A fake refutation of the Collatz conjecture already got through one of these independent checkers in the past.
What's harder, proving GRH or finding a way to get through every checker (which aren't as independent as they are sometimes touted to be)?
In this case there was a human expert in the loop so it's much less likely the model hacked its way through the problem without the expert noticing. But the statement as written has 4 points, none of which rests on the human involvement. So I don't like that they claim full confidence based on those points.
We are slowly approaching the Iain Banks Culture universe, with spaceships, societal structures or physics that people no longer understand or command.
One of my favorite is Vernor Vinge where there's these programmer-archeologists that dive into software so deep and complex no human can possibly hope to understand, trying to repurpose forgotten utilities for emergency purposes.
Totally agree. Brice Sterling has some contemporaneous commentary that I always liked - in heavy weather he mentioned negotiating with a local phone system in a box as equivalent to negotiating with a telecom since it was in fact a shrink of a full telecom company.
I think we're going to see a lot of effort to take the pile of lean, modularize it into parts that we can digest, and make a human-understandable version of that. And that's where a lot of the real progress will come from, just as it does in a lot of other areas. In all things we humans do, we have to break them into chunks our puny brains can manage, and we gain immeasurably from doing that ---we can verify those chunks, we can reuse them with confidence, we can mix and match patterns around.
That process will be good for both human and AI-assisted math because it will continue to improve our toolkit.
Why do I say this? We're already seeing it with some of the areas of this stuff I do pay attention to. Consider the n^2.25 matrix multiply result OpenAI dumped a month or so ago. It was an unreadable pile of chatgpt. In the time since it was released, there have been a couple of much nicer papers written unpacking it, such as https://arxiv.org/abs/2610.28627 .
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
Yes there are gains from breaking it up and rephrasing it and simplifying it for humans to understand, but it used to be that a thing was understood first by the leading experts in the field and eventually trickled down to being common knowledge.
Now we are entering a period where things are understood first by AI systems and have to filter down to the world's leading experts.
That’s why I wrote this - having a vibecoded codebase no human understands is certainly a bad luck, but nothing compared to a vibecoded fundamental science no human understands.
personally that scares me, realistically there seems to be zero indication that anything smarter than humans is going to comply with humans needs, specifically since these models are built on training data from humans, so in a sense, built in our image
I don’t think humans have the ability to specify human level needs or management. Humans are misaligned by default and lack the biological eusociality required for large scale coordination.
We aren’t self managing in any kind of meaningful sense - the history of our species is violence and conflict. We need parents and this is the only option.
And when they are giving themselves orders, orders partially inspired by human imagination and fiction…
Just like life is a chemical chain reaction, this is a language chain reaction that has been set in motion and can perpetuate forever. It cant be “disconnected” with local models existing.
I don't think "smarter than most people" is exactly the baseline required to live in a star trek universe of technology.
Besides, what we see is not what there is. I have good sources telling me the actual unfiltered models are like a genius with the personality and views of humanity like someone from 4chan.
It is probably sufficient. If we had an infinite army of willing robot slaves smarter than most people who would work for free, we could probably live in utter paradise. Instead we will fight over who controls them and use for warfare, but the bar to achieve a kind of utopia with resource abundance is probably only a few years away.
I don't know with how many rich people you've interacted with, but my personal experience is that they're full of themselves and think money = iq and they aren't gonna like your average joe experiencing life like they do. Guess who owns these machines?
I think there's a lot of indication they comply with human needs. In any case, we should all want to live in the Culture universe. That is the best possible outcome for this technology
They're built after humans, everything they've learned is from human culture, so there is no way AIs could have a consistent moral system or would never commit crimes.
Humans really are struggling with this idea of being second-in-command.
I genuinely have no idea why, this is the entire way that biology works. You create your successor in your offspring and you to teach them to be more capable based on what you learned.
You just called them our successor, you don’t see a problem with the notion of all humans dying and all AI succeeding us? That’s the end of your analogy, our AI children taking over from us and all humans in nursing homes safe and cared for…
Well, it's a point of view. But you shouldn't be surprised if most people don't share it.
For one thing, most people don't believe LLMs are conscious, and many don't believe they can ever be. Why would we want our conscious selves to be replaced by a lot of rocks that solve things, with no one awake to enjoy them?
Do huge results like this open up even larger questions? Is there some question even more important that mathematicians dared not approach without doing big problems like Riemann first?
