Achievements of Artificial Intelligence (AI)

Lean's used pretty widely now by mathematicians so I assume it gets a lot of scrutiny. But there can always be bugs. Wiki says:
In 2026, AIs designed to spot bugs in software were directed at Lean, and found several loopholes which were then fixed. Perhaps related to this effort, a purported disproof of the Collatz conjecture was announced as verified in Lean. However, this proof was soon determined to rely on a bug in Lean, and once the bug was fixed the proof was found invalid.
 
re: Proof of Riemann Hypothesis

Do we know that proof checking programs (LEAN, in this case) are both complete and free from error?
Not yet independently verified. There’s been a lot of ongoing discussion of the attempts here for the Alpöge work which was already formalised in Lean.
And on the work formalising Masullo’s work here.

There’s also more on Navier-Stokes, it's been a busy day it seems with various claims at different stages of development. Should keep people busy for a while!
 
In 2026, AIs designed to spot bugs in software were directed at Lean, and found several loopholes which were then fixed. Perhaps related to this effort, a purported disproof of the Collatz conjecture was announced as verified in Lean. However, this proof was soon determined to rely on a bug in Lean, and once the bug was fixed the proof was found invalid.

Two thoughts:

First, there seems a danger from a kind of circularity developing between these two entities cross-checking each other. Errors may tend to compound.

Second, proof checking software is produced by humans, at least so far (and if produced by AI then we are back to the first point). But for human produced software to work on AI produced results requires us to understand how AI works, and how to check its working/reasoning. Might be a serious risk of its formal proofs becoming unverifiable by humans. Not sure where that leads.
 
There’s been a lot of ongoing discussion of the attempts here for the Alpöge work which was already formalised in Lean.
A bit surprised by the confidence of the following statement but this seems big.

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.

I read the full statement and it seemed they did basically every single common sense test. Bar a major major bug affecting multiple implementations of Lean and other proof checkers, I think GRH is solved. That’s ridiculous.
 
Last edited:
A month ago Terence Tao was asked: Is the Riemann hypothesis going to fall soon? And he responded:

Source:

I’m a fan of numberphile but this video is really good. Not just on this particular event and the ongoing debates in mathematics but on the wider description of AI usage and the partnership between humans and the tools we now have.

Thoroughly recommend anyone listens/watches.
 
A bit surprised by the confidence of the following statement but this seems big.



I read the full statement and it seemed they did basically every single common sense test. Bar a major major bug affecting multiple implementations of Lean and other proof checkers, I think GRH is solved. That’s ridiculous.
Yes. Concerns over LEAN tend to be overblown, especially in comparison to human proofs even if current work will undoubtedly still often contain erros (see also this post by Buzzard, one of the fathers of LEAN). The Collatz conjecture example, seems to not be of too much relevance anymore (see also the above post by Buzzard). Almost nobody is concerned about this anymore and the AI companies seem to be at the end of their LEAN investments as well.
 
As an ignorant observer, I was wondering: is any of the ambivalence we’re seeing from mathematicians related to LLM-generated proofs due to them being less elegant? In Levent Aplöge’s own statement regarding his proof of the Riemann Hypothesis, he refers to it in its current form as
[…] a massive unreadable travesty of ‘slop.’
Are these LLMs proofs brute forced or ugly? Or will their elegance become apparent when mathematicians have had time to understand them?
 
In Levent Aplöge’s own statement regarding his proof of the Riemann Hypothesis, he refers to it in its current form as
That was interesting. And I’m not sure. Did he consider the LEAN elegant? Was it the lack of time and perhaps the time needed to construct a human readable argument that was missing? He definitely considered what Claude was doing towards the end something quite special.

Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line

I’m no expert here at all. But I think he may have been referring to the long conversation he had with Claude as ‘slop’ or perhaps more likely any half baked interpretation he tried to write in human legible formulae now. His use of slop is quite different to that which some anti AI people use when referring to the output of LLMs

The parts in which my vague intuition worked out due to Claude’s technical mastery are of course less familiar to me today, and probably I’ll have to edit some of these descriptions on reflection. The parts in which Claude is the expert and not me I am sort of paraphrasing a "slop" summary, which might make zero sense, and I apologize, though the Lean should suffice to clarify everything. I’m just too inexpert in the relevant stuff at the moment, I hope the reader sees the breadth of this argument!

Or will their elegance become apparent when mathematicians have had time to understand them?
That seems the hope of many. That there is more discovery more elegance and more understanding to come. But it may take some time. I’m interested to see how it goes and what the experts come up with.
 
Last edited:
There are 2 kinds of AI proof happening but the answer is sort of yes for both.

AI-generated LEAN proofs: I think LEAN can be a drag to read at the best of times (maybe analogous to machine code or assembly for programmers), and on top of that Levant's LEAN proof is over 15 million lines long. So yeah, pretty impenetrable at the moment.

AI-generated human-readable-math proofs: in some cases AI is being used to generate readable math directly, or in conjuction with a LEAN version. These are also supposed to be very unpleasant to read according to all mathematicians I've heard from (which I assume means it adds unnecessary complexity, uses confusing notation, is hard to follow, doesn't break down the argument into nice lemmas etc.). But people have been cleaning these AI proofs up and then sharing them like normal math. (Unclear how this will go with the GRH human-readable-proof as it's apparently 250,000 pages long atm.)

There's also the aspect that's been mentioned here before that what mathematicians are really excited about is understanding the new tools someone invented to solve the hard problem. I think maybe from the outside we'd think a major open problem being solved in math is like cancer being solved for medical science, and in some ways it's like that, but in many ways it's a little closer to AlphaGo beating Lee Sedol -- many of these problems are important because we culturally decided they were important, not necessarily because nature forced us to care about them.
 
As an ignorant observer, I was wondering: is any of the ambivalence we’re seeing from mathematicians related to LLM-generated proofs due to them being less elegant? In Levent Aplöge’s own statement regarding his proof of the Riemann Hypothesis, he refers to it in its current form as

Are these LLMs proofs brute forced or ugly? Or will their elegance become apparent when mathematicians have had time to understand them?
A 250,000 page proof seems like a problem that exceeds mere elegance
 
I think maybe from the outside we'd think a major open problem being solved in math is like cancer being solved for medical science, and in some ways it's like that, but in many ways it's a little closer to AlphaGo beating Lee Sedol -- many of these problems are important because we culturally decided they were important, not necessarily because nature forced us to care about them
Thinking perhaps in a different way on the Lee Sedol angle, do you think we’re seeing novel approaches here to some of the recent discoveries? Do you think we may see different proofs too or in this case will the different people will end in the same place via the same path?
 
Thinking perhaps in a different way on the Lee Sedol angle, do you think we’re seeing novel approaches here to some of the recent discoveries? Do you think we may see different proofs too or in this case will the different people will end in the same place via the same path?
Yes, there are novel approaches and we are seeing completely new proofs that people wouldn't have come up with. Initially people were very concerned that they weren't sufficiently novel enough, but more just based on lots of data, involved smart tricks or counterexamples. Opinions still diverge slightly. Initial discussions were all about the AI just filling out "convex hull", this isn't the case any more and the opinions are also changing quite swiftly, most don't seem to hold onto the "convex hull hypothesis" anymore and I think opinions will adapt quite quickly as well. There was never a really meaningful concept of "convex hull of mathematics" in the first place.
 
Back
Top Bottom