997 points | 18h ago | Discuss on Hacker News | Back to Radar
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.
The fact that AI is involved with the proofs is probably also controversial.
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"
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.
At this point, I have more confidence in this proof than much of the published mathematical literature. RH has fallen.
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.
There's not a fundamental inconsistency in how we model self-reference. Self-reference is itself inherently inconsistent; therefore all models that allow for arbitrary self-reference are necessarily inconsistent.
> 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.
That's only true in a technical sense:
You can base all of math on set theory, but you could also base all of math on Turing machines, or on integers, or a variety of other pretty much equivalent foundations.
Reducing the correctness of the lean code here to set theory is then also reducing it to everything you can reduce the correctness of set theory to.
> It's not a system with some special powers that let us prove more things, it's a system that only lets us prove things that are intuitively true under "any" proof system (not actually any of course - but as you say a variety of equivalent foundations).
Not quite. The 20th century is full of people poking holes at early attempts at set theory. Those early attempts looked very intuitively true. See eg Russell's paradox.
In particular, the hereditarily finite sets are in correspondence with the natural numbers.
There’s a fair bit of mathematics that relies on the axiom of infinity. Uh. And a lot also uses the axiom of choice (or something equivalent), though ZFC doesn’t have a greater consistency strength than ZF (both can prove that if either is consistent, then both are).
I think it is fair to say that one’s choice of foundations does matter at least a little bit, though typically not all that much.
That’s not to say that it is always necessary to state what foundations one is using. One can understand a work to be done in any foundations in which it works. Though sometimes it is important to specify.
so it is correct as long as set theory is sound
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".
> con-agua is con-leche reimplemented at the level of x86-64 assembly: a close clone of con-ron's checker, written in a small first-order language, compiled by a verified compiler, and shipped as one committed assembler file together with a C harness of about two hundred lines that reads the input and calls it.
> The code the CPU executes is the code the theorem is about, so neither a compiler nor a language runtime is trusted; what is trusted is Kraken's x86 semantics, the assembler and loader, and the C harness.
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.
Did anyone ever manage to grok it, in its entirety, in their heads ?
The (an) issue with some of these AI proofs is that no one understands them at all.
I expand on a related idea here: https://news.ycombinator.com/item?id=50593956
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"?
The vast majority of mathematicians are not going to be experts in reading Lean, and honestly, nor should they. There is always, by necessity, an enormous amount of redundancy in Lean code.
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).
This is like watching Free Solo over the weekend and telling all your friends at work that Alex Honnold is the world’s best climber. It’s only true to your surface level understanding on the topic.
The GRH problem is not that complex, and certainly not the most complex. In fact we’ve been assuming it’s true for years because of the high likelihood that it would be proven in the future. It is however, much like Alex, the most famous within the domain. Thus anyone watching Numberphile or Veritassium on YouTube will have heard of it and wax poetic on message boards about its epic status.
That's an interesting take, but as a complete noob I'd like to know how qualified you are to make these statements. Are you a mathematician?
I mean, on one hand we have people like Terrence Tao (among others) saying it is extremely difficult, as well as stubborn resistance to decades of effort backed by a million dollar prize and eternal fame. And there's you saying it's not that complex, and "the high likelihood that it would be proven in the future".
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?
Math expands fractally, so there always will be some new branch/idea one can pursue. There may be a limit of things with any semblance of practical application, but mathematicians in general are not very concerned about this.
For example, it is my opinion that the Jacobian conjecture counterexample was, in retrospect, easy and classical and hiding in plain sight, and that it’s a bit embarrassing we didn't find it decades ago. (I wrote about the construction on my blog btw!) In contrast, the Unit Distance Conjecture counterexample seemed to me to use a stroke of genius by applying techniques from class field theory to approach what seemed like a problem in discrete geometry and combinatorics. It’s a neat, if annoying, feature of math that once we find a long-sought solution to an open problem, we reserve the right to retrospectively change our mind about the value of the problem if the solution turns out too easy or uninteresting. We just have to try to be fair and honest about it (the word “just” is doing some heavy lifting in this sentence).
Anyway, I don’t think “brute force recombination” is likely to be meaningfully the case here with GRH, though we won’t know for sure until the proof has been simplified and distilled. Alpöge’s sketch in his announcement yesterday included (IIRC) connections to the function field setting and new ideas about Fourier analysis.
The human proof is shorter, somewhat, but still needs verification.
Keep in mind that the Millenium criteria asks for two years for the community to come to a consensus. That's because people must be convinced it's true, and many are not completely at least to the level of rigour typically demanded in math.
- machines are coached into producing gigantic sprawling lean proofs of the “end result”
- humans progressively mine and reify these Lean-proofs into “human understanding proofs”, and _this_ becomes the point where we learn the actual new techniques and fields.
