Posted by jlebar 10 hours ago
Provides great context on this accomplishment, what it means but also doesn't mean.
I'd really like to make it the top link (and relegate https://www.anthropic.com/research/formalizing-fermats-last-... to the toptext) since HN has been tracking the work of https://news.ycombinator.com/user?id=kevinbuzzard for a long time and we're big fans. But I guess that would be overkill.
Gives you an idea of the scale...
Ugh we still don't know if this is true and it's nearly impossible to calculate without a full understanding of the real CAPEX cycle. Stop spreading these rumors until we know for sure.
Building the LLM that could do this work in 11 days cost multi billions.
The economics probably only make sense if LLMs prove to be a benefit to almost everyone in a way we can all accept.
Otherwise this cost a lot more than we’d otherwise pay. It was incredibly fast though. But we all know: cost, speed, quality. Pick two.
It's like having new solar panels installed every week. Sure you're "profitable" on the $0.20/kWh you're selling your "free" energy at when you ignore the cost of the solar panels you're buying every week.
Basically there was a choice between taking the money, and growing. They chose growth.
The US doesn't pay too much to healthcare, they pay too much to health insurance. Too much for too little value
Spending on health insurance is spending on health care.. Americans want free healthcare but no tax bump so health insurance is a compromise.. when even just ACA was passed and premiums increased, democrats got destroyed at midterms so Americans might be living in la la land.
I see funding of chatgpt as one of small part of a history where governments and industry fund basic science and moonshot programs, not to generate revenue, but to explore what is possible.
LLM funding is not aimed at improving our understanding of the world, it's aimed at making people reliant so that they may extract wealth through subscriptions for shareholders.
Americans don't get good healthcare and education because that's what they vote for, in elections and wallets. I am hopeful that that changes, but we shall see.
^ this section should have been in the first few paragraphs imho. Explaining why this is relevant shouldn't be so far down.
Explaining the value of what you are showing should always go towards the start. Else, why would anyone bother with the rest?
Giving such a blanket "responsibility" to the author at all is just such a bummer! I say let them do whatever they want, there is always more than one way to express oneself. Someone who was never taught to write a clear thesis in the first paragraph for whatever reason doesn't inherently have less to say.
Never heard of Abstract section? First semester on a college or last year on high school.
> In 2026, AIs designed to spot bugs in software were directed at Lean, and found several loopholes which were then fixed. Perhaps related to this effort, a purported disproof of the Collatz conjecture was announced as verified in Lean. However, this proof was soon determined to rely on a bug in Lean, and once the bug was fixed the proof was found invalid
However it's a bit different than the usual 'bugs' we encounter in normal software development. Lean is more like a type checker. If you can write a false proof in Lean then the bug is in Lean itself, not your code.
In other words, Lean can have bugs, but the amount of code we need to check scales with Lean itself, not with the length of proof. Just like the chance that C compiler has bugs doesn't increase as we write more C code. So the 13M lines of code doesn't really matter here.
You could imagine the typechecker has bugs (and indeed another comment mentions examples of bugs!). Crucially though anytime the typechecker has a bug fixed you could rerun the typechecker on the code to see if it still type checks.
This is the whole promise of formal verification. It reduces the problem of verification purely to the typechecker. If the typechecker is correct, then the proof is verified, no matter how many lines of code the proof is. As a sibling comment puts it, the chance of bugs mainly scales with the number of lines of code in the typechecker, not in the amount of lines of Lean code.
Your question is akin to asking, "yes this spellchecker ran fine on your essay, but are you sure it runs fine on War and Peace? That's 1000x more words!" To which the answer is the number of words doesn't matter if the spell checker is correct (which it might not be! And longer passages might reveal more bugs! But you can always rerun it). The main source of bugs is more lines of code in the spell checker, not in number of words in the text.
If the compiler certifies that the code indeed produces a term of that type, then the proof is correct.
So, only need to trust: (1) That theorem statement is correctly encoded (FLT has a very short 1 liner description really)
(2) Lean compiler is correct
As someone not very familiar with Lean, does it really just depend on the entry point / theorem being correctly encoded? Can intermediate statements ever be mis encoded or misinterpreted, or is this what would count as a “bug in the Lean compiler”?
My question to any mathematician reading this: does the above make ANY sense to you?
I ask that because I can read most technical material related to computer engineering, programming, hardware specifications etc. Even if I don't fully understand all details, I can follow them pretty well. So I wonder if professional mathematicians can look at the above and still make sense of it like experienced software engineers do for computer stuff.
This is very different to believing the proof, which would require at least a pass understanding the general approach, seeing that it all actually fits together, then going deeper. At some point you transition to relying on the Lean all hanging together, but as mathematicians we all draw that line somewhere.
But yeah, makes sense. Same thing if you saw news on someone's new database technique to improve performance. If they say the right words, don't say the wrong words, and if you cared enough you'd do spot checks proportional to the claim. If pressed you'd examine the source code, and run independent checks. But if smells roughly right, that's a good first approximation.
But not an expert on this.
