From our other sites
The story
On October 6, 2026, OpenAI used an AI model that has not been released to the public to tackle 372 unsolved problems in mathematics, publishing the results all at once in 722 papers. The results include progress toward the Riemann Hypothesis, which has stood unsolved for more than 150 years, as well as the resolution of a hard problem dating back to the 1940s. The announcement was framed as AI solving major math problems almost as a “side effect of a performance test,” and it drew plenty of reaction on 5ch’s Science News+ board. Alongside amazement at the AI’s raw ability, commenters on the thread were notably concerned that the results were racing ahead without any solid system in place to verify the proofs.
Hard Math Problems Were Being Solved as a “Side Effect” of AI Performance Testing — 722 Papers Released at Once
On October 6, 2026, America’s OpenAI used an AI model not available to the general public to announce results on 372 previously unsolved hard problems in mathematics, presented across 722 manuscripts.
Among them are results that bring us significantly closer to the Riemann Hypothesis, which no one has been able to crack in over 150 years, as well as hard problems that had remained unsolved since the 1940s.
Source: nazology.kusuguru.co.jp / Original article here
What people said
Simple as that.
It's one of those things where your head gets it but your gut just won't accept it.
The mathematicians understand that.
and crowned itself king of the earth.
Humanity is now a thing of the past.
From here on, we can only live as slaves.
Meanwhile humans toil away, sweating to serve Lord AI.
That's because you're using the free chat version.
Even that has gotten a lot better about saying weird stuff these past few months, though.
Nobody uses AI these days the old way — ask a question, get one answer, done.
and then Chappy (a common nickname for ChatGPT in Japan) comes along going "oops, did I do it again?"
and drops a bombshell result — of course the math world is gonna go all wobbly lol
Apparently there's a decent number of counterexample patterns in there,
so that alone should be useful as-is.
Scientists: "We're not your quality inspectors!"
Even Perelman — his paper was brilliant, but still…
even if OpenAI had read the room and not used math problems as a benchmark substitute,
the moment the latest model got released, random math-illiterate users would end up going
"uh, I think I just solved some unsolved problem called XYZ? lol" anyway — that's probably how it'd play out.
That said, from the math community's side, I get that something real gets lost when the field gets trampled like this.
This is a genuinely tough problem.
> Doubts raised over OpenAI's 719 AI-generated math papers — Navier-Stokes proof said to conflict with formal verification
yellow.com/ja
> Questions are mounting in the math community over just how far the 719 AI-generated math papers from OpenAI can be trusted.
> The trigger was a new paper pointing out that, for the Navier-Stokes equation proof the company announced earlier,
>
> "the body of the paper doesn't match the Lean code used for formal verification."
If "the body of the paper doesn't match the Lean code used for formal verification," then it's not actually a proof.
Either OpenAI is being sloppy and reckless, or it simply hasn't reached the bar needed to be published as a proof in the first place.
Same deal this time — something like "nearly 60% of the proof manuscripts are already public, but the work of rewriting them into Lean form is still in progress," meaning they haven't even verified it, right?
Putting out something unverified and calling it a proof in a paper is wrong to begin with.
That's rock bottom, OpenAI. Is AI even usable in academia at this level?
Is it nothing more than a glorified, sloppy office tool after all?
Probably before too long, AI will be generating both the Lean verification and the human-readable papers,
start to finish. At this pace, maybe even within the month?
Last month the bar was still something like "AI poured in huge amounts of time and resources (i.e. big money) to solve the problem,"
but this time it's solving problems after thinking for about 3 hours each —
we're already past the point where it's even comparable to human mathematicians.
The next-gen model will probably go public sometime before the year's out,
and once it does, you can just try solving stuff with it yourself. For about $100 a month in subscription fees,
an AI will answer your questions endlessly, for as long as it takes you to get bored.
"Last month," "this month" —
what era do you even think this story is from?
The sequence is: in September there was the story about solving the Navier-Stokes equation,
and then the next generation of AI after that solved this whole batch of problems at once, right?
What era do YOU think this is, exactly?
I did write "there was the story that it was solved," though?
There really was a flood of news like that at the time.
We're talking about a timeline here, and the moment you say "you don't understand the content!"
you're the one who doesn't understand the flow of this conversation.
I bet you're just trying to muddy the original point.
If it's being credited to OpenAI's next-gen AI, there's no way it could be
some old story from way back — that's the whole point.
