Dwarkesh Podcast · Dwarkesh Patel

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

March 20, 2026·1 hr 24 min·5 clips
Kepler turned Tycho Brahe's data into ellipses, equal areas, and the third law.
1. Dwarkesh Podcast brings Terence Tao into a discussion of Kepler, Newton, and AI for mathematical discovery. 2. Tao traces Kepler through Copernicus and Aristarchus to Tycho Brahe's decades of naked-eye observations from a Danish-funded island observatory. 3. Kepler first tried to fit planetary spacing to the cube, tetrahedron, octahedron, and dodecahedron as Platonic solids. 4. Brahe guarded the data, and Tao says Kepler only got enough of it after copying observations and fighting over access with Brahe's descendants. 5. The measurements broke Kepler's circle theory by about 10%, pushing him toward ellipses and the equal-areas law. 6. Tao says Kepler's third law came later, after he had only about six data points for orbit size and period. 7. Newton later explained all three laws with F = ma and centripetal acceleration, giving the inverse-square law. 8. Tao and Dwarkesh use Johannes Bode's planetary-spacing rule as a warning, because it fit Uranus and Ceres and then failed with Neptune. 9. Tao argues that AI has driven the cost of idea generation down to almost zero, much like the internet did for communication. 10. He says the bottleneck has moved to verification, evaluation, and deciding which of thousands of AI-generated theories deserve attention. 11. Tao notes that journals are already flooded with AI submissions, so peer review cannot scale as a filter on its own. 12. He compares this shift to the 20th century's move from theory-and-experiment to simulation, big data, and regression. 13. Tao says Kepler was an early data scientist, but even he began with preconceived theories before fitting Brahe's data. 14. He argues that deep learning, transformers, and even base-10 numbering all became dominant through adoption, inertia, and test of time. 15. Tao points to Bell Labs and the 1940s as an example of one result affecting probability, computer science, and signal transmission at once. 16. He says human science still needs ways to identify unifying ideas among millions of papers, not just one-off technical fixes. 17. On AI for math, Tao says roughly 50 of about 1,100 problems have been solved, but the burst of pure one-shot solutions has paused. 18. He describes the remaining problems as a mountain range of walls, where AI can jump some low cliffs but still struggles with partial progress. 19. Tao and Dwarkesh compare software and math, with Tao saying code prototypes are easier for AI than maintenance, integration, and mathematical intuition. 20. He ends by suggesting AI could revolutionize the experimental side of math by testing thousands of problems instead of one.

As heard by us

Kepler's geometry gives the episode a clean frame for AI's limits in math.

The piece opens with Terence Tao tracing how Kepler built on Copernicus and Aristarchus, using the solar-system puzzle to frame a wider discussion of AI and math.

Read the full review in PlayNext →

Why you'd press play

Kepler, Copernicus, and AI for math: your geometry brain gets the full tour.

Read the full recommendation in PlayNext →
Listen to the show on