Truman Dickerson
- Mathematicians at Caltech are taking a new approach to the existential crisis facing math.
- They’re making their own AI models — by mathematicians, for mathematicians.
- Their approach offsets the current balance of power, which is dominated by OpenAI.
Sergei Gukov said he could see the writing on the wall.
So when OpenAI in October solved hundreds of famously thorny math questions that had stumped the field’s best and brightest for years, the Caltech professor wasn’t surprised.
In fact, his AI models had solved a few of them first, he said.
“The more we work with the systems, the more we understand them,” Gukov told Business Insider of AI. “As a result, the better we can control them.”
Gukov runs a lab at Caltech where he and a handful of students design and build their own open-source AI models from scratch. The models are specifically engineered to solve math problems, he said over sushi at a Japanese restaurant in Pasadena.
The approach is turning heads in the math community at a time when antipathy toward major AI companies is high, as the giants rapidly publish solutions without running them by academic luminaries.
Mathematicians writ large have torn into OpenAI’s way of going about it in Substack essays and lecture halls across the country. Put simply: They say that the companies, OpenAI chief among them, are brazenly exploiting their field simply to drum up hype for an IPO.
And of the actual solutions proposed by OpenAI, many say they’re slop — completely meaningless unless a human researcher spends weeks deciphering how they were reached.
If mathematicians build the models themselves, however, that puts the ownership of AI-generated solutions back into their hands.
Tasmin Chu, a Ph.D. math student at Caltech who runs a popular anti-AI Substack, called mathematician-created AI systems an effort by mathematicians to “own the means of production.”
“It’s complex,” she told Business Insider. “I guess it depends on how it really plays out.”
Though farsighted, Gukov’s models are not without their shortcomings. He likens them to Ferraris, luxury cars that, for all their speed and craftsmanship, are largely inaccessible to the general public.
For instance, most of his models can only solve one specific math problem. If you asked one to branch out into a different area, it would sputter.
By comparison, Anthropic models can handle math prompts across a variety of disciplines. Gukov said a more generalized model is in the works, but isn’t ready yet.
Even so, the notion of mathematician-created AI models offers hope to a community that was left divided in the wake of OpenAI’s solution dump.
On one side is a stalwart group diametrically opposed to all forms of AI in their work. Chu leans toward this group. “I’m personally invested in it remaining viable to do mathematics without AI at all,” she said.
On the other side are students going all in on AI, whether because they believe doing so will save the field, or because they don’t believe the field is worth saving.
Terence Tao, a math phenom based at UCLA, has repeatedly advocated that mathematicians create their own AI systems.
At Caltech on Friday, he demoed a platform created by a nonprofit he founded. The Open Math OS, as the makers call it, provides access to open-source frontier AI models, cloud computing, and collaborative tools.
It also, crucially, allows mathematicians to control how their data is used — an issue that has been important ever since two mathematicians accused OpenAI of using their chat logs to help solve the Navier-Stokes problem.
Truman Dickerson
For Gukov, who was born in Moscow and immigrated to the United States to study math, the existential dilemma of AI in math is not so much a crisis, but an opportunity for growth.
“This AI research aspect is super exciting,” he said. “You feel like Columbus exploring new territory.”
Â