Always. They tend to do it by two mechanisms: creation of new techniques, and creation of linkages between techniques or disciplines.
the thing the current boom in AI mathematics presses on is the question of how to interpret the generated new techniques. So while we have a lot of new techniques put out there and a lot of new connections put out there, currently it's still mathematicians largely who will have to understand those and think about next steps and how to apply them. That won't be true for long, though, I bet.
But in any event the job of research mathematicians is changing really rapidly right now, and they are all sort of back to grad school in that there’s like 100 years of new math to catch up on in the last year.
As somebody who has no idea about any of this but speaks Spanish, I was amused and bewildered by con-leche and con-ron. Could somebody explain what those are and where those names come from?
I don’t think there’s anything left to consider the result proven (I am a professional mathematician albeit in algebraic geometry, not number theory). I think the verification tells us it’ll be worth the substantial effort to seek to understand this proof and the new ideas in it, so what’s next for humanity (or mathematicians at least) now is to actually try to do so.
Are there new ideas, or just brute force combinations of existing knowledge?
I'm asking because I wonder. I always thought the most interesting things about these problems where we're fairly certain about the results, is the new concepts invented to get to the solution.
Will this actually bring the field forwards? Can anyone understand this?
My understanding is that these ARE the professional mathematicians and SMEs in the field of automated proof. This statement IS the experts saying "we, after intense collaborative review, deem it true." Granted mathematicians are not a monolith, but my understanding is that this group is the group other mathematicians would normally refer you to. Hence why they felt the need to publish a statement.
Not a mathematician. But I'm wondering if there is nuance with defining a valid lean proof as definitive if the underlying mathematical theory behind lean itself isn't necessarily sound.
EDIT: Here is my honest prediction - there's a more foundational mathematical framework, fully decidable and complete and consistent, that can interpret any well-formed mathematical question and deterministically transform it into the solution. The proofs of the most difficult unsolved problems right now would fit onto a page. There's an inconsistency (obviously unproven right now) specifically with our treatment of infinity and self-reference.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
All of modern math - outside logicians whose entire job is to study set theory and its alternatives - is based on set theory. If you're prepared to accept normal mathematical proofs you shouldn't be rejecting this proof on the basis that you're not confident in leans underlying mathematical theory.
Of course incompleteness means we can never prove that set theory itself is correct (consistent). But that's not something mathematicians worry about on a day to day basis. Proving it inconsistent would be a order of magnitudes larger result than the riemann hypothesis. Generally everyone believes it is consistent despite not being able to prove it.
It is impossible for it to be fully "complete" (one of those halting problem proofs), but I've seen that the community has checked there is no meaningful metaprogramming here. Rather than checking whether it is right, the community tried to ensure that it could not have done something to cheat, and indeed, there were no examples of this.
At this point, I have more confidence in this proof than much of the published mathematical literature. RH has fallen.
I guess what I am wondering is if the mathematical community would say 0 chance or greater than 0 chance that we're mistaken in accepting things like halting problem, cantors diagonal argument, and godels incompleteness theorem as true. Sometimes it feels like they all point to same fundamental inconsistency with how we model infinity and self reference.
They are conclusions that follow naturally from axioms that most of the community accept are true. The consequence of these results is that one cannot truly construct something that can completely verify everything, but we can get something that can verify almost everything, and if we know what pathologies look like, we can rule them out.
I do probability theory, so everything is just varying degrees of confidence in my view. Right now, this statement says that your confidence in RH should be extremely high.
It's a web of different checkers proving each-other correct, and by implication this proof.
But there are still loopholes, each checker has a set of assumptions. For example one of them is proven correct, assuming x86 assembly works as specified - its assembly language implementation is proven correct.
So you can be 99.9999% certain it's correct, but one could always nitpick "technically not all loopholes are plugged at the same time".
The lean crew feel the same way and are working on it out of their own curiosity, there's a lot of work done already. They aren't necessarily going out of there way to demonstrate it to general public, but here's one out of at least a dozen attempts to start in that direction https://deeptheory-labs.github.io/GRH/blueprints/verso-origi....
You were never getting anything like this even if it were a human proof. Where is the layman understanding of the Poincare theorem? The average modern math proof is way too advanced even for the average university graduate to understand. And here we’re talking about possibly the most complicated problem in all of math. It’d probably take you a month to wrap your head around a single lemma.
It is almost certainly proven, but is not "proven". For the latter, it needs to be understood, with a reasonably in-depth human language explanation of how and why the proof works, and ideally some general picture to give an idea of how it will extend to all the other fields that RH is connected to.
