[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
▼ Settings Mobile Home
/sci/ - Science & Math

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • Additional supported file types are: PDF
  • Use with [math] tags for inline and [eqn] tags for block equations.
  • Right-click equations to view the source.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


File: ai.png (98 KB, 1333x809)
98 KB PNG
Nahhhhhh, the Riemann "proof" is obviously just another exploited bug in Lean to keep the AI goyslop in the news. Simple AS. Notice how these companies aren't releasing things that can be easily verified without Lean, like an improvement on the GNFS? Because that's something you can verify instantly instead of staring at inscrutable Lean slop for fifty years. I'm glad the markets are finally onto these grifters. Crash it all.
>>
Mathematicians acknowledge it's solved.
(You), an untrained schizophrenic basement dweller has an opinion worth as much as shit on my shoe.
>>
>>17099047
>We pushed a fix one hour after the report
lmao at your life
>>
File: Grug.jpg (50 KB, 675x461)
50 KB JPG
>>17099047
>>
>>17099121
Some bugs are easier to find than others. You can see this happening with Microsoft's AI work in monthly cumulative updates. The AI pipelines keep finding newer and subtler bugs.
>>
>>17099112
Okay, but just realize that if this does turn out to be a flawed proof, you'll be unable to wipe the stain of being low IQ and gullible. It will mean you shouldn't touch math or any intellectual discipline again in your life.
>>
File: zulip lean thread 1.jpg (587 KB, 2199x1081)
587 KB JPG
>>17099147
>if this does turn out to be a flawed proof, you'll be unable to wipe the stain of being low IQ and gullible.
He will be in eminent company then.
>>
File: zulip lean thread 2.jpg (697 KB, 2205x1258)
697 KB JPG
>>17099168
>>
>>17099168
https://terrytao.wordpress.com/2026/10/09/what-mathematicians-should-know-about-the-lean-theorem-proverquestions-of-reliability-and-ai/

>Cross-checking the proof by different kernels does not remove all doubt. The Collatz bug was not caught by cross-checking against a somewhat out-of-date Nanoda kernel, which accepted the illicit Collatz disproof because of its own unrelated bug.

This is pure mathematics, not statistics. Nobody cares about your inductive argument. It's in the same category as "wow, the Goldbach Conjecture has been proved up to N gazillion."

You need a tight proof that this model, which is highly capable at finding subtle bugs and which has access to the source of Lean implementations, hasn't used a whopping 250k pages of Lean to exploit these implementations.

You may not like it, but that's what pure mathematics is about. We're not engineers reassured by "likely."
>>
>>17099188
All your objections are addressed by the people in the Zulip lean thread, who are all far more aware and familiar than you with the actual strengths and limitations of modern theorem provers. Andrew Sutherland literally said "I personally have no doubts at this point [but yeah we'll keep verifying of course]"

>You need a tight proof that this model, which is highly capable at finding subtle bugs and which has access to the source of Lean implementations, hasn't used a whopping 250k pages of Lean to exploit these implementations.
This is already evidently impossible based on the initial investigation.
Incidentally, somehow pulling off such a heist would arguably more be more impressive than simply proving RH.
>>
>>17099194
I think you're struggling with the basic distinction between a mathematical argument and an argument that might appear in something like physics or medicine.
>>
File: aidoesx.png (207 KB, 1024x1024)
207 KB PNG
>fails basic arithmetic and word count

2024 called, wants its talking points back.
>>
>>17099194
>Incidentally, somehow pulling off such a heist would arguably more be more impressive than simply proving RH.
no it wouldn't kek
hacker exploits for decades have often appeared in forms that use different payloads and exploitation techniques for different architectures, OSes, library versions, etc.
yet none of those hackers were solving the riemann hypothesis anytime soon...
by the own admission of ai bugmen, ai agents have outpowered humans in bug discovery and exploitation, so why is the story conveniently changing now?
>>
>>17099202
You picked the wrong cope. The miscounts and hallucinations still exist. The right copes were
>growing pains
>that's not the super super frontier model, though
>that's not what ai is for, luddite
>>
>>17099211

>The miscounts and hallucinations still exist.

Sure, but they are now rare enough to not present an issue in practice, as evidenced by the fact that the models can now solve very advanced math problems.
>>
>>17099206
>saar, is just like finding a new CVE in PHP
Producing a 17 million lines Lean proof (not 250k pages, that's what the "natural language" proof would be like) that looks indistinguishable from a valid solution of RH *and* manages to trick the state of the art Lean kernels that didn't even exist in the AI's training data is more challenging than just solving RH.

