DeepMind’s New AI Found A Strange New Way To Think

About this video

• Video Title: DeepMind’s New AI Found A Strange New Way To Think
• Channel: Two Minute Papers
• Speakers: Dr. Károly Zsolnai Fehér
• Duration: 00:07:30

Overview

This video discusses DeepMind's AI, AlphaProof Nexus, and its attempt to solve mathematical problems posed by mathematician Paul Erdős. While the AI had a high failure rate, the video argues that this is a significant achievement because it tackles problems that have remained unsolved for decades. The speaker highlights the novel approach of using an unreliable AI within a "harness" or "loop" involving a judge AI to iteratively improve solutions, likening it to a chess tournament. The video also touches on the evolution of AI capabilities and the importance of the surrounding systems in achieving reliable results.

Key takeaways

• DeepMind's AI, AlphaProof Nexus, attempted to solve 350 of Paul Erdős's open mathematical problems, succeeding in 9 out of the solved problems, with a cost of a few hundred dollars per problem.
• Despite a high failure rate, the AI's ability to solve problems that have stumped mathematicians for decades is considered a significant achievement.
• The AI utilizes a formalized mathematical language called Lean for verification and employs a "judge" AI that compares and selects the "better" of two potentially incorrect solutions, iteratively improving them in a tournament-like system.
• The intelligence is not solely within the AI model itself but also within the "loop" or "harness" system surrounding it, which includes a reliable judge to ensure correctness.
• The video contrasts the current AI progress with past skepticism, showing a rapid advancement from basic arithmetic to solving complex, long-standing mathematical problems.