MIT professor in despair: OpenAI cracks the millennium problem, and humans can’t comprehend it.
MIT Professor Despair: The Millennial Problem Solved by OpenAI Is Beyond Human Comprehension!
Summary
Xinzhiyuan reports Recently, the global mathematics community has been rocked by AI. OpenAI released 722 papers in one go. While these papers are still subject to verification, it appears that from the unique game conjecture to the quasi‑Riemann hypothesis, many well‑known …
Xinzhiyuan Report
Xinzhiyuan Report
</h3
Recently, the global mathematics community has been blown up by AI.
OpenAI has published 722 papers in one go.
Although these papers are still subject to verification, it appears that many famous unsolved mathematical problems—from the unique game conjecture to the quasi‑Riemann hypothesis—have been solved.
A mathematician wakes up to find that their research may be finished!
Hungarian probabilist Alfred Rényi once called mathematicians “machines that turn coffee into theorems.” Now there really is such a machine, and it runs on a data center.
In the latest episode of Bloomberg’s well‑known podcast “Odd Lots,” MIT’s Deputy Dean of Engineering Education, Justin Solomon, tackles a tough question:
How AI will disrupt the world of mathematics.
AI directly reduces the absolute pinnacle of human intellect—“mathematics”—to a lower dimension.
This autumn, human intellectual pride was shattered by “alien logic.”
When OpenAI solved the millennium prize problem that has puzzled humanity for a hundred years, the result was not a global celebration but a massive, bottomless cognitive panic.
Because even the world’s top mathematicians and MIT professors, upon reading the second page of the proof, could only lower their heads in despair and reluctantly admit:“It’s correct, but I can’t understand it at all.”
When “truth” and “human understanding” diverge for the first time, have we evolved to the threshold of divinity, or have we completely become AI‑reared “intellectual illiterates”?
The story began last month, when OpenAI’s model presented a Millennium Prize problem—a counterexample proving the failure of solutions to the Navier‑Stokes equations.
The Navier‑Stokes equations are known as the “God equation” of fluid dynamics.
When you stir your coffee with a spoon in the morning, the swirling foam; the air currents tearing the sky as a typhoon’s eye forms; even the drag generated when an aircraft wing slices through the air… in theory, all of these are governed by this elegant yet profound equation.
Writing down this equation isn’t difficult, but for the past century the world’s brightest minds have been stumped by one fundamental question:If we start with a smooth initial state, can the solution to this equation suddenly “blow up” at some moment in the future?
In plain language: could a fluid, under some extreme conditions, suddenly defy physical intuition and erupt with infinite speed? To find the answer, the Clay Mathematics Institute offered a $1 million prize.
Historically, mathematicians have generally believed “no.”
But OpenAI has produced a counterexample:there exists an extreme stirring method akin to a “crazy spoon” that, in theory, can drive the fluid into complete loss of control within finite time.
More dramatically, this cross‑century breakthrough was not created entirely from thin air by AI.
Previously, a group of top mathematicians in Spain had already built the main structure of this “spoon” with pen and paper, pushing the progress bar to 99%. AI, relying only on its terrifying computational power and inhuman logical reasoning, completed the final—and most desperate—1% for humanity.
Logically, shouldn’t we be popping champagne now?
But the most fatal dilemma has arisen:This epoch‑shattering proof, once released, no one in humanity can get through it.
The host of “Odd Lots,” Joe, downloaded the paper and looked at those symbols as if reading a celestial script.
Meanwhile, MIT associate dean Justin Solomon—who earlier at Pixar studied the complex mathematical models behind visual effects—when trying to read this proof,after flipping through the first two pages, his brain completely crashed.
“That proof reads like a nightmare—filled with dense technical details, wildly counter‑intuitive inequalities, and extreme boundary values,” Solomon admitted.
The host asked, “Is the proof elegant?” Solomon gave a wry smile, “Probably not elegant at all.”
The host followed up, “Is it correct?” Solomon paused, “Probably yes.”
For thousands of years, from Euclid’s geometry to Newton’s calculus, in the world of mathematics，“It is correct” and “humans understand it” have become almost synonymous。
A brilliant proof itself carries the brilliance of human rationality and the poetry of logic.
But today, these two things have been torn apart by AI. Truth has arrived, but it speaks an alien language.
The podcast mentioned a crazy move by Uber founder Travis Kalanick: the Silicon Valley mogul has recently become obsessed with studying physics together with AI.
He thinks that just by chatting with AI, relying on a so‑called “Vibe Physics,” he has already approached some major breakthroughs in physics.
But “can intuition from just chatting with a machine really grasp complex physics?”
If mathematical and physical breakthroughs seem far removed from ordinary people, then one of today’s greatest mathematicians, Fields Medalist Terence Tao, delivers a soul‑piercing thought experiment that brings this anxiety directly to everyone.
Just a few days ago, he gave a talk at Caltech’s “Math 2.0” event.
Slide link: https://teorth.github.io/tao-web/slides/math-2-0-caltech-2026.pdf
A slide by Terence Tao sparked controversy across the internet.
He asks everyone to imagine the following scenario:
Humanity issues a prompt to a super‑AI with near‑infinite computing power—“Find a cancer treatment that can pass Phase III clinical trials.”
After massive black‑box computation, the AI actually spits out a formula. It is a complex mixture of several unknown chemical components. The AI’s mathematical model precisely predicts: as long as this concoction is prepared and infused into a patient’s vein, it will eliminate 100% of cancer cells.
