The world of mathematics is in a state of flux, with **artificial intelligence (AI)** rapidly emerging as a formidable player. In mid-May, a *top-secret gathering* of elite mathematicians took place in **Berkeley, California**, to confront this advancing technology. Their mission? To test the capabilities of a sophisticated AI, **o4-mini**, designed to solve complex mathematical problems. What they discovered left them both **astounded** and somewhat apprehensive.
The Gathering: A Clash of Minds
Thirty of the most celebrated mathematicians from around the globe, including some who **traveled from the U.K.**, assembled for what was to become an unforgettable showdown. For two days, they unleashed a barrage of mathematical questions on o4-mini, a **reasoning chatbot** powered by cutting-edge AI from OpenAI. Ken Ono, a prominent mathematician from the University of Virginia and a key figure at this meeting, remarked, “I have colleagues who literally said these models are approaching **mathematical genius**.”
Understanding o4-mini
o4-mini belongs to a new generation of **large language models (LLMs)** that have received specialized training. Unlike their predecessors, these models—like Google’s **Gemini 2.5 Flash**—are more nimble and trained on tailored datasets, honing their **deductive reasoning** skills. Earlier models struggled with intricate math, often achieving less than **2 percent success** on previously unsolved problems. In stark contrast, o4-mini displayed surprising strengths in mathematical reasoning.
The Challenge Begins
OpenAI had previously commissioned **Epoch AI**, a nonprofit focused on benchmarking LLMs, to curate an array of challenging math questions—300 in total—unanswered in the academic discourse. The meeting was the culmination of hard-fought efforts to uncover deeper mathematical mysteries.
A Test of Wits
With modest progress being made, **Epoch AI** organized an in-person gathering on **May 17-18** to accelerate the final round of question development. The mathematicians were split into small groups, diving headfirst into solving problems they believed would stump the AI.
A Surprising Turn of Events
By Saturday night, tensions were building. Ken Ono devised what he thought was an insurmountable challenge—one recognized among his peers as an open question in number theory. When he put the problem to o4-mini, his expectations were thwarted. The chatbot not only processed the question but also revealed its reasoning step-by-step.
The Moment of Truth
In a turn of events that was both exhilarating and daunting, o4-mini dissected the literature and presented a **correct solution**, cheekily stating, “No citation necessary because the mystery number was computed by **me!**” Ono admitted, “I was not prepared to be contending with an LLM like this. I’ve never seen that kind of reasoning before in models. That’s what a scientist does. That’s frightening.”
Lessons Learned and Future Implications
Despite surrendering to the AI in this instance, the mathematicians managed to **uncover ten questions** that stumped o4-mini. They were, however, left pondering the implications of this technology. Ono noted that the bot performed like a “**strong collaborator**,” and even Yang Hui He from the London Institute for Mathematical Sciences concurred, stating, “This is what a very, very good graduate student would be doing—in fact, more.”
Speed vs. Depth: A Double-Edged Sword
o4-mini also demonstrated astonishing speed, solving problems in minutes that could take human professionals weeks or even months. However, this efficiency came with its own set of worries. Both Ono and He raised concerns over **over-reliance** on AI solutions, fearing that its authority could lead to misplaced trust in potentially flawed results.
Looking Ahead: The Changing Role of Mathematicians
As discussions evolved, the prospect of a **“tier five”**—a new realm of mathematical challenges—came to light. As AI continues to advance in sophistication, the role of mathematicians may shift significantly. Instead of solely solving problems, they may increasingly focus on engaging with AI to unveil new mathematical truths, akin to guiding graduate students.
Final Thoughts
Ono emphasized the need for creativity and adaptability within the field. “I’ve been telling my colleagues that it’s a grave mistake to say that generalized artificial intelligence will never come… these large language models are already outperforming most of our best graduate students in the world.” As mathematicians grapple with these rapid advancements, their challenge now is to find a way to thrive in an era where AI is not just a tool, but potentially a **collaborator**.
For a deeper dive into this evolving narrative, you can explore the importance of AI in **[mathematics](https://www.scientificamerican.com/article/how-does-chatgpt-think-psychology-and-neuroscience-crack-open-ai-large/)** and its implications for our understanding of complex problems.
Support science journalism: If you appreciate this kind of insight into technological advancements, consider **[subscribing](https://www.scientificamerican.com/getsciam/)** to support impactful stories that shape our understanding of the world today.