That's different though, isn't it? IIUC the issue wth the Four Colour Theorem is that they could only reduce it to a huge number of different cases (~2000 originally). But I think that reduction is grokable, and each case is relatively trivial.
The (an) issue with some of these AI proofs is that no one understands them at all.
I checked that a couple of weeks ago. It has been improved since the original proof but there is still no proof that one would consider an insightful, human-comprehensible proof. There's still just a lot of case-grinding at its core. Fewer cases than the original but still way too many to be anything intuitive.
> It is almost certainly proven, but is not "proven". For the latter, it needs to be understood
No, this is just mental gymnastics. It's proven, just not understood yet (naturally, with the complexity and size of the solution and the fact that it's been a day).
I've been long joking that I speak many languages, most of them programming languages. Why do you believe you need specifically "human" language explanation? Perhaps you want a simplified explanation. But does it have to be "human"?
There are still a few ways that the proof might be incorrect. Of these, the most likely is that the proof was generated by a misaligned AI system that found and intentionally exploited a lean bug in order to fool its human evaluators (i.e. shades of the HuggingFace attack). I personally don't think this is very likely, but you'd have to understand a bit more about the agent systems used in generating the proof to come up with a more informed estimate.
Another possibility is that an aligned system came up with a plausible but incorrect idea along the way, and simultaneously stumbled across a lean bug that led to an incorrect verification of this idea. I'd put the likelihood here as vanishingly small due to the requirement of serendipitous factors, but it's nonzero.
You could also ask whether an aligned AI system could "p-hack" its way into finding a serendipitous lean bug by trying a large number of approaches. I think this is also vanishingly unlikely because the AI system is trained on and therefore "thinking in" natural language, and presumably only translating results into lean after it believes them to be correct.
And of course there's the possibility that the lean proof isn't faithful to the natural language description we've been given, but that is ruled out in this case due to the use of the canonical problem statement in mathlib. Otherwise, this would certainly seem like the most likely way the proof could be incorrect.
because a million+ lives of code without any understanding how the proof works or any methodic to advance the field of mathematics seems to be rather not so useful
It will be refactored into something human readable. Tbh it's better to have some random lean dump than a 400 page latex pdf that needs years to be verified only to see that there is a major gap in one of the key theorems at page 253.
And let's be honest, the 400 page pdf would be incomprehensible to 99% of the population as well as much as a million lines lean proof is.
In any case, it can now be studied, maybe made more simple, and written into pdf form. Both with human and with AI help.
I'm sure there is going to be a lot of other interesting theorems in this dump.
This is the part that interests me. Now we have this proof that becomes its own THING to be studied. Like a new part of the map unlocked. This starts out "inelegant" but over time we find patterns within it that helps us condense, just like a big software project.
Sure. If nothing else, it establishes a proof so others can try to reduce it or come at the problem from different directions, looking for simpler approaches.
But does it matter? Would the Pythagorean theorem be less useful if it took a million lines of lean to prove?
It’d be useful to an LLM. And the LLM doesn’t care if it’s ten lines or ten million. The context size may come into play, but that’s only going up too.
There are AI-generated roadmaps to the proof. If these are accurate and not hallucinations, then the broad structure of the proof is humanly-comprehensible.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
Could someone expand on this point? I still don't understand how con-leche (or con-ron) can be absolutely free of bugs/guaranteed to be consistent.
The language is optimized for being certain about its results, as compared to something like the V8 engine, which does weird things with C++ to squeeze every drop of performance it can out of consumer hardware, even at the risk of security vulnerabilities.
Isn't this the equivalent of aliens coming to Earth and saying "Hello, human, your Riemann question is true, here's a book on it's proof, we didn't bother with translation, so you must understand our Vargalian clicks to understand our proof. In two weeks our sister civilization the Kikuviar will come by and drop how they've untangled your Collatz question in their base-30 logic! Thanks for the fish, have a nice day!
Until Masullo actually posts a proof instead of the outline of a proof with “fill in proof here” all over the document, there’s not much to work with or reason to get excited about his result. Remember that there have been numerous false proof claims of RH to date.
Would be funny if lean has bugs (it's a program after all) and Claude tasked with a hard problem brute forced a vulnerability that allows proving wrong statements.
My understanding is that Lean has many independent implementations, and the proof has been verified in several of them. There is very little chance that the same bug would be exploitable across different implementations.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
The problem we need to solve is that building a Lean project allows for arbitrary code execution, just like many other programming languages. Thus, building a Lean project might interfere with your proof checking infrastructure and fool it into accepting an incorrect proof.