While I don't know the specifics, and someone more "in-the-field" than me would recognize all the "named" theorems etc
I am aware that there have been minor issues that have come up with the formalization specifically, and that previous proofs for lower values of n were always needed.
Though it used to be n=5 and lower needed to be checked.
Vaguely. It's describing connections between a number of other mathematics results than can be connected to prove FLT. I assume all the work described is being done to make the proof more presentable, smaller, basically "prettier".
It sounds like they established a minimum and maximum bounds for n in x^n + y^n = z^n, where one proof works for n greater than or equal to 17, and another proof for n < 37 (when prime).
I believe the case (remembering back 40 years here) n is even is very easy, and n is composite and odd slightly less so. Neither really being in the ballpark of what they describe here.
Wiles-Taylor-Wiles was the original proof by Andrew Wiles, and its corrections.
Galois representations is about vectors over Galois extensions, which are essentially adding roots to regular numbers (rationals, integers, etc). That ties into the Langlands program, which is a big area in number theory (that I don’t know much about).
Together with flat deformations and Frey curve, I think they’re talking about a topic in algebraic geometry as applied to number theory.
I also recognize the name Eisenstein from my time as an undergrad, though two decades out and not working in the field I’ve forgotten what his work on ideals implied here. Ideals are a well-known topic though, a sort of structure inside a ring (set with + and *) that is closed under operations — like evens in the integers are the 2Z ideal.
So I’d describe it as “sensible with an undergrad background”.
It also convinced me I had no interest in that path. Setting aside the grinding work of producing a proof that can only be reached by existing years in the abstract and hyper niche isolation of the problem space (not to mention that you might never discover it or that it DNE), the anguish of the output being a paper or presentation or some other artifact of human symbology (_words_, really) that could at any moment be refuted by a single observation of a single mistake—-that sounded like hell to me.
An equivalent high schooler today probably sees things differently, in light of this news and the undeniable implications of LLMs on mathematics. Sturdy autoformalization tooling should with time completely dispel the aforementioned anguish, once our confidence in converting a human proof to Lean/etc. reaches that of a compiler translating Java application language to bytecode. Errata may always exist, but in practice these new methods will do wonders for rigor and peace of mind.
(I’m far less confident re novel discoveries. There’s too much chance of derivative findings based on something part of the training looking like genius but really just tiptoeing on the shoulders of humans, whereas autoformalization is absolutely convincing to me as transformative, particularly to check correctness of AI outputted proofs as mentioned in the post.)
Pretty insane. I suppose it lends further credence to the idea that anything that can be shown to be correct can be done by a model.
My strong hunch is that it was a joke - he knew how difficult the problem was and claiming he had a solution was I think a huge motivating factor for many mathematicians trying to prove it. The greatest nerd snipe troll in history.
How have we not merely substituted one verification problem for another?
13M lines does seem extreme and there is probably a lot of inefficiency given the way the proof was developed. Cutting it down is probably a long road, but is also a very well defined problem that AIs can probably just go do with enough time and budget now.
How can you be so sure its not result of inefficiency?
I have used Fable for formalisation and it will, unless I catch it, reprove results it previously had proven, inline, in other results.
> Pretty insane.
I don't think the count of "intermediate theorems" tells you anything. Here's something from an algebra textbook:
---
Let G be a group, let H be a subgroup [of G], and let N be a normal subgroup [of G]. Then
H ∨ N = HN = { hn | h ∈ H, n ∈ N }.
---
This says that the subgroup closure of H and N, the smallest subgroup that contains them both, is identical with the set consisting of all products of an element of H (on the left) and an element of N (on the right).
Part of the proof:
---
Suppose that x and y are elements of [the set of products hn]. Then x = h₁n₁ and y = h₂n₂, where hᵢ ∈ H and nᵢ ∈ N. Now h₂⁻¹n₁h₂ = n₃ ∈ N, as N is normal in G. So n₁h₂ = h₂n₃. In this case
xy = (h₁n₁)(h₂n₂)
= (h₁(n₁h₂)n₂)
= (h₁(h₂n₃)n₂)
= (h₁h₂)(n₃n₂),
---
This will translate directly into lean. If you do it this way, you will prove at least 10 of what would be described in lean as 'intermediate theorems':
∃ h₁ ∈ H, ∃ n₁ ∈ N, x = h₁ * n₁
∃ h₂ ∈ H, ∃ n₂ ∈ N, y = h₂ * n₂
h₂⁻¹ * n₁ * h₂ ∈ N
n₁ * h₂ = h₂ * n₃
x * y = (h₁ * n₁) * (h₂ * n₂)
(h₁ * n₁) * (h₂ * n₂) = (h₁ * (n₁ * h₂) * n₂)
(h₁ * (n₁ * h₂) * n₂) = (h₁ * (h₂ * n₃) * n₂)
(h₁ * (h₂ * n₃) * n₂) = (h₁ * h₂) * (n₃ * n₂)
h₁ * h₂ ∈ H
n₃ * n₂ ∈ N
At $50/M output tokens, this would have cost on the order of $300k (plus a bit for input/prefill tokens) at API rates.
