edit: I need to learn to read, that's the first part of their statement.
The statement of the Riemann Hypothesis is correct. This statement is part of Mathlib: RiemannHypothesis and has been carefully vetted by experts in Lean and number theory.
While I have not studied this topic enough to really understand _how_, but I've been told that type systems laid out in frequently cited papers by the likes of Girard and Martin-Löf were found to be unsound more than a decade after publication. These issues were then "patched", so to speak.
Can anyone better versed in type theory explain how likely that is for the type system of Lean?
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.
In short, Lean has formally verified checkers now, and the RH proof was checked with them. So any such soundness hole would have to be exploited in the correctness proofs of the checkers, which is a much smaller attack surface than the RH formalization.
Depending on how the checker works, you might be able to run it on itself. But that doesn't say much; maybe it only approves itself due to a bug.
Not at all saying it is inconsistent, but I personally find the liberal use of introducing inaccessible cardinals axioms a bit concerning.
I don't think so. They depend on some kind of type theory and formulate everything from those primitives.
Many problems may not require even full zfc. zf, peano or even weaker may be enough.
There are also many theorem provers not just lean. lean depends on type theory, but metamath for example depends on zfc directly.
It may be a good but likely tedious exercise to convince yourself by translating the current proof to a zfc dependent prover.
(1) A lot of the mathematics that was done in the 18th century was
obviously inconsistent and yet still useful.
(2) Trying to verify logical consistency of a theory is basically impossible
because of Goedel's 2nd Incompleteness Theorem.
Hitting a wall like this makes me think we're asking the wrong question.The theory and implementation of Lean is very much so a result of human involvement, on a massive scale even. Just because it is something repeatable does not lower that in any way.
Distrust of LLM output should not turn into distrust of automation, I think. The former I totally get (and deeply hold myself), the later is likely to hold one back - imagine banning calculators because there is not human in the loop, to make an extreme example.
This is pretty much it. Pre LLMs very little maths had been formally verified and a lot of it relied on proofs from humans convincing other humans, and issues did arise which would occasionally mean chunks of maths we previously thought were true had to be reconsidered.
Proving something in lean probably involves writing a lot of proofs for new areas, and the formal proofs are usually significantly more verbose than a human proof where a lot is often left for the reader.
The main theorem's cluster directly imports only three others: a zero-free region, primes in progressions and Mertens-type zero sums. Most of the code sits 3–5 import steps below those, and no theme is more than about 4% of the code.
(I made the map. Cluster names are written by Claude from code samples, so they can be wrong.)
Have you tried asking it to?
If everyone got to do that then we would really know how rare his talent is. But we don't, and he doesn't post his prompts or workflows. So how can you be so sure.
If he was at least posting his prompts we could acknowledge that. But honestly I'm not going to give him the benefit of the doubt on this. Highly accomplished and successful yes, but uniquely talent who knows.
Most of the discoveries in the last 2000 years were made by individuals who were born into wealth, had the right set of circumstances, or were blessed by the powers that be at the time.
Imagine all the smart minds throughout human history who toiled away in fields collecting food and will never be known about.
Unfortunately scientific discovery has always been unevenly distributed.
Poincare Conjecture (Perelman) just lived with his mom in a small apartment.
Erdos was a homeless dude who couch surfed.
Yitang Zhang (Bounded Gaps between primes) truly has a humble background of a farm labourer born in Chinas Cultural Revolution.
Some inequality is fine, world isn't fair, but this a little bit ridiculous.
I'd heard something like this in terms of results (the "glass ceiling" of there being male CEOs gets talked about, but not the "glass floor" where the most dangerous and shitty jobs are also done by men, to an even larger extent), but hadn't heard something like that applied to IQ.
There's pretty widespread consensus in feminism and gender studies that traditional gender roles impact our decisions gravely. One such widely-held traditional gender role is that men are the “protectors” and the “fighters”.
So no, I don't accept that men in dangerous jobs are stupid. I think for the most part they are in those jobs because they feel their society expects it of them in order to be seen as “truly masculine”.
It's not just the proof that's valuable to scientific discovery, it's the process of using a powerful tool at an advanced level that others could learn from. It would also help to show concrete evidence of any sudden shifts in the AI that might indicate training data was siphoned from other researchers (which I don't necessarily think happened here, but is always an ongoing question).