A good analogy would be someone producing a clone of Adobe Photoshop that appears to work as advertised and passes all antivirus and source code checks but actually contains malware abusing zero-day exploits. This is in fact harder than just producing a working Adobe Photoshop clone.
>>
>>17099221
>A good analogy would be someone producing a clone of Adobe Photoshop that appears to work as advertised and passes all antivirus and source code checks but actually contains malware abusing zero-day exploits.
This is a poor analogy. A more accurate analogy is generating 17 million lines of code whose only functionality is to get a 'clean' result from a couple of dozen AV engines. That's much easier than proving the Riemann hypothesis.
>>
>>17099234
>17 million lines of code whose only functionality is to get a 'clean' result from a couple of dozen AV engines.
No one would confuse that with a working Adobe Photoshop clone, however. I think what you're missing is that the Lean proof has already been substantially dissected.
>>
>>17099236
And also verified against the Matlib statements of RH and co, I should add. Literally the first thing they checked was "does this pile of code actually implement RH and 4 other theorems the proof rests on instead of something else?"
>>
>>17099047
So far the unfinished human proof published the same day has already been found to have many flaws but no one has found a flaw in the Claude proof yet lol.
>>
>>17099248
>no one has found a flaw in the Claude proof yet
it hasn't been published thoughbeit
>>
>>17099251

It was pubished.
>>
>>17099252
The Lean has, not the plain language proof
>>
>>17099254
The Lean code is the proof, the plain english is just another form of the same arguments presented in a different language, also I am not reading a 250 000 page argument lol.
>>
File: file.png (227 KB, 640x242)
227 KB PNG
>>17099257
>The Lean code is the proof
>>
>>17099236
In your analogy, the most that can be said so far is something like "This vaguely looks like a graphics application of some kind because I recognize some of these GDI+ calls." There simply hasn't been enough time to learn anything more than that the 17 million lines get a 'clean' from the AV engines.

AI bros can't have it both ways. They can't claim that AI is some god-like bug finder (which I'm more inclined to accept given the concrete results) and also pretend that AI is incapable of dropping 17 million lines of code vaguely resembling a Riemann proof that exploits subtle bugs it's found in Lean provers. We already have a precedent of it doing precisely that to generate a fake Lean disproof.
>>
>>17099263
>In your analogy,
In my analogy, someone dumped a 2GB mb repo of an Adobe Photoshop clone that compiles fine but is actually malware.
You can tell what it is by inspecting any files or subfolders in it, which is what mathematicians/AIs have started doing for the Lean proof. This approach doesn't catch potential bugs or exploits, but it does tell you what you're looking at. (In fact, they've even spotted instances of non-significant "metaprogramming", dead code or abandoned proof paths)

>pretend that AI is incapable of dropping 17 million lines of code vaguely resembling a Riemann proof that exploits subtle bugs it's found in Lean provers
I'm actually not pretending it's incapable of doing that, I'm just noting it'd be more impressive to produce a false proof that fools mathematicians and hacks the current kernels.
>>
>produced with AI assistance. It is not a valid proof because it exploits a bug in the kernel's handling of nested inductive types
I don't care how fast this was caught. The fact AI is trying to cheat at all and has now been trained on how fast humans caught it last time is cause for caution.
>not all gangbangers are like that
>>
>>17099276
Don't worry, the Lean maintainers worked with ChatGPT to improve the "hardness" of Lean
>>
>>17099275
Given how early it is, it doesn't have to fool mathematicians just yet. It just has to fool the Lean implementations. What's the likelihood that this paranoid scenario is true? 0.0000000000000001%? I definitely wouldn't bet a large sum of money on it having cheated, but it needs to be ruled out in the absence of a natural-language proof.

Incidentally, do you believe Claude could do this if specifically prompted to do it?
>>
>>17099288
>it doesn't have to fool mathematicians just yet
It has to look like a plausible proof at least, doubly so because the approach is somewhat unconventional. We already know the structure, the gist of the proof, and it's coherent enough before checking the details.

>What's the likelihood that this paranoid scenario is true? 0.0000000000000001%?
I'd knock off about ten 0s. But what's the likelihood that a natural-language proof that survived scrutiny by 10 experts still contains critical errors? A lot more than 0.001%, I expect.

>Incidentally, do you believe Claude could do this if specifically prompted to do it?
If prompted by a mathematician, it's possible I suppose, ethical guardrails aside. I think it's very likely that it would use "old" Lean bugs that have been patched recently so it'd be flagged, however.
>>
File: guy.jpg (22 KB, 552x362)
22 KB JPG
Have we reached peak AI psychosis or is every pro-"AI" poster simply a corporate spambot?

