Grant Sanderson – AI and the future of math
Summary
Grant Sanderson discusses AI's rapid progress in mathematics, noting it's the fastest-advancing field for AI. While AI can "cold-solve geometry" problems in 19 seconds for the International Math Olympiad (IMO) and achieve gold-level performance, it still struggles with combinatorics, indicating a "spiky frontier." He posits that solving Millennium Prize problems might involve deep domain knowledge, cross-field connections, or building entirely new mathematical theories, rather than just brute-force computation. Sanderson highlights that future AI benchmarks in math will likely shift from problem-solving to generating novel conjectures and definitions, which are harder to quantify. He also explores AI's advantages like parallelization and systematic context manipulation, contrasting them with human limitations. The discussion attributes AI's success in math and coding to "verifiability" and "grindability," allowing for extensive parallel rollouts and deterministic credit assignment.
Key takeaway
For research scientists evaluating AI's impact on scientific discovery, recognize that AI's current mathematical prowess, driven by "grindability" and verifiability, will increasingly automate theorem proving. Your focus should shift towards generating novel conjectures, definitions, and distilling complex AI-derived insights into human-understandable frameworks. Embrace AI as a tool for systematic exploration and connection-making, but prioritize developing the "curatorial" and "explanatory" skills essential for navigating an AI-accelerated mathematical landscape.
Key insights
AI's rapid mathematical progress stems from verifiable, grindable domains, shifting future benchmarks to conceptual innovation.
Principles
- AI progress is "spiky," excelling in some math areas over others.
- "Grindability" and verifiability drive AI's success in math and code.
- Human understanding remains crucial for complex AI-generated proofs.
Method
AI can systematically explore problem spaces by manipulating context, such as trying both positive and negative proofs or applying different biases.
In practice
- Use LLMs as advanced search tools to find human-curated learning resources.
- Consider AI for "auto-research" in formal systems like Mathlib.
Topics
- AI in Mathematics
- Mathematical Discovery
- Automated Theorem Proving
- AI Benchmarks
- Formal Verification
- Large Language Models
Best for: AI Scientist, Research Scientist, AI Ethicist
Related on AIssential
See Counsel's argued verdicts on the open AI decisions leaders are weighing →
Editorial summary, takeaway, and curation by AIssential. Original article published by Dwarkesh Podcast.