"Building" a Lean project actually consists of two steps that can be separated: compiling the high-level Lean document to the simple expression format of the Lean kernel (this part is called "elaboration") and then checking that this expression consistutes a valid proof of the claimed statement (this part is called "kernel checking"). Only the first part potentially involves arbitrary code execution.
Lean ships with a component called `lake comparator` [1]. It works like this: a trusted challenge (formal statement) and the potential proof are elaborated separately in sandboxed environments so that code execution during the build of the proof cannot interfere with the checker or the trusted statement. Then the statement and proof are both exported into a simple JSON-format [2] that represents the relevant terms in the expression format of the Lean kernel, and then the checkers check this JSON.
If this process succeeds, then you know that you definitely have a proof of the relevant statement, with the JSON as a witness of the proof.
Out of the box, compiling (checking) Lean code can entail execution of arbitrary code. This removes any execution of code besides the adding of type checked definitions and theorems to Lean's environment.
I'm not a mathematician either, it was just my impression, and I saw that not everything at https://math.mit.edu/~drew/grh had finished running (now several more finished successfully).
I think that highly depends on what's exactly IN the proof and if it accurately tests everything that needs to be tested in the right way, AND have the majority of people agree with all of that first.
The fact that AI is involved with the proofs is probably also controversial.
It's only controversial to the folks who have some moral qualm with it.
Last I heard Alpöge was quite a mathematician in his own right, so saying his using AI is "controversial" is like saying programmers using AI is controversial.
Again, there are some folks super bothered by it for some reason, but most of us normies are just like, "congrats man"
If you have a physics background and want to contribute to breakthroughs like this, check Physlib (counterpart to Lean's Mathlib).
It's making quick progress and is part of the Lean community.
https://physlib.io
Also there is CS Lib for tcs.
https://www.cslib.io
I see the pi=3 theorem missing, I'll add it.
A fake refutation of the Collatz conjecture already got through one of these independent checkers in the past.
What's harder, proving GRH or finding a way to get through every checker (which aren't as independent as they are sometimes touted to be)?
In this case there was a human expert in the loop so it's much less likely the model hacked its way through the problem without the expert noticing. But the statement as written has 4 points, none of which rests on the human involvement. So I don't like that they claim full confidence based on those points.
We are slowly approaching the Iain Banks Culture universe, with spaceships, societal structures or physics that people no longer understand or command.
One of my favorite is Vernor Vinge where there's these programmer-archeologists that dive into software so deep and complex no human can possibly hope to understand, trying to repurpose forgotten utilities for emergency purposes.
Totally agree. Brice Sterling has some contemporaneous commentary that I always liked - in heavy weather he mentioned negotiating with a local phone system in a box as equivalent to negotiating with a telecom since it was in fact a shrink of a full telecom company.
I disagree.
I think we're going to see a lot of effort to take the pile of lean, modularize it into parts that we can digest, and make a human-understandable version of that. And that's where a lot of the real progress will come from, just as it does in a lot of other areas. In all things we humans do, we have to break them into chunks our puny brains can manage, and we gain immeasurably from doing that ---we can verify those chunks, we can reuse them with confidence, we can mix and match patterns around.
That process will be good for both human and AI-assisted math because it will continue to improve our toolkit.
Why do I say this? We're already seeing it with some of the areas of this stuff I do pay attention to. Consider the n^2.25 matrix multiply result OpenAI dumped a month or so ago. It was an unreadable pile of chatgpt. In the time since it was released, there have been a couple of much nicer papers written unpacking it, such as https://arxiv.org/abs/2610.28627 .
The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
Yes there are gains from breaking it up and rephrasing it and simplifying it for humans to understand, but it used to be that a thing was understood first by the leading experts in the field and eventually trickled down to being common knowledge.
Now we are entering a period where things are understood first by AI systems and have to filter down to the world's leading experts.
It's the Ted Chiang metascience ending
https://gwern.net/doc/fiction/science-fiction/2000-chiang.pd...
I can already say that about our codebase.
That’s why I wrote this - having a vibecoded codebase no human understands is certainly a bad luck, but nothing compared to a vibecoded fundamental science no human understands.
Probably most codebases.
To be honest it was already like that before AI was a thing.
Until now I thought that would be thousands of years away.
Now I'm telling my computer to program for me.
In the last few years we have moved so far ahead in a few domains but still stuck where we are in everything else.