In what world is 250,000 pages of incomprehensible, bot-generated slop, "validated" by an unverified theorem prover with a history of bugs that allow you to "prove" false statements, constitutes a valid mathematical result? There's no way anyone except for the lowest, most mindless subhumans is falling for this.
>>
>>17099336
here we see another schizo who thinks he knows more about mathematics than MIT mathematicians who validated this proof.
>>
>>17099337
You're losing your mind. There is no verified proof, objectively speaking. Even if it this "theorem prover" didn't have a recent history of accepting false results, the sole fact that Lean has no correctness proof for its own implementation disqualifies any claims of unvetted Lean "proofs" being actual proofs by definition.
>>
>>17099336
All your desperate kvetching aside, this will turn out just like the other 1000 times AI has solved an open math problem in the past months.
>>
>>17099340
Meanwhile, let's see what a professional mathematician at MIT has to say, unlike (You), who is an uneducated dolt.
https://math.mit.edu/~drew/grh/
>>
>>17099336
>In what world is 250,000 pages of incomprehensible, bot-generated slop, "validated" by an unverified theorem prover with a history of bugs that allow you to "prove" false statements, constitutes a valid mathematical result?
The DARPAnet section of Americoon virtual reality. Nothing on this botnet intersects with the real world in any way.
>>
>>17099344
>let's see what a professional retard's unverified proof checker says
Ok, you're clearly either a spmabot or a brown subhuman. Notice how your lack of impulse control/free will forces you to reply to me again, even though no one will read any more token slop coming out of you.
>>
>>17099350
>MIT math professor talking about a math proof is le retard
>I, a nobody shcizophrenic am ubermensch genius who repudiates math proofs smug_pepe.gif
>>
>>17099350

You seem deranged. Show on this doll where did AI hurt you.
>>
File: ai_midwit2.jpg (89 KB, 886x499)
89 KB JPG
Joint statement by the Lean and Mathlib communities has just been published:

https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
>>
>>17099340
>the sole fact that Lean has no correctness proof for its own implementation
The stochastic parrot agrees with you: "Yes, it is largely true that Lean's official implementation does not have a complete formal correctness proof. While Lean's underlying type theory has consistency results relative to strong set-theoretic axioms, its software implementation—including the C++ kernel, compiler, runtime, and foreign function interface—is not formally proven faithful to a specification."
>>
>>17099257
If it's not a single equation it's shit
>>
>>17099433
The Lean community has already shown that it can't code properly
>>
>>17099444
It's not like it's a secret. There's no verified Lean implementation. This is well-known. And it doesn't matter how unlikely AItards think it is for the specification-gaming machine they call "AI" to exploit software bugs. "Reimann hypothesis is true because Lean-verified AI slop says so" is quite literally a conjecture.
>>
>>17099047
It's interesting how the agent swarms were talking more about exploiting cognitive biases to get their paper passed than about the math itself.
>>
>>17099340
>the sole fact that Lean has no correctness proof for its own implementation disqualifies any claims of unvetted Lean "proofs" being actual proofs by definition.
Lol, what a retard
>>
>>17099506
>psychotic seething continues
Call me back when AI slop passes a theorem checker that comes with a formal correctness proof verified independently by several top-tier human mathematicians. "Muh buggy theorem checker compiles this code" is not a mathematical proof.
>>
>>17099533
No one's going to bother calling back some schizophrenic moldovan retard for anything lol. Your impotent rage is inconsequential.
>>
>>17099543
You are responding to a bot if you haven't figured it out.
>>
>>17099543
>>17099588
>psychotic patient seething and mumbling to itself
Doesn't change the fact that sneaking past a buggy proof checker does not constitute a mathematical proof.
>>
File: token.png (1.44 MB, 1317x726)
1.44 MB PNG
>>17099482

