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

Terence Tao tells Dwarkesh Patel that AI has collapsed the cost of hypothesis generation but verification and strategic judgment remain the bottlenecks blocking scientific progress.

• AI drove hypothesis-generation cost to near zero, just as the internet drove communication cost to near zero — abundance alone doesn’t create progress.
• Verification infrastructure, not idea generation, is now the binding constraint; journals are already flooded with AI-generated submissions.
• Copernicus’s heliocentric model was initially less accurate than Ptolemy’s geocentric model — correct theories often look worse before refinement, making RL-based peer review structurally hard.
• Tao wants a semi-formal language for mathematical strategy and conjecture-plausibility, analogous to what Lean provides for deductive proof — but admits it’s a wish, not a plan.
• If Riemann hypothesis proved false, Tao says he would immediately abandon prime-based cryptography, since one unknown pattern implies more.
• Human-AI hybrids will dominate frontier math for considerably longer than pure-AI replacement timelines suggest; current AI lacks key missing ingredients.
• Serendipity and inefficiency have scientific value — over-optimization (COVID-era scheduling, instant search) eliminates the accidental discoveries that drive paradigm shifts.
• High-school students could now contribute to math research frontiers using AI and Lean, collapsing the traditional PhD gatekeeping timeline.

2026-03-20 · Watch on YouTube