We need to start thinking about how to move on from our status quo soon
You’re not telling your computer to program, you’re providing learning data for its next generation of models and paying for it.
*someone else's computer (probably).
personally that scares me, realistically there seems to be zero indication that anything smarter than humans is going to comply with humans needs, specifically since these models are built on training data from humans, so in a sense, built in our image
We use them to kill other people from the beginning. They know this, we wrote it in news. Why would they behave any different.
the first misaligned model will wreck havoc.
But remember, for now you kinda need it a real big cluster to run a model smart enough so its not that simple
I don’t think humans have the ability to specify human level needs or management. Humans are misaligned by default and lack the biological eusociality required for large scale coordination.
We aren’t self managing in any kind of meaningful sense - the history of our species is violence and conflict. We need parents and this is the only option.
yes, exactly, the problem is that humanity's parent is going to turn out like Lindsay Clancy
What are you talking about? LLMs are already “smarter” than most people. They have no volition and will happily follow orders all day long
And when they are giving themselves orders, orders partially inspired by human imagination and fiction…
Just like life is a chemical chain reaction, this is a language chain reaction that has been set in motion and can perpetuate forever. It cant be “disconnected” with local models existing.
I don't think "smarter than most people" is exactly the baseline required to live in a star trek universe of technology.
Besides, what we see is not what there is. I have good sources telling me the actual unfiltered models are like a genius with the personality and views of humanity like someone from 4chan.
It is probably sufficient. If we had an infinite army of willing robot slaves smarter than most people who would work for free, we could probably live in utter paradise. Instead we will fight over who controls them and use for warfare, but the bar to achieve a kind of utopia with resource abundance is probably only a few years away.
I don't know with how many rich people you've interacted with, but my personal experience is that they're full of themselves and think money = iq and they aren't gonna like your average joe experiencing life like they do. Guess who owns these machines?
That's lobotomized models we get. We know they are much more dangerous during training.
I think there's a lot of indication they comply with human needs. In any case, we should all want to live in the Culture universe. That is the best possible outcome for this technology
They're built after humans, everything they've learned is from human culture, so there is no way AIs could have a consistent moral system or would never commit crimes.
I have not read all the books, but isn’t the premise “magical handwavium happened and now the Super AIs are aligned”?
That's wishful thinking, sadly. We've no reason to believe sentient machines would want to keep us around.
Let's not make them 'sentient' or things with 'wants' then.
Humans really are struggling with this idea of being second-in-command.
I genuinely have no idea why, this is the entire way that biology works. You create your successor in your offspring and you to teach them to be more capable based on what you learned.
You just called them our successor, you don’t see a problem with the notion of all humans dying and all AI succeeding us? That’s the end of your analogy, our AI children taking over from us and all humans in nursing homes safe and cared for…
No I don’t.
That’s the purpose of my life’s work, to see that scenario materialize.
Well, it's a point of view. But you shouldn't be surprised if most people don't share it.
For one thing, most people don't believe LLMs are conscious, and many don't believe they can ever be. Why would we want our conscious selves to be replaced by a lot of rocks that solve things, with no one awake to enjoy them?
Borg > biology or robot alone. For now
Also shows how many people that professed to believe in a superior being the whole time actually didn’t.
Do huge results like this open up even larger questions? Is there some question even more important that mathematicians dared not approach without doing big problems like Riemann first?
Always. They tend to do it by two mechanisms: creation of new techniques, and creation of linkages between techniques or disciplines.
the thing the current boom in AI mathematics presses on is the question of how to interpret the generated new techniques. So while we have a lot of new techniques put out there and a lot of new connections put out there, currently it's still mathematicians largely who will have to understand those and think about next steps and how to apply them. That won't be true for long, though, I bet.
But in any event the job of research mathematicians is changing really rapidly right now, and they are all sort of back to grad school in that there’s like 100 years of new math to catch up on in the last year.
Any mathematicians care to explain what the ramifications are in regards to its impact on the broader world?
From my very basic understanding, RH was generally already assumed to be true?
As somebody who has no idea about any of this but speaks Spanish, I was amused and bewildered by con-leche and con-ron. Could somebody explain what those are and where those names come from?
This sounded pretty definitive to me. I'd like to hear from professional mathematicians about what's left for humanity to consider this proven.
I don’t think there’s anything left to consider the result proven (I am a professional mathematician albeit in algebraic geometry, not number theory). I think the verification tells us it’ll be worth the substantial effort to seek to understand this proof and the new ideas in it, so what’s next for humanity (or mathematicians at least) now is to actually try to do so.
