DiscussionTechnical

Limits of AI - explained by mathematicians | Lex Fridman

Lex Clips

Mathematicians discuss the 'solved axiom' (the idea that AI could eventually solve all mathematical problems) and debate its plausibility. They explore the distinction between AI's mechanical problem-solving through statistical patterns versus human mathematical understanding, emphasizing that humans currently still ask the important questions while AI executes solutions.

Summary

The transcript captures a discussion between Lex Fridman and two mathematicians about the limits and capabilities of AI in mathematics. One mathematician expresses skepticism about the 'solved axiom,' arguing that proving a theorem is different from truly understanding why the proof works—there remains unexplored 'magic' in mathematical discovery. They emphasize that good mathematics is characterized by simplicity, generalizability, and the asking of meaningful questions, which currently remain human-driven activities.

The conversation then explores whether AI could eventually discover mathematical truths autonomously. One speaker notes that research programs already exist that systematically discover minimal physical laws by observing natural systems (like deducing gravity from falling objects or Navier-Stokes from fluid dynamics). However, there's debate about whether AI could independently identify which questions to ask—the core creative act in mathematics.

Ankur expresses moderate skepticism about the solved axiom, warning against anthropomorphizing AI capabilities. He illustrates this with deep learning examples, particularly the 'Adversarial Examples Are Features, Not Bugs' research showing that neural networks solve problems through statistical patterns and 'trapdoors' fundamentally different from human reasoning. He argues that while AI succeeds remarkably well, its mechanisms differ qualitatively from human problem-solving, suggesting AI is better viewed as a complementary tool for 'centaur math' (human-AI collaboration) rather than as a replacement for human mathematicians.

Key Insights

  • One mathematician argues that proving a theorem is not the same as understanding why the proof works, and that unexplored 'magic' in mathematical reasoning prevents them from fully accepting the solved axiom
  • The gold standard in mathematics has always been finding simple results with high generalizability, and humans currently retain the role of asking which questions are worth exploring
  • Research programs already exist that systematically discover minimal physical laws by observing natural systems—for example, deducing gravity (F=MA) from observing balls dropping
  • Neural networks often solve problems through statistical 'trapdoors' or patterns invisible to humans—demonstrated by research showing networks trained on mislabeled data with adversarial perturbations still learn to recognize natural images
  • While AI solutions look qualitatively different from human mathematical work, this difference presents an opportunity for collaboration ('centaur math') rather than a threat, assuming we avoid over-anthropomorphizing how AI actually solves problems

Topics

The solved axiom and its plausibilityUnderstanding vs. proof in mathematicsQuestion formulation as the core human mathematical contributionAutomated discovery of physical laws and mathematical structuresDeep learning and adversarial examplesAI reasoning mechanisms vs. human mathematical reasoningFuture of human-AI collaboration in mathematics

Transcript

[0:03] - By the way, and so to, sorry, to go back to your point, I don't actually believe the solved axiom. I don't think we're there. I can prove a theorem, but I don't understand why the proof works. There's some magic that happens over here, and until I generalize somehow the phenomenon and understand it more deeply, then I'll see why the, the proof works the way it does. Maybe that's one of the reasons why I don't actually believe the solved axiom, because the things that we do are much more exploratory and what constitutes a good question beyond just, "Here's the Riemann hypothesis, make some progress on it." [0:35] - So I think you also brought…

Full transcript available for MurmurCast members

Sign Up to Access

More from Lex Clips

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.