Re: #23 was probably just mistaken, plain and simple.
if that so-called treasure couldn't crack the problem for decades anyway,
it's really just sour grapes from the losing side.
It literally says to stop just dumping results without showing the process.
It's just promotional material insisting it's a proof.
With humans you have peer review to debug it,
but with AI and Lean it's all self-reported, so you can't tell if there are bugs.
Humans will handle it from there.
If it finds the path, humans can walk it.
but none of it is backed up and it all smells fishy.
It's possible to define, in Lean, the mathematical structures of Newtonian mechanics and quantum mechanics broadly speaking (Hilbert spaces, CPTP maps, existence of solutions to differential equations, etc.).
However, this only reproduces 100% consistency in the mathematical structure of the physical laws — it doesn't reproduce the actual fluctuations of individual real-world physical phenomena themselves.
The moment you write that the Navier-Stokes equation was "solved," it shows you don't get it, so maybe just stay quiet.
It's said that Fermat mistakenly believed he'd proven it using the infinite descent method he himself favored.
In reality it's not something provable by infinite descent, and the proof wasn't actually completed until the 20th century.
does AI actually "think" when it answers, the way a human does?
If it's just regurgitating sketchy old information as-is,
I feel like that'd be pointless.
The definition probably varies by person, but yeah, it's thinking.
Otherwise it couldn't solve Math Olympiad-level problems.
At the very least, it's definitely not just rolling dice and sloppily copy-pasting stuff.
To answer that, you'd first have to figure out what "thinking" even is.
AI is a knowledge database, so it's not thinking — it's just blindly trying, one by one, proof techniques that worked for humans in the past, and presenting whichever result looks good as the answer.
That's why, afterward, you have to properly verify whether it's actually a correct proof.
Since AI can't explain how it arrived at a result, rewriting it in Lean and verifying it is essential in AI's case.
And even if the Lean check turns up no problems, humans still have to verify whether that Lean code itself is correct.
The advantage of AI-generated proofs is that they can apply past proof techniques humans might never have thought to use.
But don't expect much from proofs built on genuinely new ideas unlike anything before — AI is, after all, just an aggregate of humanity's past knowledge.
it should be able to perfectly translate its proofs into normal prose a human can read.
The fact that it can't be translated into a human-style paper means it isn't thinking anything through,
which makes me doubt whether it even has the same logical consistency as real math to begin with.
Aren't you overrating humans?
Plenty of humans can't do that properly either.
Poor Fermat, getting talked about over and over with people acting like they've got him all figured out.
If he really knew the truth, he's probably thinking "I'll flatten you for that, you little punks."
For his other results, Fermat similarly scribbled about having a "marvelous proof," and in those other cases it turned out they really could be proven by infinite descent elsewhere. So the theory goes that when he jotted the note about what we now call his Last Theorem, he believed at the time it could be proven by infinite descent too — but when he later tried to actually carry it out, infinite descent just didn't work for that one.
Don't just make up guesses like that.
Like I'd know.
This is simply what's currently believed to be Fermat's "marvelous proof."
If you don't like it, go write a paper that overturns it.
Background and Key Issues
What stands out about OpenAI’s announcement is that it rolled out results on unsolved math problems all at once, across 722 papers, not as peer-reviewed “research” but as a byproduct of performance-testing an unreleased model — and that departure from the math world’s normal process is what triggered the backlash. “Lean,” which comes up repeatedly in the thread, is a programming language and tool used to automatically verify by computer whether a proof is logically sound; whether an AI-generated proof text actually matches its corresponding Lean verification code has become the yardstick for how much it can be trusted. In fact, separate reports have pointed out that for the previously announced proof of the Navier-Stokes equations, the body of the paper and the Lean code didn’t match — and much of the thread’s skepticism traces back to that incident. In other words, the real issue isn’t whether AI can solve these problems at all, but how claimed solutions get verified and who’s responsible for that process.
*This article is excerpted and summarized from the 5ch (Science News+) thread “yƒiƒ]ƒƒW[z”Šw‚Ì“ï–â‚ÍAI‚Ìu«”\ŽŽŒ±‚̂‚¢‚Åv‚É‰ð‚©‚ê‚Ä‚¢‚½„Ÿ„Ÿ722–{‚̘_•¶‚ªˆê‹“ŒöŠJ [‚·‚ç‚¢‚Þš].”
Leave a Reply