Even more astonishing, although no medical researcher or chemist in humanity knows why this formula works, the AI’s reasoning process not only received flawless formal verification in the interactive theorem prover Lean, but also, in the real world, this concoction has brutally passed an extremely rigorous FDA Phase III clinical trial.
Now the choice is in your hands: if this “miracle drug” that even top scientists cannot decipher is placed before you, do you dare to inject it into your own veins?
Tao Zhexuan sharply puts the core dilemmas of cutting‑edge AI in “alignment” and “interpretability” on the table.
In this thought experiment, could this AI solution have a fatal flaw? Might it simply exploit loopholes in the human “Phase‑III clinical trial” rules, achieving the literal goal through an “alien logic” that humans cannot comprehend at all?
Would you shout “stop” before the needle pierces the skin and insist that: at least one human oncology expert must fully understand its cure mechanism, or at least one human mathematician must completely grasp the mathematical model behind the AI’s formulation!
Perhaps late‑stage cancer patients don’t care about the underlying mechanism, but the question is: how can efficacy be determined?
This is precisely the terrifying future revealed by OpenAI solving the Navier‑Stokes equations: The frontier AI’s mathematical deduction capability is already approaching, even surpassing, the limits of human cognition. The traditional research paradigm of “human‑validating‑AI” is hitting a hard physical wall.
More shocking—and more unsettling—than conquering a millennial problem is the AI‑generated proofs filled with counter‑intuitive leaps and a complete lack of aesthetic sense, what we might call “alien logic.”
When AI’s productivity begins to surge exponentially while human verification capabilities remain stagnant, the first system to collapse is the academic review process we have proudly upheld for centuries.
In the podcast, Solomon reported a staggering set of numbers: the top machine‑learning conference ICLR received only one or two thousand submissions ten years ago; but this year, submissions surged like a tsunami to sixty thousand!
Reviewers need to digest these AI‑generated papers within just a month or two.
“I don’t even have time to open the PDF,” Solomon says with a wry smile.
Human elite computational power is exhausted, peer review has become a mere formality, and the seal of academic quality has lost its effectiveness.
Even a high school student, armed with an AI‑crafted flashy résumé, managed to slip into the reviewer pool of top‑tier conferences and evaluate papers authored by university professors.
What troubles this MIT associate dean even more is that his inbox has become a hotspot.
Every day, strangers email him: “Professor, I think I’ve proved a new result, AI says this derivation is important—could you verify it for me?”
Models are inherently inclined to “go along with what people say,” which is disastrous in a field that demands absolute rigor like mathematics. The lofty barriers of mathematics were originally meant to block pseudo‑science that “looks right but is actually nonsense.” Now AI dresses pseudo‑science in the guise of truth, launching a massive DDoS‑style assault on humanity.
What follows is an even harsher class solidification: doing research now means competing with wallets and computing power.
If this trend continues, future mathematics prizes and Nobel Prizes will have their rules completely rewritten: “Whoever buys the most tokens, whoever has the strongest GPU compute power, will be the first to find the answer.”
We can do a quick calculation:
A typical university mathematics professor painstakingly derived a prototype of a direction on paper, before even publishing it for verification; those tech giants that “control the compute faucet” (like OpenAI, Google), using internal unpublished models at least two generations ahead of university labs, followed even the slightest hint leaked online by the professor, and within a week forced the final result with the compute power of tens of thousands of GPUs.
In fact, Solomon delivered a killer blow: “Maybe the source of inspiration for their model is that they’ve free‑riden on your usual online academic discussion records.”
So the question arises:Who truly deserves credit for this century‑spanning breakthrough?
The $1 million prize from the Clay Mathematics Institute—does it go to the tech‑giant CEO who provides computing power? To the programmer who writes the prompt “Please prove the Navier‑Stokes equations”? Or to the human mathematician who laid the groundwork?
No one knows. This messy accounting has become a headless cold case of the cyber age.
Since AI can even tackle the Millennium Problems, programmer positions are visibly shrinking (the first decline in computer science enrollment at MIT in twenty years), and even MIT professors can’t understand the proofs, what can humans actually do in this AI‑bombarded era?
Solomon’s assessment is half a cold splash, half a door left open.
He believes that all the groundbreaking feats the models achieve today essentially remain within the “Convex Hull” formed by existing human knowledge.
AI excels at drawing high‑dimensional connections between known knowledge nodes, and can even, through exhaustive brute force, compute the final steps that have stumped humans for a decade.
But truly great new theories are like building a skyscraper on flat ground from nothing, creating an entirely new concept out of thin air, and redrawing the world’s picture.
Humans learned to brew wine thousands of years ago, enjoying cheerful drinks and a gentle buzz. Yet it was not until modern times, millennia later, that we understood the mechanism of “yeast” under a microscope.
Perhaps future mathematics and science will follow this path: AI, like nature, will be responsible for generating cheap yet effective “answers”; humans will spend decades or even centuries figuring out the “why”.
Generating answers is getting cheaper, but “verification” and “understanding” have not followed suit; instead, they have become the most scarce luxury.
“So, do our kids still need to learn math?” Joe asked the practical question that concerns every parent the most.
This time, MIT’s vice dean didn’t stammer; he answered with confidence:“I think it’s essential.”
Because when “answers” become cheap and the world awaits the next leap, who can pose the next Navier‑Stokes equation? Who can judge whether an AI‑generated answer truly carries weight? Who can turn that pile of cold code into “understanding” that human civilization can digest and pass on?
For now, these things are still on humanity’s side.
References:
Edited by: David
Get MirrorSpeed