Will AI Prove the Riemann Hypothesis Without Understanding It? – Terence Tao
TL;DR
Tao thinks Riemann likely needs new math, not just brute force — he contrasts it with the four color theorem, which was essentially proved by exhaustive case analysis, and says Riemann “doesn’t feel like” that kind of problem.
A disappointing AI outcome is possible: the hypothesis could just be false — Tao mentions the unlikely but real scenario where a computer simply finds a zero off the critical line, turning the whole saga into a giant verification exercise instead of a deep conceptual breakthrough.
The most plausible path is human-AI collaboration, not a one-shot autonomous Lean proof — Tao says the winning dynamic may be a type of collaboration “that doesn’t exist yet,” with AI helping generate variants, patterns, and transformations humans can turn into theory.
Lean may actually make machine proofs more inspectable, not less — because every lemma can be studied atomically, mathematicians can isolate which steps are boilerplate and which ones contain genuinely new constructions or ideas.
Tao expects a future profession of proof refactoring and proof interpretation — he imagines mathematicians and AIs doing “ablation” on giant Lean proofs, removing parts, compressing arguments, and even using other AIs to judge elegance.
Once a proof exists, post-processing it may be easier than people fear — citing the Erdős problem website, Tao notes that AI-generated verified proofs have already been followed by AI summaries and human rewrites, suggesting incomprehensible first-pass proofs can still be unpacked.
The Breakdown
Could AI prove Riemann in inscrutable Lean code?
The clip opens with the core anxiety: if we keep training AIs to solve harder and harder problems in Lean, could one eventually prove the Riemann hypothesis in a way that works formally but gives humans almost no insight? Dwarkesh frames the nightmare version as “assembly code gobbledygook” — a proof that verifies but doesn’t deepen mathematical understanding.
The four color theorem is the cautionary example
Tao’s answer starts with humility: “we don’t know.” He points to the four color theorem as the classic case where brute force more or less won, and says we still don’t have a conceptually elegant proof. His broader point is unsettling but clear: some theorems may only yield to gigantic case splits plus computer checking, and maybe that’s just how reality is.
Why Riemann still feels different
Even so, Tao says part of why mathematicians prize the Riemann hypothesis is the strong intuition that solving it should require “some amazing” new mathematics or a surprising bridge between previously unconnected areas. We don’t know the shape of the solution, but it doesn’t feel like something you settle by checking cases. Then he adds a darkly funny twist: if the hypothesis is false and a machine simply exhibits a zero off the line, that would be mathematically decisive — and emotionally “very disappointing.”
Tao bets on collaboration, not full autonomy
Tao says he doesn’t think “fully autonomous one-shot approaches” are the right model here. Instead, he expects much more mileage from humans working with extremely powerful AI tools, though the exact collaboration style may be unlike anything that exists today. His example is very concrete: imagine generating a million variants of the Riemann zeta function and using AI-assisted data analysis to discover patterns that reframe the problem in a different branch of mathematics.
How do you spot the hidden gem inside a formal proof?
Dwarkesh asks the key follow-up: what if the breakthrough is latent in Lean code, but nobody recognizes its significance — like Descartes inventing coordinates, except in formal syntax it just looks like some boring type signature such as “R to R”? Tao’s answer is that Lean’s real advantage is granularity. You can inspect each lemma atomically and ask which parts are standard boilerplate and which parts feel genuinely new and load-bearing.
Lean proofs can be dissected, simplified, and rewritten
Tao imagines an entire future profession built around taking giant Lean-generated proofs and doing “ablation” on them — removing steps, finding cleaner routes, and using reinforcement learning or other AIs to optimize for elegance. He also notes that modern AI is changing the economics of mathematical writing: papers used to be costly to refactor, but now once you have one version, you can generate hundreds more. So a massive messy proof may be ugly at first, but it becomes raw material for interpretation.
He’s less worried about incomprehensible proofs than other people are
Tao ends on a surprisingly optimistic note. He points to examples from the Erdős problem website where an AI generated a proof, thousands of lines of code verified it, then other AIs summarized it and humans rewrote it in more intelligible forms. His takeaway: a first proof artifact may be messy, but once it exists, we increasingly have tools to analyze, compress, and understand it.