models 4:15 What is the standard model of AI and why does it fail? 12:08 What is provably beneficial AI? 17:49 Teaching AI to learn from humans, even when we make bad choices 20:11 How can we steer AI development in a safer direction? 28:01 International cooperation on AI safety 30:41 What a good AI future could look like 35:38 Protecting human purpose: should AI step back? 39:15 How young people can get involved in AI safety 40:47 Stuart Russell's best advice On the Existential Hope Podcast hosts Allison Duettmann and Beatrice Erkers from the Foresight Institute invite scientists, founders, and philosophers for in-depth conversations on positive, high-tech futures.

Full transcript, listed resources, and more: https://www.existentialhope.com/podcasts

Follow on X. Hosted on Acast. See acast.com/privacy for more information.