AI solves 20-year math challenge that researcher thought machines could not crack

Researchers tested the problem with GPT-5.4. (Image credit: Alamy)

• AI solved a 20-year-old math problem designed by Naskręcki
• GPT-5.4 succeeded on its eleventh attempt, surprising experts
• AI solutions still require human verification for accuracy

Did our AI summary help?

A research-level mathematics problem that took nearly twenty years to design has been solved by an artificial intelligence system, surprising the mathematician who created it.

The problem was developed by Bartosz Naskręcki, a mathematician at Adam Mickiewicz University in Poznań, Poland. He designed the challenge as part of FrontierMath, a benchmark used by researchers to test how well AI systems can handle extremely difficult mathematical reasoning.

Naskręcki had long believed that artificial intelligence could not handle such problems. In earlier remarks, he described AI as little more than a “very advanced calculator” capable of performing calculations but lacking the deeper understanding needed for genuine mathematical insight.

The challenge itself was not a short puzzle. Its solution required roughly thirteen pages of detailed mathematical reasoning and drew on advanced fields such as number theory, combinatorics and algebraic geometry. Even experienced mathematicians might take weeks to identify a workable approach.

Recent testing changed that assumption.

Researchers ran the problem through GPT-5.4, a new generation AI model designed for advanced reasoning tasks. The system attempted the problem eleven separate times. Ten attempts failed, but on the eleventh run the model produced a correct solution.

While the success rate was modest, the result still surprised the mathematician who had spent years refining the challenge.

After reviewing the output, Naskręcki wrote online that he was “deeply impressed” and described the solution as “very nice, clean, and almost human.”

The episode reflects the rapid improvement in AI systems that specialise in reasoning tasks. Earlier models were able to solve only a small share of advanced mathematical benchmarks, but newer versions are beginning to tackle research-level questions set by academics.

Still, experts say the breakthrough should not be overstated. Solving the problem once out of eleven attempts shows that such capabilities remain unreliable and experimental.

Human verification also remains essential. Even when AI produces a solution, mathematicians still need to check every step of the reasoning.

For Naskręcki, the moment was both exciting and unsettling. After spending two decades crafting a problem meant to challenge machines, he found that an algorithm had finally crossed the barrier he thought would last much longer.

Rather than replacing mathematicians, he suggested, the new generation of AI may instead become a powerful research tool for them.