The world is too loud. Read what matters.

Andrej Karpathy

If AI proves the Riemann Hypothesis and no one understands, mathematicians say it doesn't count

A thought experiment about incomprehensible proofs reveals: proof's value isn't about settling truth to 100%, but about whether humans can understand it—that's what mathematicians truly care about.

Riemann HypothesisMathematical proofAI interpretabilityLeanAI alignmentPhilosophy of mathematics
The episode revolves around a hypothetical question with multiple perspectives clashing; no new results or data, but it articulates the divide between ‘proof’ and ‘understanding’ clearly—worth a quick listen.

The argument · tap a timestamp to hear it

0:12

Incomprehensible proofs horrify some mathematicians but amuse others

When AI produces a billion-line Riemann Hypothesis proof with no explanation—just empty definitions like ‘let A be a function, let B be a function’—reactions are completely divided. Some find this absurdly funny; but mathematical romantics find the result ‘horrible’ because they want human participation in proofs and want proofs themselves to be beautiful. Still, the search for elegant, conceptual proofs doesn't become less important.

1:12

A proof no one understands cannot count as a proof

A logician studying metamathematics is blunt: this scenario isn't even dangerous, it's not worth caring about at all. If a Lean proof runs to two hundred billion tokens and nobody can read it, it's not a proof of the Riemann Hypothesis; ‘I simply don't care it exists’. This doesn't mean the problem is solved, only that focus shifts to understanding that gibberish or finding another way to get at the truth.

2:13

Multiple Lean kernels verify computational proofs with near-zero error risk

First, take the established formal statement of the Riemann Hypothesis, run it through the verifier, check whether the proof holds and has no gaps. If two or three different Lean kernels all conclude ‘this really is a proof’, credibility is already very high. Even if at first nobody fully understands, a proof isn't a static thing—people break it apart, isolate pieces, pour huge time into understanding these objects. Once convinced it's correct, that itself will be a great day for mathematics.

3:13

Some theorems can only be solved by computational brute force

The Four Color Theorem is a famous example: no conceptually elegant proof has been found and may never exist—some problems can only be split into vast cases and checked one by one by computer. The Riemann Hypothesis is treasured precisely because everyone believes solving it requires creating entirely new mathematical tools or bridging two previously unrelated mathematical fields; nobody knows what the answer will look like. And there's an unlikely but real scenario: the conjecture is actually false, and some massive computation finds a zero off the critical line—that would be devastatingly disappointing.

5:14

Moving certainty from 99.9% to 100% adds minimal value

If AI produces a Lean proof but no explanation, and that explanation would exceed a human lifetime of reading, the proof barely matters—mathematicians are already highly convinced the Riemann Hypothesis is true, pushing that certainty from 99.9% to 100% is just ‘a bit of information’ and won't truly change a mathematician's life. But if the proof is merely poorly explained rather than fundamentally incomprehensible, that's completely different: almost all mathematicians would drop what they're doing and pour everything into translating it into something humans can understand.

7:16

AI systems must explain proofs, not merely produce them

Someone points out this is their biggest disagreement with other conference speakers: these models should be superhuman explainers on top of being superhuman provers—they already explain to professionals with great patience and without judgment knowledge they ‘should have known but didn't’. If that explanatory ability extends from teaching textbook facts to explaining entirely new proofs, models could break down massively complex proofs and translate them into human language—exactly what they'd be specially good at.

9:18

Understanding matters more to mathematicians than formal proof

Even if another mathematician claims to have proved the Riemann Hypothesis, the first reaction isn't congratulations but ‘what's the answer, explain it to me’—understanding is the whole point. Proof is core to mathematical epistemology precisely because humans have suffered too many losses betting on intuition, like false proofs that look right but prove 1 equals 0. Human intuition is powerful but unreliable, so strict proof within axiomatic systems is needed. But understanding happens elsewhere; proof is just the gold standard for translating ‘what we see very clearly’ into rigorous argument. The discussion also notes that one of the strongest models, like O3, has been called ‘a very famous liar’, raising alignment concerns.

10:18

Problem-solving becomes commodified, only rare hard problems retain value

Historically, once an ability becomes cheap or commodified in society, what people value shifts toward what's orthogonal to it. Certain types of problem-solving ability are becoming and will continue to become cheaper and less valued, but landmark big problems like the Riemann Hypothesis might still survive as scarce hard problems—though the speaker admits they could be wrong.

11:19

Mathematics has not yet experienced its Move 37 moment

Move 37 is the move AlphaGo played in defeating a top human player, commonly used to ask whether a field has had its moment of complete machine superiority over humans. For mathematics, the answer is ‘resoundingly no’—the real Move 37 might have to wait for machines to prove the Riemann Hypothesis or some Millennium Prize Problem, or even propose a deep theory bridging different mathematical fields at once. And even when that moment comes, solving a big problem is never the endpoint: what solving a problem produces will be used to generate new problems and conjectures, and exploration of integers and primes will continue.

In their own words · checked verbatim

you press a button and you get this billion line long proof of the Riemann hypothesis and it's just like, let A be this function, let B be this function, let C be this number, let D be this function

if there's a lean proof of the Riemann hypothesis that's like, you know, 20 billion tokens long and nobody can read it, that's not a proof of the Riemann hypothesis and I just don't care that that exists.

The four color theorem is a famous example. We have still not found a conceptually elegant proof of this theorem.

Raising that confidence from, you know, 99% or 99. 9% to 100% is a little bit of information. Would it like change my life as a mathematician? Not super much.

Human intuition is incredible and it's also very faulty.

many of the most capable models like O3, for example, was a very notorious liar.

Every time that a theorem or open problem is solved, to be good mathematics, the solution isn't the end.

Glossary

Riemann Hypothesis
A famous unsolved conjecture in number theory about prime distribution; one of the Millennium Prize Problems.
Lean
A formal proof assistant programming language that automatically verifies the correctness of mathematical proofs.
Move 37
A famous move AlphaGo played in defeating top human players; symbolizes a moment of complete machine superiority over humans.

How to listen

Who it's for

Engineers, researchers, and founders interested in AI's mathematical limits, interpretability, and philosophy of mathematics.