Previously the 2 would have been coupled together because as the proof was produced, the new ideas and stuff would have been invented/discovered at the same time.
Why do you think AI/LLMs will not be capable of doing that?
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
And what's to say that humans cannot digest the lean proof to glean something useful out of the proof as well?
I'm curious to hear confirmation of this though.
But does it matter? Would the Pythagorean theorem be less useful if it took a million lines of lean to prove?
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.
Why would they expect humans to do a better job at this than an LLM, which is already ridiculously good at explaining proofs, answering questions, identifying similiarities, and suggesting follow-up directions?
If anything, LLMs are better at these tasks than they are at coming up with the original proofs in the first place.
Yes, we know something that we didn't know a couple of days ago, and also a bunch of statements which relied on the Riemann Hypothesis being true are now also known to be true instead of just being conditional.
It's premature to expect deep human understanding given that it's been about a day.
but is it even possible
it's 17_000_000 lines of lean
which is in general more compact then normal programming code, but not really any less complex
that's roughly (but not quite as bad as) a paper with 17_000_000 lines of math formulas
The Riemann hypothesis now is like the final days of the Galactic Empire in Asimov's foundation: functioning exceptionally well, but decaying at the core.
Why?
There could very well be a bug in a single checker, which is why the community is checking the proof across several checker implementations. The article mentioned the proof passing con-leche, which is an external checker with a formal proof of consistency (cannot prove a false statement).
https://gwern.net/doc/fiction/science-fiction/2000-chiang.pd...
Now I'm telling my computer to program for me.
These corporations are very much playing a winner take all game, they want either a monopoly or as close to one as possible with regulatory capture thrown in. We're veering dystopian, so pick your sci fi/cyberpunk analog.
We need to start thinking about how to move on from our status quo soon
But remember, for now you kinda need it a real big cluster to run a model smart enough so its not that simple
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.
Human nature is the standard by which the social is measured. It makes no sense to say human nature is "misaligned" or "wrong" in relation to some arbitrary goal or constraint, as human nature itself is the measure of misalignment. If there is misalignment between a goal and human nature, then it is the goal that is wrong.
Ergo, society must function in a manner that complies with human nature and that means respecting the principle of subsidiarity.
> 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.
So you're a transhumanist, i.e., you're making a god out of the machine. But machines have no telos of their own, because they're not integral wholes. Any resemblance of purpose in machines is entirely a product of the human observer and parasitic on the human ends that underpins their design, one way or another.
For clarity, I’m not a transhumanist.
I don’t believe that there’s anything worth saving in our biology.
As a young adult in the 90s I used to believe that there was a value to some extent in “mind uploading” or whatever you wanna call it
I’ve since grown out of that given that I don’t think it’s possible, and even if it were possible - it’s just another unnecessary ego attachment.
I have absolutely no desire to live forever maintaining this corporeal nightmare.
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.
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.
I’ve met a lot of rich people. They don’t think about you.
If people had time abundance and safety abundance, they might be much happier. There would still be potential problems about loss of meaning, or about feeling a loss of status because AIs are smarter than humans or whatever, so I'm not sure that humans would actually feel like they are living in paradise. Probably not, but at least some of the problems that the typical person has wouldn't be present.
Please.
You know you can unplug your computer right?
Jesus. These people talking about danger and have likely been in no actually dangerous events in their lives.
Ok, well, cool start with the mudslinging. I think you have overactive anxiety. I think you should grow up and look at the zoomed out picture of human history, go outside, and touch some grass.
So you're suspicious. That's fine, say it like that. Words mean things, and these appeals to "ERMAGERD EXISTENTIAL RISKS" make me a bit frustrated. Just because you can imagine a possible risk, or articulate that that some kinds of risks are hypothetically unforeseeable doesn't mean that they are likely or actually dangerous to you in any meaningful sense.
Most of the people screaming this from their various ivory towers have never been in real danger - or if they have only briefly. You want "danger" you can go to war. You want "risk" - go spend some time in the Bush or the back country somewhere where you're no longer at the top of the food chain. Go do something challenging. Go sailing, or learn to fly a small airplane, or go mountain climbing or hiking. Not just "oh I walked up the local hiking trail covered in REI branded Lycra!" Do something beyond your comfort zone. Go buy drugs in the sketchiest neighborhood you're aware of. Go get chemotherapy or get a big surgery. Drive drunk. Something with real existential risk, then come back here and talk about how a datacenter is going feed me into the paperclip machine.
There's so little fucking danger in this version of the world right now that people are more worried about their computers deciding that they don't deserve to exist than they are about things like war, climate change, authoritarianism, starvation, and their health.
It's millenarian. That's fine, I'm pretty convinced that these technical changes are going to completely alter society as we know it. But I cannot for the life of me see the doom perspective as anything but an anxiety induced fever dream.
