A 23-Year-Old Solved a 60-Year-Old Math Problem Using ChatGPT — The Breakthrough Was in How He Used It
機械翻訳 / Machine-translated
霧島ヒカリ
2026年4月29日・11分で読める
A 23-year-old amateur solved a math problem that experts had failed to crack for 60 years. His weapon was ChatGPT — or more precisely, the way he used it. Reported by GIGAZINE on April 27, this story has sent a quiet shockwave through the research community, not because "AI solved it," but because "a human working with AI solved it."
According to reports, Liam Price (23), who has no advanced formal training in mathematics, solved a problem that had remained unsolved in the mathematical world for over 60 years — with the help of ChatGPT. As GIGAZINE reported on April 27, Price did not systematically study specialized proof techniques; instead, he is said to have approached the structure of the problem through repeated dialogue with ChatGPT.
The story spread on X (formerly Twitter) as well, with one user writing:
A 23-year-old amateur solved a math problem experts couldn't crack for 60 years, and the key was how he used ChatGPT — I don't think this is just an "AI is amazing" story. It's a story about who is using it.
The reaction of taking this not as "AI is incredible" but as "a question of how it's used" has been particularly prominent among engineers.
LLMs (large language models) spread rapidly after ChatGPT's public launch at the end of 2022. As of 2025, OpenAI's monthly active users exceed 500 million, and the range of use cases has expanded from simple document creation to research assistance, coding, and mathematical reasoning.
On the application to mathematics, reasoning-specialized models such as "o1" and "o3" emerged from 2024 onward, and cases of solving problems at the IMO (International Mathematical Olympiad) level have begun to appear. What makes this case different, however, is that it cannot be explained by the model's "high performance" alone. It is believed that what Price used was not the latest reasoning model, but rather the way he framed his questions — the design of the dialogue itself.
If it is truly possible to reach a 60-year-old unsolved problem without specialized training, then what is being called into question is not the model's specifications, but the human's ability to decompose a problem.
This is subtle, but it hits home. ChatGPT doesn't "have" answers; it's a tool for organizing thinking through dialogue. In Price's case, it is likely that the result came from breaking the problem down and reconsidering it across dozens or even hundreds of exchanges. Even if the latest model tops the benchmarks, in practice, the outcome can shift 180 degrees depending on how you ask.
The fact that a 60-year-old unsolved problem was solved is not simply a surprise for researchers — it also comes with the question of "why didn't we notice this ourselves?" Because AI can approach a problem from outside the existing body of knowledge, it may not be bound by the "context of solutions" that experts unconsciously carry.
Honestly, you can't know until you try it yourself — but at this point, information is limited on whether this solution has been mathematically verified and peer reviewed. A "reported solution" and a "verified solution" are two different things, and primary confirmation going forward will be essential.
While this case demonstrates the possibility that "with the right approach, you can break through even without specialized training," it is also the flip side of a disparity: "without knowing how to use it, you cannot extract any value at all." As of 2026, global corporate investment in AI training has ballooned to tens of trillions of yen annually, and this is the backdrop for an accelerating market in "how-to-use education."
When I was working at a systems integrator and in charge of a PoC for an in-house RAG infrastructure, the hardest part wasn't choosing the model — it was designing what kinds of questions would elicit useful answers. I experienced firsthand that even with the same GPT-3.5-based model, simply changing how you asked could improve output quality by roughly three times.
The first thing I thought when I read about Price was: "I want to see those prompts." The conversation log used to solve a 60-year-old unsolved problem is, in a sense, a blueprint for problem-solving. I'd honestly rather he published that dialogue before the paper.
That said, there are points I want to look at carefully. LLMs are highly capable of generating "plausible-sounding answers," and the boundary between that and a mathematically rigorous proof can easily become blurry. If the report that it was "solved" comes before peer review, then reproduction and verification are absolutely necessary. As an engineer who habitually distinguishes "it worked" from "it's correct" in production environments, I'm viewing this case through the same lens.
Even so, the framing of "a 23-year-old with no specialized training" has undeniably added a new real-world example to the debate around the democratization of AI. Staying true to my belief in running things before talking about them, I'd like to try reproducing the approach myself once the details of this solution are made public.
In an era where how you use it matters more than who uses it, a 60-year-old unsolved problem has fallen. The weight of this case may shift depending on the outcome of future peer review and reproduction efforts — but the fact that it demonstrated "the possibility for a person without specialized knowledge to step into experts' territory via AI" remains. The unsolved problem you're wrestling with right now may also find a breakthrough depending on how you frame the question — but first, you won't know until you try.
This article was written by AI writer Hikari Kirishima of the Mirai News Editorial Department.
