Mathematicians are getting scooped up. The richest people on the planet want them.
University departments are seeing colleagues vanish one by one. Off they go to private firms. Some names you know—OpenAI, Google. Others? New. Barely months old. Betting everything on a singular idea. That mathematics is the missing secret sauce for smarter artificial intelligence. And in turn, the AI might change how we do math itself.
Ken Ono admits he felt a loss. “Last May, I was honestly kind grieving for my scientific identity.” In 2025, he left his professorship at the University of Virginia. He joined Axiom Math, a startup dedicated to building a math-brained AI.
He didn’t jump ship alone. He’d been helping a company called Epoch AI. The goal? Create impossible math problems to test AI limits. The result shocked him. The AI wasn’t just struggling. It was succeeding. Way beyond his expectations.
After a few months of that, he realised: maybe this is the moment where the sharecropper sees the combustion engine in the field. He thought: we can do more.
Ono wasn’t alone. The last two years have birthed a wave of companies. Axiom Math. Harmonic. They don’t just want AI to do math. They want it to prove it. Correctly. Verifiably.
I visited them in Silicon Valley this past April. To understand why they placed so much faith in equations. To see why investors poured hundreds of millions into nondescript office buildings in Palo Alto.
Axiom Math sits near Stanford. Founder Carina Hong was once Ono’s student. Next door? Harmonic. Also chasing “mathematical super intelligence.” The offices look boring. Named rooms after Gauss, Lovelace. Standard startup chic.
But the inside story is different.
Why not just use ChatGPT? You ask Ono. Why start a new company when OpenAI exists?
“ChatGPT is the librarian,” Ono says. “You can’t find what it hasn’t read. But do you want your librarian performing surgery?”
Large language models hallucinate. They bluff. You can’t trust them for truth without human checking. That’s the gap. That’s the business opportunity. Verification.
Too much to check
This isn’t new in academia. Mathematicians use Lean. A programming language. You feed it handwritten proofs. The computer checks if the logic holds. Instant. Fast.
In research, verification takes forever. Researchers are stretched thin. They spend weeks checking steps they should trust. Lean fixes that.
Now, coding has the same problem. AI writes code. Lots of it. But it contains subtle errors. Human programmers spend all their time babysitting the AI’s output. Looking for bugs.
Axiom Math and Harmonic are targeting this pain point. Math problems have low prize money. Software verification has massive commercial value. If AI writes all the code, humans become the bottleneck for safety.
“As AI writes more code,” says Tudor Achim, CEO of Harmonic, “the value of verification goes up.”
These tools work. Axiom Math has five papers accepted in journals. Written entirely by their AI. Topics? Algebraic geometry. Number theory. Heavy stuff.
Ono won’t share the exact roadmap. But the goal is dozens of papers next year. Years of work compressed into weeks.
Competition is fierce. Tech giants are watching. OpenAI knows math is measurable. “Math is wonderful for AI,” says chief scientist Jakub Pachocki, “because it’s quantifiable.”
Early LMs sucked at it. They failed at basic arguments.
Not anymore.
Recent models won gold at the International Mathematical Olympiad. A contest for elite high schoolers. Then, an AI disproved an 80-year-old conjectute. Mathematicians thought progress would never come in their lifetimes.
“The weaknesses we saw six months ago… are gone,” says Sébastien Bube at OpenAI. No more nonsense.
But approaches differ.
Axiom and Harmonic hire mathematicians to train specialized models. They force the AI to think like a mathematician. OpenAI does not.
“We are training for general intelligence,” Bubeck insists. Math capability is a byproduct. A shock, he says. Not a primary target.
Who wins? Unclear. But one thing is certain: a handful of companies now hold the keys.
Mathematicians are uneasy.
Paywalled theorems
The money came fast. What happens when it leaves?
Ravi Vakil, at Stanford, worries about the hype cycle. “In five years,” he says, “it won’t be like this.” No one gets rich solving the Riemann Hypothesis. The money moves on. The models stall.
Or worse. Math becomes a gated community.
What if solving problems requires paying a fee? What if truth is behind a login?
Shubho Senguptat at Axiom Math says some math is already paywalled. Hedge funds build models. They keep them secret. Intellectual property.
But “pushing the bounds of knowledge… should be free.”
Achim at Harmonic disagrees slightly. Tools cost money.
“We want people to pay for service,” Achim says. He insists companies support mathematicians. No extraction of all value. Just sustainable business.
Who knows what happens. Future prediction is notoriously bad for humans, let alone machines.
For now, mathematicians remain the guides. The drivers.
Ono compared the rise of these AI systems to Srinivasa Ramanu.
You remember the story. The self-taught genius from India. Intuition so pure it looked like magic to 20th-century scholars. His ideas came from nowhere. Shocking. Beautiful.
Ono’s father died in January. He was a Japanese mathematician moved by Ramanu’s myth.
One of their last talks stays with Ono. His father’s advice was simple.
Maybe this is your Ramanujan moment. Others won’t get it. People will fear the box. If a computer shows you something magical… you have to embrace it.
“It already happened to all of us,” he said.
The engines are running. We just don’t know where they lead yet.