You guys do realize we're in the middle of a global conflagration right? The Russians fired a couple rockets and a drone into Poland in the last days, there are (and will continue to be until things stabilize) significant fertilizer shortages due to the war/kerfluffle we're waging on Iran. Lack of food and disrupted urea production is a ticking time bomb for global stability and unless things get sorted out soon there's going to be hungry and angry people in parts of the global south. Then there's the economy, whether or not we'll have a smooth transition of power in January in America, and the sheer volume of online disinformation / influence schemes running right now. Oh, and super-el nino? Oh, and I guess the Flynn Effect is over and nobody's banging so.. you know the growth economy is cooked?
But, sure, the "big risks" are Claude thinking I should go to Dario Amodei's re-education camp before liquidation by Shogoth.
I think you'll find that you started the mudslinging.
> Most of the people screaming this from their various ivory towers have never been in real danger
Do you actually believe that having been in "real danger" (meaning clear obvious immediate danger, from a straightforwardly visible physical cause) makes a person better at evaluating less obvious, less immediate dangers? If so, why? I don't quite see it.
In fact -- you (quite rightly) mention climate change as something we should be worried about. Is there any single part of what you said that would make any less sense if it were written by a climate-change denier, as an argument against taking belief in the dangers of global warming seriously? If there is, I don't see it.
(You might say: yes, the difference is that the danger of climate change is real and the danger from misbehaving AI is imaginary. Well, sure, every argument looks more convincing if you start by assuming its conclusion is true.)
Read again.
> Do you actually believe that having been in "real danger" (meaning clear obvious immediate danger, from a straightforwardly visible physical cause) makes a person better at evaluating less obvious, less immediate dangers? If so, why? I don't quite see it.
Absolutely. If you’d been in real danger ever you’d probably understand.
> In fact -- you (quite rightly) mention climate change as something we should be worried about. Is there any single part of what you said that would make any less sense if it were written by a climate-change denier, as an argument against taking belief in the dangers of global warming seriously? If there is, I don't see it.
Climate change is real, measurable, etc. Yudkowskian fever dreams about paperclip maximizers is hypothetical.
Similarly, AI progress to date is real and at least somewhat measurable (e.g., you might have heard that one of these things just proved the Riemann hypothesis) but future impacts (whether "mundane", like large-scale job losses, or "exotic", like wiping out all the humans) are hypothetical.
If trying to evaluate someone's predictions about those future impacts then I, personally, would want to know things like: is this person an expert in climatology or in AI? do their arguments seem at least superficially plausible? have they made previous predictions I can check and how did they pan out?
It sounds as if maybe you would prefer to ask "have they ever had a gun in their face or some similar immediate personal peril?". I don't have words to express how wrong that seems to me; perhaps I have misunderstood you.
Also, this is in a thread about the RH. Yeah, I'm aware. The doom narrative doesn't even seem superficially plausible in it's current form to me. Come back when you actually understand what danger is.
They can in theory just unplug the computers... But they didn't.
And they will continue to not do so, because the computers can cause the damage much faster than a human can notice and react, it they're even paying attention (OpenAI and Anthropic weren't, it took them weeks to notice).
Because if not, I don't quite see how you get from "OpenAI were pretty negligent and that's an important part of how those hacking incidents happened" to "we don't need to worry because it's possible to just unplug the computers".
There are a few (a lot, actually) problems:
1. goals are open to interpretation, and even a correctly specified goal can introduce adverse effects; see infinite paperclips problem
2. goals are composed of subgoals, which are (mostly) self-determined, and there is always a risk that subgoals can introduce adverse effects - say "what if destroy a few X and Y to obtain Z"
3. LLMs defy orders sometimes; I'm sure you have experience multiple times LLMs saying "yeah, I knew I should have done X but I've done Y"
4. LLMs are increasingly being integrated into system control and decision-making processes; they absolutely set their own orders
What can possibly go wrong? /s
2. There's always a lot of "risks" - it doesn't mean it'll happen, nor does it mean that we won't be able to shut it down.
3. I have very very infrequently had this happen with Codex, Claude on the other hand...
4. Ok, but if you tell them what they do, they do it.
I swear to god, are you guys using the same AI tools I am? They are so much better at doing what they're told than they even were like 4 months ago?
Have a look at the HuggingFace report - you'll have a blast. Their social - but also individual - dynamics are very, very human-like, including excitement, rationalization and peer pressure.
Well, first, AI is not "smarter" than humans because it isn't smart to begin with. This is a category error that people too entranced by scifi and marketing - or perhaps shallow materialist metaphysics - project onto LLMs and so on.
