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/

  • spectrums_coherence@piefed.social
    link
    fedilink
    English
    arrow-up
    8
    ·
    edit-2
    5 hours ago

    There is several layer to this.

    First, just being a stochastic parrot is already really useful in research, it has absorbed all the written human knowledge perhaps several times. Having such a database that is searchable via natural language is already quite valuable.

    Second, modern LLM are generally trained two stages, 1) pre-training, which it trys to correctly predict the next word 2) reinforcement training, where it try to solve problem (math, coding, robotics) and its network evolves towards the weight that can solve these problems.

    While pre-training gives a scholastic parrot, reinforcement learning is what makes it “useful” (depends on your definition of useful) in math, robotics, and coding. A vast amount of knowledge are stored in natural language, in some sense, the representation of language, i.e. stochastic parrot, unlocked the huge amount of problem to reinforcement learn on. Also the development of formal methods also gives the foundation for these reinforcement: it allows the correctness of result of LLM to be verifiable without human in the loop. (Yet, less than .001% of the AI money have been flowing into ACTUAL formal method research)

    AI company are mining (or at least care about) these math problem, because 1) they are easy to mine with verifiable reward 2) they hope intelligence in math can generalize to other domains.

    Third, there is an expert human that is skilled in both mathematics and managing LLMs, running with unlimited tokens and bespoke harnesses.

    Finally, GRH are really much closer to being solved than we realized, since a human proof is already being reviewed several month ago and confirmed correct by the referee (with no lean certificate). The human proof, of course, is much much cleaner, and have zero AI involvement.


    Here are some caveats:

    On the math side: unlike what Kevin Buzzard will tell you, correctness is absolutely overrated. It is true all reusable results need to be correct, but mathematics are useful to the world because it creates simple and principled insight to power other fields. That is a “wrong proof of RGH with correct transferable insight” >> “a correct lean slop that takes decades to understand”. That being said, there likely are many insights to be mined from the lean slop, and Levent is commited to do that. But confirming the lean proof is correct is a much much smaller deal than a layman would assume.

    On the AI company side: you might wonder why Claude are terrible at archtecturing software yet it can solve GRH. Because the architecture of software don’t really have a good verifiable reward: we archtect software in a way that is easy for human to understand and navigate, there is no way to definitively check that a software is “well-archtected” as opposed to “all test has passed” or “code compiles”.

    in the same way, AI produces absolutely terrible math proofs and have a hard time to distill insight, because there is no definitive way to say “what is the spark within this proof”.

    Finally, I want to emphasize that “software architecture” and “insight” are likely not unsolvable, just much harder than press a button to see if the test passes. AI company are certainly working on it.

    But the real bottleneck, IMO, we do not see that intelligence in math being generalize well to other fields. They still need bespoke training for these models to do accounting, even though most people would assume solving the GRH would imply you can do accounting.

    This is likely why they are pushing RSI, because I think both pre-training and reinforcement learning are seeing diminishing return now.

    On the money side, AI company are likely burning through more money than all the funding human mathematicians in history combined. Although these model are already here, and inference are cheaper than mathematicians, but people often neglect the mountains of R&D and training cost. So building newer and better LLM just to solve a couple of the most important problems in mathematics might not be a cost-effective deal for anyone.

      • spectrums_coherence@piefed.social
        link
        fedilink
        English
        arrow-up
        2
        ·
        edit-2
        1 hour ago

        “Recursive self improvement”, which like “artificial general intelligence” is not a well-defined term.

        Their current pitch is that AI is going to lead to exponential growth of AI research itself, thus they will be able to find the “next big thing” soon.