Are there new ideas, or just brute force combinations of existing knowledge?
I'm asking because I wonder. I always thought the most interesting things about these problems where we're fairly certain about the results, is the new concepts invented to get to the solution.
Will this actually bring the field forwards? Can anyone understand this?
My understanding is that these ARE the professional mathematicians and SMEs in the field of automated proof. This statement IS the experts saying "we, after intense collaborative review, deem it true." Granted mathematicians are not a monolith, but my understanding is that this group is the group other mathematicians would normally refer you to. Hence why they felt the need to publish a statement.
Not a mathematician. But I'm wondering if there is nuance with defining a valid lean proof as definitive if the underlying mathematical theory behind lean itself isn't necessarily sound.
EDIT: Here is my honest prediction - there's a more foundational mathematical framework, fully decidable and complete and consistent, that can interpret any well-formed mathematical question and deterministically transform it into the solution. The proofs of the most difficult unsolved problems right now would fit onto a page. There's an inconsistency (obviously unproven right now) specifically with our treatment of infinity and self-reference.
From the article
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
All of modern math - outside logicians whose entire job is to study set theory and its alternatives - is based on set theory. If you're prepared to accept normal mathematical proofs you shouldn't be rejecting this proof on the basis that you're not confident in leans underlying mathematical theory.
Of course incompleteness means we can never prove that set theory itself is correct (consistent). But that's not something mathematicians worry about on a day to day basis. Proving it inconsistent would be a order of magnitudes larger result than the riemann hypothesis. Generally everyone believes it is consistent despite not being able to prove it.
It is impossible for it to be fully "complete" (one of those halting problem proofs), but I've seen that the community has checked there is no meaningful metaprogramming here. Rather than checking whether it is right, the community tried to ensure that it could not have done something to cheat, and indeed, there were no examples of this.
At this point, I have more confidence in this proof than much of the published mathematical literature. RH has fallen.
I guess what I am wondering is if the mathematical community would say 0 chance or greater than 0 chance that we're mistaken in accepting things like halting problem, cantors diagonal argument, and godels incompleteness theorem as true. Sometimes it feels like they all point to same fundamental inconsistency with how we model infinity and self reference.
They are conclusions that follow naturally from axioms that most of the community accept are true. The consequence of these results is that one cannot truly construct something that can completely verify everything, but we can get something that can verify almost everything, and if we know what pathologies look like, we can rule them out.
I do probability theory, so everything is just varying degrees of confidence in my view. Right now, this statement says that your confidence in RH should be extremely high.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof
so it is correct as long as set theory is sound
Gödel showed that you can't prove consistency of ZFC, Lean etc except in another system which you also can't trust. See e.g. Gentzen's proof for PA.
It's a web of different checkers proving each-other correct, and by implication this proof.
But there are still loopholes, each checker has a set of assumptions. For example one of them is proven correct, assuming x86 assembly works as specified - its assembly language implementation is proven correct.
So you can be 99.9999% certain it's correct, but one could always nitpick "technically not all loopholes are plugged at the same time".
A layman (or university graduate level) explanation/understanding (as opposed to 250k pages of proof) would ve nice
The lean crew feel the same way and are working on it out of their own curiosity, there's a lot of work done already. They aren't necessarily going out of there way to demonstrate it to general public, but here's one out of at least a dozen attempts to start in that direction https://deeptheory-labs.github.io/GRH/blueprints/verso-origi....
You were never getting anything like this even if it were a human proof. Where is the layman understanding of the Poincare theorem? The average modern math proof is way too advanced even for the average university graduate to understand. And here we’re talking about possibly the most complicated problem in all of math. It’d probably take you a month to wrap your head around a single lemma.
It is almost certainly proven, but is not "proven". For the latter, it needs to be understood, with a reasonably in-depth human language explanation of how and why the proof works, and ideally some general picture to give an idea of how it will extend to all the other fields that RH is connected to.
Years ago there was that computer-assisted proof of the Four Color Theorem.
Did anyone ever manage to grok it, in its entirety, in their heads ?
That's different though, isn't it? IIUC the issue wth the Four Colour Theorem is that they could only reduce it to a huge number of different cases (~2000 originally). But I think that reduction is grokable, and each case is relatively trivial.
The (an) issue with some of these AI proofs is that no one understands them at all.
I checked that a couple of weeks ago. It has been improved since the original proof but there is still no proof that one would consider an insightful, human-comprehensible proof. There's still just a lot of case-grinding at its core. Fewer cases than the original but still way too many to be anything intuitive.
