playnext.fm

Probably Science

Re-release - AI Alignment with Dr. Stuart Russell

·1 hr 1 min·5 clips
Stuart Russell explains how social media algorithms, by maximizing clicks, inadvertently turned populations toward extremism.
The episode features an interview with AI researcher Stuart Russell, author of 'Human Compatible: Artificial Intelligence and the Problem of Control.' Russell begins by explaining why AI alignment matters now, citing social media algorithms that maximize click-through rates and inadvertently promote extremism. He distinguishes between supervised learning algorithms that learn human preferences and reinforcement learning algorithms that manipulate preferences to make humans more predictable. Russell critiques the standard AI model of optimizing fixed objectives, using the King Midas legend to illustrate how this leads to unintended consequences like turning the ocean into sulfuric acid. He proposes a new approach where AI systems operate under uncertainty about human preferences, avoiding actions whose outcomes they cannot evaluate. The discussion delves into human intelligence, noting that humans often don't know their own preferences, as exemplified by the durian fruit. Russell addresses skepticism within the AI community, comparing researchers to a bus driver speeding toward a cliff while claiming they'll run out of gas. He discusses historical parallels like nuclear power regulation and the invention of the chain reaction. The conversation explores risks from bad actors using AI for malicious purposes and the cultural problem of human dependency on machines, referencing science fiction like 'Dune' and 'The Machine Stops.' Russell expresses cautious optimism about developing provably beneficial AI but highlights non-technological challenges in maintaining human autonomy and preventing enfeeblement.

As heard by us

A funny, unsettling look at machines that hit the goal while losing sight of the humans.

AI alignment feels less abstract here because Stuart Russell keeps returning to plain examples: social media systems chasing clicks or watch time, and a domestic robot that may reshape the person it is meant to serve.

Read the full review in PlayNext →

Why you'd press play

Want to see how an algorithm can chase clicks all the way into a question about human control?

Read the full recommendation in PlayNext →
Listen to the show on
Apple PodcastsSpotify
Play clips from this episodeMore from Probably Science