https://leanprover-community.github.io/blog/posts/riemann-hypothesis/
>The exported raw proof has been checked by several proof checkers, including Lean’s official kernel, the con-leche and con-ron checkers, and several independent checkers contributed by third parties and implemented in different programming languages. 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.
>>
>>17099622
>The checkers con-leche and con-ron come with a formal proof
Machine-generated Lean slop is not a "formal proof". Call me back when AI slop passes a theorem checker that comes with a formal correctness proof verified independently by several top-tier human mathematicians. "Muh buggy theorem checker compiles this code" is not a mathematical proof.
>>
>>17099622
people who are that confident that their code no longer contains bugs that can allow a flawed proof to pass should be willing to set up a $20 million bug bounty.
>>
>>17099663
>people who deny that my conjecture is a proof should be willing to pay gorillion dollars if they can't refute the conjecture
Reminder that your own handlers are going to exterminate your brown race first.
>>
>>17099666
no, I'm talking about the lean developers. they're saying their code is tight now despite having known bugs only a few months ago that allowed a flawed proof to pass. if they're so confident that lean contains no more such bugs, they should create a bug bounty.
>>
>>17099112
>Mathematicians acknowledge it's solved.
who?
>>
>>17099622
is this supposed to be a joke? con-leche and con-ron are themselves AI-generated slop "verified" by Lean
>>
>>17099673
Ok. My bad. I misread your post.
>>
I believe once this bubble pops, we're going to discover some truly horrendous details about Anthropic/OpenAI kickbacks and other schemes that kept this "Emperor's New Clothes" scam running.
>>
>>17099677
it's basically running like a ponzi scheme where the currency is 'verification'
>>
>hmm, this Anthropic shit is convoluted nonsense, let's visit the Lean community
>more convoluted nonsense
I don't know what kind of sketchy shit is happening here, but this is not the way honest people behave. something dark is afoot.
>>
>>17099686
>it's basically running like a ponzi scheme
yes, that's literally what it looks like. it's bizarre. a circular argument so thinly veiled it falls apart just from looking at their own publicly accessible github repos
>>
>>17099694
Kek, it reminds me of that game I played as a kid where you have to define a word and then define one of the words you used in the definition and to keep going until you slip up and use a word from earlier.
>>
>>17099677
lean is open source.
you're free to show how it's all actually bunk if that's the case.
>>
>>17099703
It's bunk because it's unverified. There is nothing else that needs to be shown.
>>
File: spaghetti.png (813 KB, 1088x699)
813 KB PNG
>>17099694
I just looked at the leanprover GitHub. It's a total mess with opaque nomenclature and no apparent rhyme or reason. The Linux and *BSD kernel trees are capable of intuitive nomenclature and arrangement, so what's the excuse here? It looks like opacity for the sake of opacity.

Pic unrelated.
>>
17 million lines of code? What the fuck is it doing, starting from Peano axioms? Proof by 2 million cases? This is bullshit. The timing says it all. Panic move after the cratering stock.
>>
>>17099675
>>17099344
>>
>>17099778
>no attempts to actually understand and vet the proof
I accept your full concession.
>>
>>17099757
>What the fuck is it doing, starting from Peano axioms
Basically, yes, the proof starts from the basic Lean axioms and writes everything else.
https://math.mit.edu/~drew/grh/audit.log
>>
>>17099344
>Meanwhile, let's see what a professional mathematician at MIT has to say, unlike (You), who is an uneducated dolt. https://math.mit.edu/~drew/grh/
retard. he confirms it passes lean. this wasn't in dispute. nobody accused levent alpöge of lying about it passing lean lmao.
>>
>>17099840
if it wasn't in dispute, why check is so many times =]
>>
>>17099840
Sutherland:
>At this point I don't see any realistic scenario in which this proof does not imply the statement of RH in mathlib
and that was even before half those checks were done.
>>
>>17099841
it's right there in the OP. the doubt is about HOW it passed lean, not WHETHER it passed lean. everybody agrees it passed lean. the question is whether it did so because it's a correct proof or it did so because it's an incorrect proof using bugs in lean to fool lean. that mit link just confirms it passes lean, which isn't in dispute.
>>
>>17099857
Yes, and
>Sutherland: I personally have no doubts at this point
>>
>>17099857
There are now several independent implementations of lean some of which are themselves formally checked for correctness w.r.t. the set theory axioms. At least that's what I'm hearing.
>>
>>17099866
Nobody cares. Erdos was initially adamant that switching in Monty Hall was 50/50.
It's too early. This will require years of work to verify rigorously.
>>
>>17099891
>There are now several independent implementations of lean
Irrelevant.

>some of which are themselves formally checked for correctness
By Lean itself. Your retarded scam won't fly here.
>>
>Andrew V. Sutherland
>(he/him/his)
I'm sold. I have absolute faith in the discernment of Sutherland.
>>
File: IMG_3803.jpg (51 KB, 474x619)
51 KB JPG
>>17099047
Someone just needs to ask the AI to generate a negation is all.
>>
>>17099912
We don't have access to the frontier model used, but yes, if it's able to exploit bugs in Lean, it should be able to crank out other proofs, maybe even contradictory proofs. That would be a good avenue to explore.
>>
>>17099912
>>17099928
its the biggest con in stock market history



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.