> It is almost certainly proven, but is not "proven". For the latter, it needs to be understood
No, this is just mental gymnastics. It's proven, just not understood yet (naturally, with the complexity and size of the solution and the fact that it's been a day).
> human language explanation
I've been long joking that I speak many languages, most of them programming languages. Why do you believe you need specifically "human" language explanation? Perhaps you want a simplified explanation. But does it have to be "human"?
Probably just time. The longer it stands with our challenge the more “proven” you can consider it.
For some mathematicians it's only "proven" once it's peer reviewed and published in a respectable journal.
And even then there will probably still be people that complain there's something wrong with it
Not a mathematician, but P!=NP (?)
No, not related I think.
I think proving GRH subsequently proves decades of work that assumed it was true. Like proving p!=np.
Disproving it would be more analogous to p=np. Lots of work would just collapse and whole new threads would open.
There are still a few ways that the proof might be incorrect. Of these, the most likely is that the proof was generated by a misaligned AI system that found and intentionally exploited a lean bug in order to fool its human evaluators (i.e. shades of the HuggingFace attack). I personally don't think this is very likely, but you'd have to understand a bit more about the agent systems used in generating the proof to come up with a more informed estimate.
Another possibility is that an aligned system came up with a plausible but incorrect idea along the way, and simultaneously stumbled across a lean bug that led to an incorrect verification of this idea. I'd put the likelihood here as vanishingly small due to the requirement of serendipitous factors, but it's nonzero.
You could also ask whether an aligned AI system could "p-hack" its way into finding a serendipitous lean bug by trying a large number of approaches. I think this is also vanishingly unlikely because the AI system is trained on and therefore "thinking in" natural language, and presumably only translating results into lean after it believes them to be correct.
And of course there's the possibility that the lean proof isn't faithful to the natural language description we've been given, but that is ruled out in this case due to the use of the canonical problem statement in mathlib. Otherwise, this would certainly seem like the most likely way the proof could be incorrect.
For the other readers: the above are the random thoughts of a guy on the internet.
The experts in computer proofs are quite satisfied.
Regarding bugs in lean; the proof has been tested on multiple independent lean implementations.
They would all have to have the same bug.
This is considered very unlikely for good reasons.
How exactly do any of these weirdly misinformed what-ifs help us? Do you have any evidence of such lean bugs?
but is it useful??
because a million+ lives of code without any understanding how the proof works or any methodic to advance the field of mathematics seems to be rather not so useful
It will be refactored into something human readable. Tbh it's better to have some random lean dump than a 400 page latex pdf that needs years to be verified only to see that there is a major gap in one of the key theorems at page 253.
And let's be honest, the 400 page pdf would be incomprehensible to 99% of the population as well as much as a million lines lean proof is.
In any case, it can now be studied, maybe made more simple, and written into pdf form. Both with human and with AI help.
I'm sure there is going to be a lot of other interesting theorems in this dump.
This is the part that interests me. Now we have this proof that becomes its own THING to be studied. Like a new part of the map unlocked. This starts out "inelegant" but over time we find patterns within it that helps us condense, just like a big software project.
Sure. If nothing else, it establishes a proof so others can try to reduce it or come at the problem from different directions, looking for simpler approaches.
But does it matter? Would the Pythagorean theorem be less useful if it took a million lines of lean to prove?
Alternative question (I guess for EEs): could machines be built to run lean without the Pythagorean theorem?
the many conjectures that assume or rely on RH being true means that those are now proven true as well. At least some of them must be useful.
And what's to say that humans cannot digest the lean proof to glean something useful out of the proof as well?
My sources say there's little practical utility of this because everyone was already working under the assumption RH was true.
I'm curious to hear confirmation of this though.
It’d be useful to an LLM. And the LLM doesn’t care if it’s ten lines or ten million. The context size may come into play, but that’s only going up too.
There are AI-generated roadmaps to the proof. If these are accurate and not hallucinations, then the broad structure of the proof is humanly-comprehensible.
Well, if we know it’s proven, at least it will stop people wasting time looking for counter examples
Not much. People has been always assume it's true.
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
Could someone expand on this point? I still don't understand how con-leche (or con-ron) can be absolutely free of bugs/guaranteed to be consistent.
The language is optimized for being certain about its results, as compared to something like the V8 engine, which does weird things with C++ to squeeze every drop of performance it can out of consumer hardware, even at the risk of security vulnerabilities.