It also uses Prove2Me, which uses a graph like previous automated theorem provers. A fact that LLM hawks have categorically denied here before, with opposition naturally flagged.
Now they have it in writing.
Yeah, because before now there's been literally zero proof of an automated theorem prover scaffold around the LLMs being used, and big counterexamples and such being found, with raw chat logs available, where no such thing was used.
> Now they have it in writing.
Yeah, because now it's actually being done. They talk about it as a novel thing, because it is. You don't get to claim being "right all along" from this
> The work certainly achieves some of the aims of the EPSRC project, and indeed it goes much further in terms of what is formalized (I only promised the EPSRC that I would reduce FLT to the 1980s; this repo proves the whole thing). But I also promised several other things to EPSRC: firstly, that I would be making pull requests to Lean’s mathematics library, adding fundamental objects from modern number theory; this is ongoing. And secondly, and perhaps most importantly, that I would be creating a dynamic document enabling humans to explore the modern proof.
> Daniel used OpenAI internal models to discover new soundness issues in the official Lean kernel and runtime
https://leodemoura.github.io/blog/2026-8-24-postmortem-for-t...
They found several bugs and they have patched them. Lots of work going into making sure lean is sound.
We simply don’t know what those 13M contain and whether it “makes sense” and doesn’t trigger Lean bugs. (There are “independent” lean verifiers, but historically they contained the same, or similar, bugs.)
I think they should spend another few billion tokens and let agents try to disprove any of those statements or links between them. Then I'd be a lot more convinced.
Meaning, people and LLMs are finding 1=0 bugs in formal verification tools. I have no idea how likely this is in this case, though!
You probably heard about Goedel Incompleteness -- the proof that the the axiomatic itself cannot be proven, like using ZFC to prove ZFC, but that's another topic.
It would be fun to play with this Anthropic/Lean formalization under different axiomatics.
Godel theorems are for systems with basic arithmetic, zfc doesn't include arithmetic, thus are not object of Godel theorems.
you understand that "expressive enough to produce" are not obvious elements of zfc, that's some average consumer napkin math and not strict formalization.
I said I am not expert, I am indeed not expert in zfc and godel theorems, but I am an expert (phd) in actual formalization theory. Formal theory is very simple concept: its alphabet, set of formulas on top of this alphabet, and function which translates one formula to another.
ZFC can't "obtain" peano, simply because it doesn't have say * operator defined. You need to do something on top of it. Additionally, zfc itself looks like loosely formalized say in wikipedia (and I am not sure if there is any strict formalization anywhere), we take it as common sense that it can utilize some simple logical rules (e.g. modus ponens), but what are exactly rules, which could be separate topic of research, this detail is skipped.
> support your point with explanation or be ignored :-)
Anyone who says "Godel theorems are for systems with basic arithmetic, zfc doesn't include arithmetic, thus are not object of Godel theorems" and isn't joking warrants a permanent ignore.
https://math.stackexchange.com/questions/1366560/why-does-g%...
https://math.stackexchange.com/questions/1090437/how-to-prov...
https://en.wikipedia.org/wiki/Zermelo%E2%80%93Fraenkel_set_t...
its hard to me to tell what this means formally(as I said I am not expert). There is no "interpret" operator in zfc. I believe what it says if you add some robinson axioms + some logical rules on top of zfc, you can carry your results.
If we take ZFC (or some other set theory) as our meta theory, we can easily see that the axiom of infinity (of ZFC) gives a set of natural numbers (using the von Neumann encoding), which, when equipped with the successor function, is a model of the natural numbers.
Also, I am not sure successor function is enough for PA.
zfc itself is not sufficient, you need some layers of extra concepts formalization to fit specific problem domain(e.g. zfc doesn't define even basic arithmetics), which also could have potential issues.
I am not entirely sure about lean, but the core algebras for systems like lean are in the 100s of lines of code.
You can likely convince yourself it is correct in a weekend or less - especially with an Ai to help you understand it.
And granted, I don't know the exact details about Lean. It might be that they don't have an incredibly simple core - as has elsewise been the norm.
https://leodemoura.github.io/blog/2026-3-16-who-watches-the-...
...and for those who are looking to roll-their-own:
https://ammkrn.github.io/type_checking_in_lean4/title_page.h...
...and some thoughts on putting stuff in the kernel:
I optimistically expect to witness the advent of a global 'panacea' in my lifetime thanks to AI's efforts. Cost effective large scale genetic engineering, a cure for every disease, potentially even a cure for aging.
The future is both beautiful and terrifying.
If you're happy to die, why be bothered by others' trying to live longer? You won't be around. And assuming people can finance it themselves, is it really a problem for society?
I'd also say people may want more life for themselves, but what does that mean at scale, forever?
There are many reasons that people living forever would be a problem for society, the most obvious being an ever-increasing population.
Child mortality is very low now compared to the past, thanks to the modern medicine and technology.
I am glad humanity "played God", and reduced this unnecessary child suffering.