# Terence Tao – How the world’s top mathematician uses AI

- Channel: [Dwarkesh Patel](https://www.youtube.com/channel/UCXl4i9dYBrFOabk0xGmbkRA)
- Published: 2026-03-30
- Source video: [Watch on YouTube](https://youtube.com/watch?v=Q8Fkpi18QXU)
- Read on ReadTube: https://www.read.tube/p/videos/Q8Fkpi18QXU
- Language: English

## Summary

**Tao Sees AI Reshaping Math**

Terence Tao argues that AI is making idea generation cheap, but verification, judgment, and scientific storytelling are becoming the real bottlenecks. Using Kepler, Brahe, and Newton as a model, he says AI will likely flood math and science with candidate ideas while humans and institutions must better sort out which ones are real progress. He thinks current AI is already a powerful breadth tool for mathematics, especially when paired with Lean and human experts, but not yet a replacement for deep, cumulative mathematical understanding.

Terence Tao frames AI as a dramatic shift in how mathematics and science work, comparing it to the historical move from theory-driven discovery to data-driven pattern finding. He uses Kepler’s discovery of planetary laws to argue that major advances often come from trying many bad ideas against reliable data, and says AI can now do that kind of broad exploration at scale. But the core bottleneck has moved: generating candidate theories is easier, while validating them, judging their importance, and turning them into useful science is much harder.

He emphasizes that many famous breakthroughs were initially incomplete, wrong in part, or less accurate than existing theories. Copernicus was simpler but less precise than Ptolemy, Darwin lacked a mechanism for heredity, and Newton’s theory left conceptual mysteries. That means future AI-generated results may look messy, partial, or even unhelpful at first while still representing real progress. Tao worries that current review systems are not designed to handle a flood of AI-generated papers and proofs, and says the field needs better benchmarks, standardized datasets, and ways to assess partial progress and explanatory quality.

On math specifically, Tao says AI has already solved a noticeable number of Erdős problems and is very good at trying standard techniques, checking details, and doing breadth work, but still struggles with cumulative reasoning and building on partial gains. He sees the strongest future in human-AI collaboration: AIs can map out large problem spaces, surface easy wins, and handle tedious auxiliary tasks, while humans focus on the hard conceptual gaps and on making sense of results. He also thinks AI will change how mathematicians learn, write, and communicate, but that the social and persuasive side of science will remain important and may be difficult to automate.