this post was submitted on 20 Nov 2026
68 points (80.4% liked)

science

30028 readers
602 users here now

A community to post scientific articles, news, and civil discussion.

dart board;; science bs

rule #1: be kind

lemmy.world rules

founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
[–] xxce2AAb@feddit.dk 1 points 19 hours ago (2 children)

That's what I'm guessing. The LLM - in a hapless attempt to please - will have shat out a Lean "proof" that does check out, but that is utterly uncorrelated to the claim.

[–] red_tomato@lemmy.world 6 points 13 hours ago

The proof uses Lean’s built-in definition of the Riemann hypothesis. It’s just a few lines long, and has been agreed to be correct for many many years.

[–] trillian@feddit.org 4 points 14 hours ago (1 children)

Have you looked at it? It uses the Mathlib formulation of the Riemann hypotheses.

[–] xxce2AAb@feddit.dk 5 points 9 hours ago* (last edited 9 hours ago) (2 children)

No (and nobody else has either). However, somebody vastly more qualified did and barely got started digging before the first sinkholes appeared:

Jenny Lorraine Nielsen: "Invalidity of the Alpoge Proposed Proof of the Riemann Hypothesis"

On 20 November 2026 Levent Alpöge announced a proof of the Riemann Hypothesis and of the Generalized Riemann Hypothesis for Dirichlet (L)-functions, released only as Lean 4 code. The archive holds 175{,}187 Lean files. Inspection of the source shows that the construction supplying the final hypothesis bundle rests on witnesses whose required properties are vacuously true. The geometry predicate is defined to be (\mathtt{True}). The geometric witness is built from a model the code itself names degenerate. The lattice witness has its required properties by definition. A statement that fixes a line-bundle degree holds only because its hypothesis cannot be satisfied. The lemmas written to extract content from these witnesses are never called. The pieces that would identify the row-side geometry with the divisor-side witness are the vacuous ones, so the chain between them is broken. On the path to the conclusion the height equation holds by (\mathtt{rfl}), because the height is assigned to be its own formula. The formalization therefore does not establish the geometric realization or the lattice identification that its own structure names. The proposed proof is invalid: what has been checked is a term whose structural witnesses are vacuously true, not the argument those witnesses stand for.

Source: https://philarchive.org/rec/NIEIOT-2

Existence of a Lean4 "proof" means very little in and of itself.

[–] General_Effort@lemmy.world 3 points 2 hours ago (1 children)

somebody vastly more qualified

No kidding. They have their own proof of Riemann, as well as a complete gauge theory.

Anyway, keep up the critical thinking.

[–] trillian@feddit.org 6 points 6 hours ago* (last edited 6 hours ago) (1 children)

This doesn't matter as long as there is a term inhibiting the type corresponding the Riemann hypothesis and which is built following the accepted axioms (so no sorry). This has been checked. Using different Lean kernels, so the probability that a bug in Lean might have been exploited is very low as well. Of course, there might be some intermediate statements/lemmas that hold vacuously, but as long as their correctness is not required for the proof to hold it doesn't matter. They might just be dead code or helpers that are not required (The lean file is ~17million lines of code, so that is expected). If you are actually interested in things you have to pay attention to when verifying lean code, take a loot at a recent blog post by Kevin Buzzard: https://xenaproject.wordpress.com/2026/11/12/and-the-proof-is-checked-in-lean/

I really don't understand the Lemmy LLM stance, dismissing all of its achievements. I am not "Pro-AI", but it's very dangerous to ignore the societal changes that will follow from its existence.

Verification progress: https://math.mit.edu/~drew/grh/

[–] xxce2AAb@feddit.dk 2 points 5 hours ago* (last edited 4 hours ago) (1 children)

I can't speak for anyone else, but in my case (and in this context) it's based on a history of overinflated claims. See, for example:

https://www.argmin.net/p/faster-fourier-transforms

https://www.argmin.net/p/galactus-the-devourer

Of course, that has no direct baring on this result, and it's good to hear different kernels were used as that addresses one of my other immediate concerns.

Thanks for the link. I'll be following this with some interest.

[–] trillian@feddit.org 2 points 4 hours ago* (last edited 4 hours ago)

Regarding the overinflated claim examples: I think this a misunderstanding between complexity theory and "everyday notion" of fast. In Complexity Theory, one is usually, interested in the asymptotic behaviour, i.e., how does the "performance" (number of steps) of an algorithm scale with the size of its input. In particular, an algorithm that runs for 10³⁰⁰ · |x| steps on an input of size |x| scales in the same way (linearly) as does an algorithm that runs for |x| steps, despite being massively slower. So the notion of "fast" or "efficient" that is sometimes used in complexity theory is not the same as one might expect if not familiar with the field.

For example, concerning the exponential time hypothesis (ETH your second link), OpenAI did in fact find an algorithm that scales subexponentially (instead of exponential), thereby disproving the hypothesis, which was very unexpected to say the least. Of course, the algorithm (as noted in that post) is not practical, but that was never the goal (here the AKS primality test comes to mind, which checks whether its input is a prime, showing that primality checking can be done scaling polynomial in the input size).

Actually, the ETH hypothesis has been considered as a way to tackle the P = NP millenium problem as ETH directly implies that P ≠ NP. Again, for these problems the question only lies in how an algorithm solving them scales with the input (i.e. asymptotic complexity) and not if the algorithms are practical. But people tend to describe/explain P as "problems which permit fast algorithms" conflating the complexity theory notion of fast with its everyday counterpart.

tl;dr: There are different notions of "fast". From a theoretical point of view, these algorithms are not practical. But that was never the goal as the problems are not concerned with practical algorithms but rather with an abstract notion of "fast"