Overview
Explore OpenAI's groundbreaking approach to formal mathematics and automated problem-solving in this comprehensive video lecture. Delve into the challenges of formal proofs for both humans and machines, and discover how OpenAI employs expert iteration to create an automated curriculum for AI systems. Learn about the model, data, and training procedures used, including techniques for predicting proof lengths and bootstrapping expert iteration. Examine experimental evaluations, scaling properties, and results on synthetic data. Witness how this innovative method enables AI to solve complex problems from international math olympiads, previously thought infeasible. Gain insights into the future of AI in mathematics through detailed discussions and expert commentary.
Syllabus
- Intro
- Paper Overview
- How do formal proofs work?
- How expert iteration creates a curriculum
- Model, data, and training procedure
- Predicting proof lengths for guiding search
- Bootstrapping expert iteration
- Experimental evaluation & scaling properties
- Results on synthetic data
- Solving real math problems
- Discussion & comments
Taught by
Yannic Kilcher