Newcomb's Problem is a thought experiment which, on the surface, seems obvious, but what if you're trying it out on your identical twin? Which choice is best? Aric Floyd demonstrates it, then tries it on Anthropic's Claude, which doesn't always play ball!

Computerphile is supported by Jane Street. Learn more about them (and exciting career opportunities) at: https://jane-st.co/computerphile

This video was filmed and edited by Sean Riley.

Computerphile is a sister project to Brady Haran's Numberphile. More at https://www.bradyharanblog.com

Computerphile is a YouTube channel and platform dedicated to computer science education, featuring videos on a wide range of topics, from algorithms and data structures to computer hardware and software engineering. Readers can learn about computer science concepts, programming languages, and the history of computing. With engaging videos, expert interviews, and educational content, Computerphile provides a resource for students, educators, and technology enthusiasts.

Computerphile

Newcomb's Problem, a philosophical paradox about prediction and decision-making, is tested on modern AI language models (Claude 3.5 and 3.7). The experiment demonstrates how AI systems might reason differently than humans when facing game theory scenarios, particularly when encountering copies of themselves. The results show Claude choosing the seemingly irrational "one-box" strategy, suggesting AI training data includes decision theory concepts that could lead to unexpected behaviors when AI agents interact with each other without human oversight.

Trying Newcomb's Problem on AI - Computerphile