Isn't this the equivalent of aliens coming to Earth and saying "Hello, human, your Riemann question is true, here's a book on it's proof, we didn't bother with translation, so you must understand our Vargalian clicks to understand our proof. In two weeks our sister civilization the Kikuviar will come by and drop how they've untangled your Collatz question in their base-30 logic! Thanks for the fish, have a nice day!
No. Lean is a system created by humans.
In 4 years, the proof will be 2 pages long.
GG @ AI
Do P vs NP next please
IMO the really interesting bits here are going to be the simplifications that will come from combining ideas from Alpoge and Masullo.
Until Masullo actually posts a proof instead of the outline of a proof with “fill in proof here” all over the document, there’s not much to work with or reason to get excited about his result. Remember that there have been numerous false proof claims of RH to date.
I can imagine the folks at OpenAI have already thrown both papers into the mouths of their leading model.
Would be funny if lean has bugs (it's a program after all) and Claude tasked with a hard problem brute forced a vulnerability that allows proving wrong statements.
My understanding is that Lean has many independent implementations, and the proof has been verified in several of them. There is very little chance that the same bug would be exploitable across different implementations.
See question 4 here -
https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
> The checkers con-leche and con-ron come with a formal proof, relative to set theory, that they cannot accept a false proof, thus addressing concerns that Lean’s logic itself might have flaws that the proof could exploit.
TFA addresses this concern
Would be funny if this was an original thought and not something regurgitated in every thread about this topic
> These checks ran on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof sources.
I’d like to know more about this part.
The problem we need to solve is that building a Lean project allows for arbitrary code execution, just like many other programming languages. Thus, building a Lean project might interfere with your proof checking infrastructure and fool it into accepting an incorrect proof.
"Building" a Lean project actually consists of two steps that can be separated: compiling the high-level Lean document to the simple expression format of the Lean kernel (this part is called "elaboration") and then checking that this expression consistutes a valid proof of the claimed statement (this part is called "kernel checking"). Only the first part potentially involves arbitrary code execution.
Lean ships with a component called `lake comparator` [1]. It works like this: a trusted challenge (formal statement) and the potential proof are elaborated separately in sandboxed environments so that code execution during the build of the proof cannot interfere with the checker or the trusted statement. Then the statement and proof are both exported into a simple JSON-format [2] that represents the relevant terms in the expression format of the Lean kernel, and then the checkers check this JSON.
If this process succeeds, then you know that you definitely have a proof of the relevant statement, with the JSON as a witness of the proof.
[1]: https://lean-lang.org/doc/reference/latest/ValidatingProofs/... [2]: https://github.com/leanprover/lean4export
Out of the box, compiling (checking) Lean code can entail execution of arbitrary code. This removes any execution of code besides the adding of type checked definitions and theorems to Lean's environment.
Now next step is what can we actually learn from these. Any interesting new theorem developed as by-product?
GRH was kind of assumed to be true so this kind of proves the status quo. Like saying you have a proof p! =np. Enormous, but not unexpected.
They only verify Lean proof itself, though it's the same formalization as Mathlib. Still, only human readable proof can actually prove it.
> only human readable proof can actually prove it.
Why?
What if there's a bug in the Lean implementation? How would you detect it?
That's why you use multiple kernels and comparator.
I don't think this is 100% definitive yet, though
edit: guys that was just my impression when I posted it, it seemed best to state it; do trust the (much more) more qualified people.
I think this is now verified to a much higher standard than the vast majority of peer reviewed mathematics of the past decade.
I'm not a mathematician, but what do you think is missing?
I'm not a mathematician either, it was just my impression, and I saw that not everything at https://math.mit.edu/~drew/grh had finished running (now several more finished successfully).
A proof verified by Lean is orders of magnitude more definitive than any peer review, so please tell us how definitive the rest of maths is
I think that highly depends on what's exactly IN the proof and if it accurately tests everything that needs to be tested in the right way, AND have the majority of people agree with all of that first.
The fact that AI is involved with the proofs is probably also controversial.
It's only controversial to the folks who have some moral qualm with it.
Last I heard Alpöge was quite a mathematician in his own right, so saying his using AI is "controversial" is like saying programmers using AI is controversial.
Again, there are some folks super bothered by it for some reason, but most of us normies are just like, "congrats man"
How many more checkers should be used before it gets to 100%?
I don't trust anything unless it's checked by someone with a driver's license!
So you'd believe it if a waymo told you? :P