If he truly did have non-trivial mathematical input to this it should be trivial for him to give talks on at least the overall strategy of the proof. Before LLMs this was a bare minimum expectation for basically all mathematicians and their results, and I see no reason why we should give up on it now, especially so since the previous primary artifacts (the papers themselves) are now on their way to being completely automatable.
I wouldn’t understand any of it but I might catch some glimmer of what he brought to this. My intuition is that is quite a lot.
Clearly not true given that multiple labs have many PhDs attempting this and one guy got the Riemann Hypothesis, a large contribution to Navier-Stokes, and multiple others just in the past few months.
And OpenAI takes the (correct) approach of attributing credit to the model rather than the mathematician.
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...
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.
It's making quick progress and is part of the Lean community.
But you probably can formalize our current understanding of physics, which seems useful?
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.
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.
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.
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.
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.
Just to say that it’s more than just scarcity of resources.
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. 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.
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.
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 airstrikes“All of humanity's problems stem from man's inability to sit quietly in a room alone.”
― Blaise Pascal
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.
Interpreting claims that a theorem is correct because it was checked in Lean | 329 points 5 days ago | 148 comments | https://news.ycombinator.com/item?id=50515058
From my very basic understanding, RH was generally already assumed to be true?
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
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
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.
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.
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.
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 ;-)
Like Deligne's "Weil II" paper.
Seems to be very much pushed by ehm AI enthusiasts/investors here and that's kinda it, doesn't matter much for everyone else.
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?
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
P=NP should fall next week to Grok and Elon then I suppose? What an exciting time for you.
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".
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.
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"?
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.
It seems the modern unsolved problems might have solutions incomprehensible by humans. So a new class of decidable problems emerges: llm-complete ones.
Can someone ELI5 how they know the proof is correct? It is not like anyone had the time to read 17 million lines.
The biggest concern that people have is that there might be a bug in the type checker that the proof is exploiting. But a) it's been compiled with multiple independently-implemented compilers, which are are unlikely to have the same bugs, and b) some of those compilers have themselves been proven to contain no bugs. So, it is exceedingly likely that the proof is legitimate.
Now, whether we can understand the proof is an open question...
You look for any unsafe blocks, and then run the type checker.
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.
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".
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?
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.
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.
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.
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?
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.
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.
Let's not sell humanity short just yet.
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.
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.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.
https://gwern.net/doc/fiction/science-fiction/2000-chiang.pd...
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
Now I'm telling my computer to program for me.
We need to start thinking about how to move on from our status quo soon
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.
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.
But remember, for now you kinda need it a real big cluster to run a model smart enough so its not that simple
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.
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.
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.
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.
I’ve met a lot of rich people. They don’t think about you.
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?
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?
Overrated. Most training of a frontier LLM is done on synthetic data and reinforcement learning.
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.
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.
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
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.
Think of what happens to our cattle, but now.. worse. Way worse.
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.
Absolutely based and, if I may say so, extremely redpilled.
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.
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'...
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.
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?
As for “a guy on the internet”, we’re now all people on the internet, the ivory towers are a thing of the past.
I hear there's a teapot kicking around between us and Mars, too
when the first wheels were invented - humans could move at speeds (10-15 miles per hour, for hours and hours) that were impossible for the last ~450-500k years (basically since humans existed). a lot of humans thought there might never be a better design; then slowly and slowly(about 5000 years) we made progress and we had cars made by benz and ford which broke all those previous records (going 30-40, even 50 miles per hour) and then only ~70 years later, humans reentered the earth's atmosphere at whopping 20k miles per hour - read that number again, 20,000 miles per hour and that is the fastest humans have ever moved in the known history. do i think that's it? hell no.
99.999999999% of the population are not even in the galaxy of the ballpark of understanding of any of it and what its purpose is.
There's no existential dread.
I felt a lot of that yesterday, for what it's worth. This morning, on a walk, I reached a similar conclusion to the original comment: it's a marvellous tool that's helping humanity reach further, faster.
The existential issue comes from wondering deeply what my own difference is from these tools, why I think I think and they can't. I actually started wondering: what if they _can_ think and the problem is that _I_ can't. It'd be pretty astonishing if it turns out that we're the ones that are truly stochastic parrots, but the LLMs turn out to be something more.
It wasn't Claude woke up and thought, "You're absolutely right, I'm going to crush some math today"
I worked at a university for a time. One thing I learned is that while people in mathematics and physics and biology etc are brilliant, they are terrible at actual programming.
I wonder if there will be a push by the CS community to review the solution, refactor it, simplify it, etc. Surely that can help remove noise and express the solution in a much simpler way.
