For those who are out of the loop, here is a TLDR of the situation:
Joint statement by the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative.
On November 20, 2026, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library comprising 17 million lines of code.
We have full confidence that Alpöge’s proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.
Source: https://leanprover-community.github.io/blog/posts/riemann-hypothesis/


It’s not a stochastic parrot or autocomplete. Well, not just those things. Yes, at their core LLMs are still just predicting tokens, but there’s been a lot of work on harness software and model improvements (though it blows my mind that they can apparently generate mathematical proofs while still being unable to consistently count the instances of “r” in “strawberry.”)
That said, I have my doubts. In a Lean proof that long, there’s ample opportunity to find and exploit unknown flaws in the kernel, and no human has had the time to check it (let alone the absurd 250,000 pages of the argument in prose).
I’m going to withhold judgment until an actual team of mathematicians can verify it.
Yeah it’s crazy but I think at this level of math a lot of the proving happens in concept space and not necessarily in number space.
It cannot count letters, but it can write a script that does it. Chatbots are toys, llms are stochastic parrots, but a well built agent using llms to write code is very capable. And I am not a mathematician but my hunch is that the AI companies basically treat these math problems as coding assignments. They are not looking for new discoveries, they want to code proofs for theorems that mathematicians formulated.
Sure. Little one-off scripts like that, they’re very good at.
Eh. I think this is less a function of them being capable and more a function of how much money you’re willing to spend to brute-force it. A limited token budget severely restricts the value of LLM coding, because no matter how good the harness and how carefully considered the prompt, it will inevitably make mistakes, and if you run out of tokens those mistakes are extremely costly.
If money is no object, you can force your harness to keep working on a problem until it eventually stumbles on the solution (or an exploit that allows it to appear like it’s found the solution). But nobody has an endless budget for that.
Agreed, that’s exactly what they’re doing. And that’s why I’m withholding judgment until there’s a human team of mathematicians who can verify the results. Lean is great, but it’s not perfect, and there’s ample opportunity in a multi-million line Lean proof for errors that the Lean kernel doesn’t catch.