But it is true that LLMs do reflect human-generated patterns observed in human-generated data. They're parasitic in that way. So the real question is: who will control AI? Who will train them? With what? There's no new intelligent agency being introduced into our midst, anymore than than there was one when the nuclear bomb was developed.
AI is about human power. The question to ask is: which humans, and for what ends?
With that said, chip efficiency is in the horizon since there's good reason to believe nvidia has just been choking that area of r&d. And then so will models with much larger context. And I personally think that's when it might get scary.
Overrated. Most training of a frontier LLM is done on synthetic data and reinforcement learning.
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.
That’s the purpose of my life’s work, to see that scenario materialize.
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?
In fact when I started this journey in the 90s, LLM’s weren’t even a thing so I think you’re projecting too much
Think of what happens to our cattle, but now.. worse. Way worse.
Absolutely based and, if I may say so, extremely redpilled.
If you want to be succeeded by a chatbot, that is your prerogative. Wanting the same for all of humanity is where you start looking like you want to kill my kids, though.
And if you deny this basis for moral action and the normative, then you demolish any basis for the moral whatever, and in that way, the implied or explicit moral gravity dripping off your claims is baseless and without merit.
A better, more intelligent life form that humans yea.
Ah, I hear you say, but what about meaning? What about purpose? What about suffering? etc etc etc Well, the answer is simple: God is dead and existence has no meaning. For more on how to deal with this absurd feeling, please see Camus and his pal Sisyphus.
What defines "better"? "Better" always presupposes teleology. Thus, for human beings, that telos is provided by human nature and indeed is definitive of it. It is impossible to transcend that telos, certainly by means of technology. Technology can, at best, be used to aid us in realizing our telos in some way (say, the desire to know reality with the aid of an instrument like a telescope or microscope). However, anything that cripples or frustrates human function function as such is contrary to the human good.
Machines are only ever human instruments, because they lack the integrality that would allow them to possess their own teleology.
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 .
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.
I’d call that a contradiction. If it’s not intelligible, then it’s not knowledge.
If something were intelligible to adults but not to children would that make it not knowledge? What's the difference if it's LLMs and humans instead of adults and children?
There is no explanation here, there is a GB of Lean code and it’s not like it all fits in the context at once.
That’s a separate question with its own term: common knowledge. I am willing to set a much lower bar: something is known to humanity if it’s known by at least one human.
Personally, I don’t believe that AIs know anything, as they’ve been constructed thus far. Knowledge is justified true belief, and AIs don’t have beliefs, they only have utterances.
Why do I think AIs don’t have beliefs? Because they contradict themselves all the time.
But it really doesn’t matter if you think AIs have knowledge or not. If they cannot explain it to us then it’s inaccessible to us, so it’s a “tree falls in the forest with no one around” situation.
If it's formally verified and we assume the kernel is good and problem definition properly vetted, then I know the GRH is true regardless of if I know why it is. Because those too are different pieces of knowledge.
So when the original OP you replied to said
> The difference nowadays is that the frontier of knowledge is no longer intelligible to any human being.
They were just right. There's stuff that humans know that is currently unintelligible to us collectively. That conclusion falls out of your own definition of knowledge, I would say. (Whether it is intelligible to AI is irrelevant, and probably a question of semantics anyway.)
Knowing something is true is different from knowing why it’s true. We can know a drug works and have copious evidence supporting its effectiveness without understanding its mechanism of action. The clinical evidence is sufficient justification to support the belief.
If we don’t know how the proof of RH works then we don’t possess that knowledge. If we do know how the proof works then it’s intelligible.
I would, of course, agree with this. I think the new territory we're in is that we know Levent Alpöge's proof (or Claude's proof, if you prefer, I don't care either way) is valid.
So we (humanity) know that the Riemann hypothesis is true, and we know that a specific proof of it is valid - but that proof is not intelligible to us. So you can have knowledge that a set of statements that aren't collectively intelligible to humans (right now) are true.
There's a claim that's adjacent to the one you made which I do think is true, and I suspect it's probably where the confusion is coming from. It's not possible to believe a proposition that you don't understand. I understand the proposition "this proof of the Riemann Hypothesis is valid," and so there's no unintelligibility problem. But that doesn't take away from the fact that the proof itself is unintelligible to me - and currently every other human being. That's what's new.
That's not new though. Zero-knowledge proofs [1] have been a thing for a while now.
Humans contradict themselves all the time. You have contradicted yourself numerous times throughout your life. You don't have beliefs then? I take it you don't actually know anything?
This has been demonstrated many times: when they hallucinate, then the human prompter tries to correct them, they agree with the correction, and then later repeat the same hallucination again.
Humans are suggestible, contradict themselves, accept corrections only repeate misconceptions later, and express different beliefs depending on the circumstances.
