When Wang Hong and Deng Yu received the Fields Medal—the highest honor in mathematics—it felt like a pure, unadulterated triumph of human intellect. It was a beautiful, thrilling moment. But quietly, across the mathematical community, a single phrase is being whispered: this might be the last Fields Medal without AI involvement.
It’s not a joke. Recently, an internal OpenAI model disproved the Erdős unit distance conjecture—a problem that had stymied top mathematicians for decades. A standard model did it. A problem a Fields Medalist might have agonized over for years, an AI just stumbled upon.
We thought AI would become our calculator. Instead, it has become our rival, and we don’t even know we’ve already lost the game.
But here is the paradox keeping every mathematician awake at night: the exact same AI that can casually solve a Fields-level problem cannot reliably complete a standard PhD thesis. Its upper limit is astonishingly high; its lower limit is embarrassingly low. It gets stuck in childish loops, negates its own correct reasoning over a minor hiccup, or endlessly restates a theorem without progressing.
This extreme variance in capability is giving birth to a new, deeply unsettling research paradigm. It’s called ‘Gacha Mathematics.’
In traditional human-AI collaboration, the human understands the problem, proposes a strategy, and the AI executes. In ‘Gacha Mathematics,’ you feed a problem you don’t even fully understand yourself into the AI, let it run autonomously, and simply check the output. You actively ignore the answers you can understand—because if the AI produces something you comprehend, it’s probably useless. You specifically hunt for the completely alien, cross-disciplinary answers, learning the underlying tools only after the fact.
The modern mathematician no longer creates knowledge; they are just pulling for SSR characters in a gacha game, praying the AI accidentally stumbles upon a mechanism they could never find themselves.
It sounds like fringe pseudoscience, doesn’t it? It essentially is. But it is pseudoscience with infinite patience and zero cross-disciplinary learning costs. The AI’s breadth and trial-and-error volume crush human capacity. It just keeps generating until it hits a key that suddenly unlocks the entire puzzle. And once that mechanism is found, the rest of the problem—even Millennium Prize problems—often shifts from ‘impossible for generations’ to ‘systematically doable.’
We have all comforted ourselves with a soothing lie: AI does the heavy lifting, humans do the thinking. ‘Humans ask the questions and have the aesthetic taste; AI does the proving.’ It’s what we tell ourselves to maintain our dignity.
It’s wrong. Asking a good question is often just a single-step judgment—spotting an analogy, noticing an anomaly. Proving it requires hundreds of sequential, correct steps. AI is learning the former. It has already begun forming its own cross-disciplinary ‘taste,’ making connections between fields that human experts missed because they were trapped in their own narrow specializations.
The comforting division of labor where humans think and AI works is a lie. Having taste, it turns out, is easier to encode than execution.
This forces a severe identity crisis on anyone who trades in intellectual labor. The mathematical world has always operated on an ‘IQ logic’ or a ‘persistence logic.’ You made something others couldn’t, proving you were smarter or more dedicated. The Fields Medal is the pinnacle of this logic.
What happens when a machine can ‘pull’ the breakthrough you spent a decade grinding for? What happens when the AI can’t even explain how it found the answer, and your intellect, your persistence, and your taste suddenly feel useless?
When a machine can accidentally stumble upon the breakthrough you spent a decade grinding for, your ‘genius’ is nothing but a vestigial remnant of compute power.
Mathematics isn’t going to die. But the pure mathematician will shrink. The researcher who grinds on a single problem for a decade will be crushed, unable to justify the risk of a decade-long bet in a world where an amateur might pull the winning ticket via AI. Research will become much more like applied math: rapid testing, simultaneous focus on many questions, and a massive reduction in long-term bets. The era of the solitary genius is ending; the era of the AI system tester has begun.
FAQ
Q: Isn't AI just doing probabilistic guessing without any real understanding?
A: Humans also rely on pattern matching and statistical recall. The difference isn't a magical human 'understanding,' but the AI's current inability to execute long-range planning. It can make brilliant single-step leaps but fails at hundred-step execution.
Q: If I'm a researcher, what should I actually do right now?
A: Stop trying to prove you are smarter than the machine. Shift your focus to defining important problems and building explicit dependency structures (like qmd formats) so AI can actually verify its own work. Become a system tester, not just a creator.
Q: Is AI really better at asking good questions than humans?
A: Yes, in cross-disciplinary contexts. Humans are blinded by specialization. AI reads everything and can spot a connection between algebraic geometry and number theory that a specialist would never see because they don't have the time to learn both fields deeply.