Explore an in-depth interview with Stanislas Polu, research engineer at OpenAI and lead author of "Formal Mathematics Statement Curriculum Learning." Delve into the groundbreaking application of expert iteration in language modeling for formal mathematics. Discover how this approach outperforms traditional proof search methods and creates a self-guided curriculum for solving increasingly complex problems. Learn about the state-of-the-art results achieved on the miniF2F benchmark, including solutions to challenging high school olympiad problems. Gain insights into the research process, experimental results, and potential future directions for AI in formal mathematics. Understand the implications of this work for machine-human collaboration in mathematical problem-solving and the broader impact on AI research.
AI and Formal Math - Formal Mathematics Statement Curriculum Learning
Overview
Syllabus
- Intro
- How do you explain the big public reaction?
- What's the history behind the paper?
- How does algorithmic formal math work?
- How does expert iteration replace self-play?
- How is the language model trained and used?
- Why is every model fine-tuned on the initial state?
- What if we want to prove something we don't know already?
- How can machines and humans work together?
- Aren't most produced statements useless?
- A deeper look at the experimental results
- What were the high and low points during the research?
- Where do we go from here?
Taught by
Yannic Kilcher