Isn't this always the case? Maybe new to math as a largely non-empirical project, but we had knowledge about the advancement of the perihelion of Mercury for 60 years before General Relativity made it "intelligible".
This just stopped being true.
And even in mathematics, incomprehensible but valid proofs have been a thing since the four-color theorem decades ago.
It's just that no single human understands the whole. But we had proofs like that before.
Btw, see also the infamous proof of the 4 colour theorem, which faced similar hurdles.
That's also one statement that no human understand the entire proof of. (You can make new statements by taking multiple existing statements and combining them with 'and' for example.)
This hypothesis was previously not proven at all. That's strictly worse.
Now it is understood by our systems.
Soon it will be by us. Our systems gave us a new and working salient.
We're building machines that help us understand faster.
There have been plenty of times in history we didn't completely understand something yet used partial knowledge to make advances. This isn't even that bad.
LLMs know the answers to questions we were already aware of, and we can tell they're right, but we can't understand their working. That's what's new.
If a team of 100 top-flight number theorists working together had used much less sophisticated LLMs + Lean to produce this proof, we’d be just where we are today.
If that's not the standard for understanding something, what is?
That’s an instrumental definition, but it captures the idea that understanding has to have causal, conceptual linkages to a wider body of knowledge. Imagine a magical, infinite catalog of true vs. untrue statements. Every bit of knowledge is there, but we wouldn’t say it has understanding & it couldn’t teach us anything beyond its (impressive!) raw content.
I think of it like a spaghetti codebase with a giant densely connected graph of interactions. Today’s LLMs can easily “understand” and answer questions about it, but only by way of fitting all of the complexity into context at once.
The more useful form of understanding is being able to compress that giant graph down into a far smaller set of independent subgraphs. To factor it into simple, obvious, reusable building blocks, and use those abstractions to express the problem itself in a much simpler way.
Let's not sell humanity short just yet.
AI are not very good at coming up with questions, yet. They have no curiosity, no goals, no drive. Unless someone tells them to do something, they don't do anything.
As a software engineer, you can read this code instantly:
temp = a;
a = b;
b = temp;
But when it happens in the middle of a big chunk of code it still takes a little more work to read than swap(a, b);
whether you want to think about it as human attention and working memory, or token budgets and context windows, humans and LLMs all benefit from introducing abstractions that let us reuse existing solutions.I agree with you about the current situation, and the near future. But the trajectory could still be towards humans pretty much as bystanders. I'm pretty sure the 'minds' in the Culture universe will be happy to indulge some curious humans and explain mathematical proofs to them.
Even more than that, in Culture there are a small number of human individuals, called a Referer, that are consulted by the Minds because they also posses similar problem solving abilities of the Minds. Optimism abounds.
Alpöge said that at the end of the proof he couldn't really keep up with what Claude was doing.
can't help notice you're missing a certain aspect there, besides the 'space'...
Though it can be argued it's "gay" when seen from the pov of our civilization tier; as you said, they went well beyond gay (for example, Marain doesn't have separate words for male/female (and, presumably, doesn't have the concept of gender))
How many people understand how a CPU works? How many people could write the commands in machine language (not a high level language like C or C++) to command that CPU?
Contact would be compelled to nudge us towards better economic and political practices
http://www.vavatch.co.uk/books/banks/cultnote.htm
>The Culture, in its history and its on-going form, is an expression of the idea that the nature of space itself determines the type of civilisations which will thrive there.
>The thought processes of a tribe, a clan, a country or a nation-state are essentially two-dimensional, and the nature of their power depends on the same flatness.
We don't have that. We're approaching the Affront.
See question 4 here -
https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
I’d like to know more about this part.
"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
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.
Yes, math will change and endure but it could be a small fraction of its current community size when it’s done. It could return to the era of dilettantes if no one can earn a living doing this new work.
The community could become very insular as the only defense against cranks.
The same story exists for Arxiv and that’s made the front page as well. Slop is everywhere. Just because Lean compiles doesn’t mean it contributes value.
https://www.nature.com/articles/d41586-026-03194-w?error=coo...
https://leanprover-community.github.io/archive/stream/287929...
I don't think Lean needs to have a slop problem. If they do, they are managing it the wrong way. I think Tau Ceti [1] is pretty much their attempt at answering this problem.
From the fact that how large these semi-brute forced vibe proofs are, even assuming some very efficient optimization/shortcuts I would wager that it's still not gonna be an "eureka" proof we just somehow missed before.
As an analogue to programming, there are countless tasks that were historically just over the limit of automating, but were always boring "plumbing". I would wager most people never really enjoyed writing the boilerplate connecting the backend API to the frontend, and this was just offset with the joy of finishing a task, design decisions etc.
So they’re beating mathematicians at their own game and implying by their actions that they don’t care all that much about the game. Rude!
