Dwarkesh Podcast

Terence Tao – Kepler, Newton, and the true nature of mathematical discovery

1h 24min

Terence Tao discusses the nature of mathematical discovery, using Kepler's empirical journey to planetary laws as an analogy for AI's potential in math. He highlights how AI excels at generating ideas and applying existing techniques across a broad range of problems, driving down the cost of hypothesis generation. However, the new bottleneck is validating and evaluating these ideas, as current AI struggles with true conceptual depth, cumulative understanding, and identifying partial progress, suggesting a future of human-AI collaboration in science.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Kepler as an early "LLM": Kepler's discovery of planetary laws, driven by Tycho Brahe's precise data and iterative testing of theories (from Platonic solids to ellipses), serves as an analogy for how AI can find empirical regularities by trying random relationships against verifiable datasets.
2

AI's impact on scientific bottlenecks: AI has drastically reduced the cost of idea generation in science, shifting the bottleneck from hypothesis creation to verification, evaluation, and identifying true progress amidst a flood of potential theories.
3

Complementary strengths of AI and humans: AI excels at breadth—applying existing techniques and solving many "low-hanging fruit" problems—while human experts excel at depth, creating new techniques and building cumulative understanding, suggesting a future of human-AI collaboration.
4

The challenge of "intelligence" vs. "cleverness": Current AI demonstrates "cleverness" through brute-force trial and error, but lacks "intelligence" in terms of evolving understanding, building cumulatively from partial progress, or retaining new skills for future problems.
5

Need for formalizing mathematical strategies: Beyond formal proofs, there's a need for a semi-formal language to assess the plausibility and confidence of mathematical conjectures and strategies, similar to how scientists communicate heuristically about statistical patterns (e.g., prime number distribution).

💬 Notable Quotes

"I think AI has basically driven the cost of idea generation down to almost zero."

— Terence Tao
"AI tools are really good at getting really good at the first part of it, just trying all the static techniques on a problem, often now actually making fewer mistakes in implementing them than than humans."

— Terence Tao
"It is complementary currently. It is not a replacement."

— Terence Tao