For reference, the US tax code is over 1m words (about 6000 printed pages), and the associated regulations, rulings and IRS guidance add about 75,000 pages more.
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 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.
But does it matter? Would the Pythagorean theorem be less useful if it took a million lines of lean to prove?
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
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.
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.
"con-leche and con-ron checkers"
Who comes up with these names? :-D
"con-ron" - con-leche in Rust - https://github.com/leanprover/con-ron
They are also wordplays on Spanish drinks viz. "Cafe con leche" and "Cafe con ron" ;-)
> The Riemann hypothesis for ζ(s) does not seem to be any easier than for Dirichlet L-functions (except possibly for non-trivial real zeros)
https://github.com/leanprover/con-leche
https://github.com/leanprover/con-ron
I think the idea here is that because they are independent of the core Lean kernel, they provide some level of reassurance that the RH roof is not somehow unknowingly exploiting flaws in the kernel. I'm not a mathematician though.
I don't know where the names come from.
https://github.com/leanprover/con-ron: "con-leche in Rust"
The closest I've found is Google Deepmind's "Formal Conjectures" repo (https://github.com/google-deepmind/formal-conjectures), but that's not quite what I was thinking of.
“Any sufficiently advanced technology is indistinguishable from magic.”
We’re just seeing more and more smart people encountering Arthur Clarke’s third law.
If so, my understanding is that it is unverified, and even if it were, it also depends on proving one of two additional spectral hypotheses concerning zeros of a quaternionic function constructed from an operator, and even then, it only proves the Goldbach conjecture for "sufficiently large" numbers, without a way to calculate that n yet.
The obvious step up is P vs NP.
Solving these extremely hard problems that humans, for decades (and more than a century in some cases), have been unable to is the mark of superhuman intelligence, in my opinion. So far only OpenAI and Anthropic have crossed this mark, and it would be great to have more to join them.
"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!"
It's fascinating to watch the skepticism in some of the reactions - "what if there's a bug", "it doesn't mean anything without human understanding" - just anything possible to undermine, dismiss and move the goalposts.
Although opening a new chat and asking if the previous output was correct, it detected the mistakes.
You also have to consider if an LLM can solve the Riemann hypothesis because it is 10x smarter than any mathematician, or because it is willing to write 1000x as many lines of code (including false starts) without giving up?
its because models are not trained on terrabytes/petabytes(?) of RSI generated data in your domain, unlike math and coding.
Also, there is a chance you/we can't always comprehend how smart they are.
LLMs exhibit what is known as "Jagged Intelligence" - https://news.ycombinator.com/item?id=48577159
Mathematics is one domain where they have high peaks. Why? Because we have "formalized" mathematics so well that starting with the precise and correct set of problem statements/assumptions/etc. if you follow the logical rules of inference (i.e. deductive system) then you will reach the conclusion. Formalization was invented so that we can solve problems purely mechanically at the symbolic syntax level using a machine. So the LLM proposes proof steps to the Lean proof assistant which then symbolically verifies it.
On the other hand, LLMs often fall down in "common sense reasoning" (https://en.wikipedia.org/wiki/Commonsense_reasoning) since this involves a lot of implicit context etc.
It's not like Anthropic simply gave internal Claude agents the task to solve RH and months later they came back with the solution.
> ... has been carefully vetted by experts in Lean and number theory.
So, if the vetting was not careful enough and done by people expert enough, what could happen?
. ... 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.
And what verified con-leche and con-ron?
> These checks ran on an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof source
Hmmm...
Guess I'm splitting hairs here; like I said, maybe this is as close as you get to formal proof verification, but still...
In normal life, multiple qualifications like that mean a lack of confidence that should be taken as lack of trust in the statements overall.
But in an area like this where it's all about formal verification and the consequence of bad link in the chain is severe, more qualifications should give an informed person more confidence.
1) Surely agents can be employed to simplify the proof now? "Make it smaller while it still verifies" in a loop; maybe, to a level where a human (research mathematician specializing in RH) can grok it?
2) con-leche is vibe coded; isn't it a concern w.r.t. trusting its soundness?
3) This is so exciting. Please prove P != NP already and spare us the tension
GG @ AI
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?
It looks like the meaning of the word "mathematical proof" changed fundamentally. Mathematics looks like a flattened animal on the asphalt by the CS steamroller.
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
See question 4 here -
https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
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).
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"