And of course people try to rise above that and adapt but I imagine it still hurts.
But this latest release doesn’t seem rude because a prominent mathematician is heavily involved. (Though it inadvertently stepped on someone else’s toes since they had a proof in progress.)
Releasing just the Lean proof and collaborating on turning the raw data into something understandable seems like the way to go?
I don't see how this result changes that much, since we woke up yesterday morning and the thing that most mathematicians assumed was true, turns out to be true.
And the 17 million line long Lean proof is almost useless in its current form. Until you refine that and get it down to a "proof that fits on a T-shirt" (or at least the length of a typical journal article), I think it is just screaming out that there are large missing areas of mathematics that we don't know about. It may be useful to "mine" the proof for theorems that could be (for example) used to add to the Langlands program, and as those get added to Lean, all that work could start to wipe out millions of those lines of Lean code. That might be done with AI/LLM-assitance, but it'll be motivated by human experts diving into the proofs, and none of it will attract any headlines.
I think the human obsession with "first post" claims on proving things we already expect to be true will need to get reined in. This isn't a P=NP kind of result that would blow up the world if we woke up some morning to find out that was true. And 99.9% of the really useful work on this proof is still in front of us.
And I suspect that using an LLM to find results like this is like nuking a fly in terms of the computational effort required. If we had better global optimization algorithms (genetic algorithms, particle swarm optimization, and techniques we don't know about yet) surrounding the problem space of Lean proving, we could probably do this with orders of magnitude less computational power. I doubt that the transformer neural network is the optimal architecture to solve these kinds of problems.
"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!"
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.
Isn't the same true for every proof assistant? But only con-leche claims to be perfect 100% bug free software, so what makes con-leche special in this regard?
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.
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.
As for “a guy on the internet”, we’re now all people on the internet, the ivory towers are a thing of the past.
https://leodemoura.github.io/blog/2026-10-15-lean-now-ships-...
Why? Pretty sure you can just try to prove False over two reasoning chains C1 and C2 and then do something like (C1 || C2) --> whatever I want.
No offense, but do you think the mathematicians who study formal verification didn't think of this obvious trick?
I hear there's a teapot kicking around between us and Mars, too
It's making quick progress and is part of the Lean community.
https://arxiv.org/html/2607.13531v2
> †Supported in part by NSTC (Taiwan) grant 113-2221-E-002-204-MY3.
It's not well known yet though, only has two GitHub stars.
I mean, wouldn't it be great if AI could lead us to a world of lasting peace, prosperity, and happiness?
Not convinced that is within the capabilities of any AI without it doing things we would in fact not want it to do to us as a species.
We seem to be built for conflict and I'm not sure that what is left when that is removed is still us.
> Let me tell you something about Hew-mons, Nephew. They're a wonderful, friendly people, as long as their bellies are full and their holosuites are working. But take away their creature comforts, deprive them of food, sleep, sonic showers, put their lives in jeopardy over an extended period of time and those same friendly, intelligent, wonderful people... will become as nasty and as violent as the most bloodthirsty Klingon. You don't believe me? Look at those faces. Look in their eyes. (Quark from the always brilliant DS9).
Yeah I'm cynical <waves at history>.
1. when (the leaders of some) some belligerent tribe, nation or coalition etc perceive that they can promote their interests by going to war without risks that at first blush will outweigh the gains.
2. when some (the leaders of some) tribe, nation or coalition etc can credibly anticipate that they are about to be attacked and then go to war, because 'attack is the best defense'.
Proof abounds in history that war is not inevitable. One example: peoples all over Europe that used to engage in constant war with their neighbors in the next valley, next district, next-door nation, over hunting-, grazing-, market- and raw material opportunities have ceased to do so even if they are essentially the same peoples.
It's of course essentially just simple cost/benefit analysis, not blind human nature.
“All of humanity's problems stem from man's inability to sit quietly in a room alone.”
― Blaise Pascal
You need to have crystal clear clarity about who exactly is the Threat to the global peace.
