# OpenAI’s New AI Is Doing Things We Barely Understand

## Executive summary

The video discusses the rapid advancement of AI in mathematics, citing OpenAI's release of hundreds of AI-generated mathematical results (722 papers). While these results demonstrate unprecedented computational power—solving problems previously considered impossible (e.g., Navier-Stokes equations)—the speaker raises critical questions regarding the nature of mathematical proof, the value of process versus result, and the crucial need for democratizing access to these powerful models through open weights technology.

## Key takeaways

- AI's Mathematical Breakthroughs: OpenAI has released hundreds of AI-generated mathematical results, including proofs for complex problems like irrational triangular billiards (where almost every ray visits every spot if the triangle has an angle irrational relative to $\pi$) and conformal transformations of planar regions with holes (1:30-2:00).
- The Debate on Mathematical Proof: The speaker notes that these AI proofs are often not 'elegant' and raises the philosophical question of whether the goal of mathematics is the final answer (result) or the illumination provided by the proof process (1:50).
- The Need for Open Weights AI: The speaker argues that while frontier AI models are incredibly powerful, their current exclusivity to 'frontier labs' must be countered by open weights models to ensure that scientific progress is accessible to all researchers (3:50).
- Community Oversight: The video concludes by noting that three papers had been withdrawn due to community feedback, highlighting the importance of democratizing and critically reviewing AI-generated research (5:20).

## Technical details

- Irrational Triangular Billiards: A problem concerning light bouncing in a triangular mirror room. The claim is that if the triangle has at least one angle irrational relative to $\pi$, almost every ray visits every spot (1:00-1:30).
- Conformal Transformation: The process of applying a stretch that preserves angles to an ugly planar region with holes, allowing the holes to morph into perfect circles or points (1:30-2:00).
- AI Progress Benchmarks: The speaker tracks the rapid acceleration of AI capabilities, noting that models progressed from being unable to reliably add numbers (GPT-3, 6 years ago) to potentially solving millennium problems (3:00-3:30).

## Practical implications

- The necessity of open-source, open-weights models to ensure equitable access to advanced AI research and prevent scientific progress from being confined to private 'frontier labs'.
- The role of the academic community in critically reviewing and validating AI-generated scientific claims, as demonstrated by the withdrawal of three papers due to community feedback.
- The increasing utility of AI as a tool for generating complex proofs and simulations, shifting the focus of research from manual calculation to model interpretation and deployment.

## Topics

Artificial Intelligence, Mathematics, OpenAI, Open Weights Models, Computational Science, Build Engineering, OpenAI Math Progress, OpenAI GitHub Repository, Lambda.ai

Source: https://www.youtube.com/watch?v=z_sXLlLVUls