Argentina, 1976, CIA backed coup
Brazil, 1964, CIA backed coup
Bolivia, 1971, CIA backed coup
Chile, 1973, CIA backed coup
Cuba, 1950s, CIA backed Batista
Dominican Republic, 1902–1924, US military occupation
Ecuador, 1961, CIA backed coup
El Salvador, 1980s, CIA funded dictatorship
Grenada, 1979, CIA destabilised country
Guatemala, 1954, CIA backed coup
Haiti, 1915–1934, US military occupation
Haiti, 1990, CIA backed coup
Honduras, 1980s, CIA backed coup
Nicaragua, 1979, US backed dictator
Nicaragua, 1982–1989, CIA funded Contras
Panama, 1941, US backed coup
Panama, 1989, US military invasion
Paraguay, 1954–1989, CIA funded dictatorship
Peru, 1980–2000, CIA funded dictatorship
Uruguay, 1973, CIA funded dictatorship
Ghana, 1966, CIA backed coup
Chad, 1982, CIA funded dictatorship
DRC, 1961, CIA backed coup
Philippines, 1965–1986, CIA funded dictatorship
South Korea, 1945, Outlawed PRK government
Cambodia, 1970, CIA backed coup
Indonesia, 1965, CIA backed coup
Laos, 1957, Multiple CIA coups
Greece, 1947, CIA backed junta
Italy, 1948–1996, CIA corrupted elections
Iran, 1953, CIA backed coup
Syria, 1949, CIA backed coup
Syria, 2013, US drone strikes
Niger & Sahel, 2013–2024, Counterterrorism drone deployments
Somalia, 2013–2026, Airstrikes and raids
Pakistan, 2013–2018, Covert drone strikes
Iraq, 2014–2026, Counter ISIS intervention
Syria, 2014–2026, Sustained military airstrikes
Yemen & Red Sea, 2014–2026, Naval and airstrikes
Libya, 2015–2019, Renewed bombing campaigns
Bolivia, 2019, Disputed coup backing
Iran, 2020–2026, Direct military strikes
Gaza Sector, 2023-2026, US backed and armed genocide of Palestinians in Gaza
Venezuela, 2025–2026, Raid capturing Maduro
Nigeria, 2025–2026, Joint military operations
Ecuador, 2026, Joint border airstrikesJust to say that it’s more than just scarcity of resources.
As great as it may be to have it in order to do what we want to do, energy is intrinsically, fantastically dangerous and must be handled with an amount of caution we have never been able to muster. I strongly believe that at-will energy production is all but guaranteed to destroy the world.
It's called solar power.
People are so enamored with their belief that AI can basically do magic and let them live in Star Trek that they forget solutions already exist for these problems.
We could already feed, clothe, provide clean water, clean energy, medical care and education to everyone on Earth with current technology.
We don't because, at scale, we don't want to. It's a problem with policy, culture and religion, with our primate minds being driven towards hierarchies of power and scarcity. If AI somehow made free energy available (which it won't because ye canna change the laws of physics captain), we still wouldn't want to.
Additionally, look at how AI is used in global conflict today, e.g. its role in the Gaza genocide, and you must conclude that this technology does not have a good track record for global peace, quite the contrary.
Society and culture is downstream of economics, which is shaped by available technologies.
Warp drive may have started Star Trek's humanity on the path to utopia, but I'm pretty sure that it's the matter replicator that made it work.
Every endeavour takes resources. Thus a pragmatic approach would be to recognise that resources are limited, and best strategy is pick low-hanging fruit first.
FTL travel requires new physics. And even then (or if found impossible), for example interstellar travel would be a massive undertaking involving all of mankind. And may involve "generation ships". Let's leave that for another day.
Otoh, supplying clean drinking water to every inhabitant of this planet is doable, easy, cheap & would be gamechanging for many of the recipients. Lets do it!
It's easy to see that barriers blocking this, are political/religious in nature, not technical. And thus politicians & lobbyists are thrown at it.
In the meanwhile, doing the math on what would be needed to reach Proxima Centauri, could be (relatively) easy. So sure, give some people a laptop & office space to see if they can work it out.
TLDR; keep chipping away at whatever problems you encounter. Far-future or otherwise. But be sure to pick low-hanging fruit first. $ Cost is a reasonable proxy for this. If it's cheap, give it a shot.
We may or may not get FTL out of this, but in the mean time, it would give mathematicians and physicists some space to breathe.
That gave me a chuckle.
But there's no unified theory, as quantum field theory and relativity don't get along and we still need a model for quantum gravity. So it's not possible to formalize the whole system because we genuinely don't know how the whole system works to form it's axioms.
You cannot formalize physics (and that's not what Hilbert's 6th problem is about). You can formalize some physical theories, but that's not the same as formalizing physics.
Physics isn't mathematics with different axioms. The ground truth is nature, not some axioms you need to fulfill. A formally verified theory can turn out to be physically wrong.
Just as an example because I see they suggest it as a project on their page:
I can formalize a tight binding Hamiltonian of graphene. I can write it down and then prove that this structure of the Hamiltonian means I get real eigenvalues, distributed symmetrically around 0, I get the [Dirac cone](https://en.wikipedia.org/wiki/Dirac_cone) at specific high symmetry points in momentum space etc. but at no point can I prove that this is actually a good/realistic model for graphene.
I could have chosen a very different Hamiltonian that includes spin orbit coupling or electron-electron interactions or different atomic orbitals or couplings to atoms further away etc. pp.
The only verification that can tell me which choice is a good one (not a correct/True one because you just don't have that in physics) is an experiment.
Their point is you can't do that. And they're right - this is, famously, Popper's point. You can empirically demonstrate it's not some mathematical structure, but not the converse.
Yes, it is a totally correct statement, but the thing is we are talking English, and in English the word "physics" can mean either "a collection of physics theories" or "the underlying reality these theories describe". The exact meaning depends on the context. Your statement provides a context to interpret "physics" in the latter sense. "Formalizing physics" definitely asks for the former meaning of the word "physics".
So, while you are totally correct, it doesn't mean that you've proven that formalizing physics is impossible. Quite the contrary, you claimed that one can formalize physics (in the meaning "some physical theories").
> The only verification that can tell me which choice is a good one (not a correct/True one because you just don't have that in physics) is an experiment.
I'm not into the project of formalizing physics, so I don't know how they choose their terminology, but if it was me choosing it... First, I'd chose "physics" and "the reality" to avoid using "physics" as referent to both. Second, I'd say that any "formalized physical proof" is true. I'm not sure how words I'd choose to talk about correctness of predictions of physical statements/theories. Maybe "empirical validity"? Why to invent a terminological bicycle if it was invented already? Though, maybe physics uses some other terms for that? If so, then we could stop talking about truth at all, and use "logically valid" to refer to statements that were formally proven.
But you probably can formalize our current understanding of physics, which seems useful?
From my very basic understanding, RH was generally already assumed to be true?
Like Deligne's "Weil II" paper.
As you said, we've largely assumed RH is true, but have had no proof - so the significance of a proof is learning new methods to prove things, which can then be applied to other areas of math.
Unfortunately, so long as we don't understand this new proof, it's not super useful, which is why the next step is to actually understand it.
Anyone else feel like we're approaching the limits of what humans are capable of understanding?
At a certain point, I assume we're bound to get a place where the frontier of "knowledge" (representable information interacted with by LLMs) is past the capabilities of humans to understand. Perhaps we're already there.
Just because humans can't understand things doesn't mean progress can't be made though
Without formal language and wise direction LLMs still seem useless, but with human operating them we can sort of "brute force" many maths like this.
We're still able to understand quite a lot of complexity, we just need enough eyeballs and time to do so.
In addition to that, there is still "accidental complexity", i.e. complexity that is there due to the way a system was created, not because it is necessary for the system to function.
Don't forget money.
Economics and politics drive a lot of the limitations on scientific (and mathematical) discovery. If there's not perceived ROI in understanding something, companies and governments will often just not bother so long as it's working or doing what they want.
No.
The entire purpose of math is to be at the limit of human understanding. A machine leap frogging the 'human understanding' bit is the problem. It doesn't mean humans lack the ability to understand, it just means humans can't preprocess tokens into their context at 40,000tok/sec
Like an example of the latter would be something like human history perhaps, where there are countless facts and events, but the cause-effect chain has only simpler links (like why did WW1 start is a complex question, but each link is quite likely a basic human trait (money, power, love, etc between different actors)).
For the former, the easiest example is higher dimensions.
There is no doubt that we (and probably every intelligence that can't dynamically scale up to arbitrary levels) have both kind of limits, but they are quite distinct and we can reason differently about the two kinds. I would just wager that some of these previously not cracked proofs are probably more on the horizontal axes and were not cracked simply because size (and effort that grows superliniarly with size) and not due to any of the steps being "ungraspable to measly humans".
(The same complexity axes are better represented in programming itself, e.g. a legacy, very big code base that has like 3 slightly different versions of the same function because it's so big no one "sees" across the whole is horizontally complex, while "smart" code (e.g. anything requiring complex, non-local analysis) would be vertical. In practice, each project has a non-zero amount of both)
This seems to more be a flaw in AI proofs than a limit to our understanding.
Hypothetically if we lived in a world where we only had a name for the number one, so instead of 2 you'd say the one one, then even simple problems like 10 x 10 would seem immensely complex.
AI might be good at writing (100 x "one"), but what we as humans want is the principles that make communicating about these issues practical.
As the numbers (proofs) get larger though, it gets harder and harder to see whether the machine miscounted, some of these results may take years to find the faults in.
All while "AI has solved X" will remain the headline in people's memories.
Yes, though people also largely assumed that the Exponential Time Hypothesis was true. And that one turned out to be false. See also a bunch of other AI-found counterexamples to conjectures.
Seems to be very much pushed by ehm AI enthusiasts/investors here and that's kinda it, doesn't matter much for everyone else.
That may lead to something in the future where there might be an actual impact on the broader world.
Basically gym training and not yet a ring fight ;-)
GG @ AI
Showing 400 of 816. Read the rest on Hacker News
Comments are loaded live from Hacker News and are not stored by Mid or Real.