A Middle Schooler Used AI to Challenge 300-Year-Old Math — What It

The Paper That Broke the Internet (Almost)

A 15-year-old in Gimpo, South Korea, published a paper on the preprint server arXiv claiming to disprove a 300-year-old mathematical conjecture. He did it alone — no professors, no research group, no formal academic affiliation. He used Claude to write code, ran the computations on his laptop for 38 hours, then used AI again to verify his results.

The conjecture at stake is the Rupert property, named after Prince Rupert of the Rhine, who posed the problem in 1693: Can a cube pass through a smaller cube? The answer, counterintuitively, is yes. The broader question — whether every convex polyhedron can pass through a copy of itself — remained open for three centuries.

Choeng Gi-hyo’s paper argues that the rhombic dodecahedron, a 62-faced solid made of 20 equilateral triangles, 30 squares, and 12 pentagons, cannot. He posted the result on Reddit, where it climbed to first place on a Claude-focused coding board. In a country where STEM pride runs deep and national exams determine futures, a middle schooler publicly challenging centuries of mathematical work is exactly the kind of story that amplifies instantly.

What happened next reveals more about the state of mathematics in the AI era than any single proof ever could.

Who Actually Wrote This Paper

The immediate reaction from established mathematicians was not disbelief so much as a careful distancing. Ohm Sang-il, a researcher at the Institute for Basic Science’s discrete mathematics group, was blunt: posting on an archive platform marks publication date recognition, not content endorsement. Seo In-seok, a professor at Seoul National University’s Department of Mathematical Sciences, said the same thing differently — an AI-assisted paper only counts as complete scholarship if the author understands every step and can defend it against questioning.

Choeng himself acknowledged the gap. He told reporters he could not independently verify every detailed step of the paper at his current level. But he offered a partial defense: other papers reaching the same conclusion were published around the same time using different methods, suggesting the result itself may be sound even if the presentation is incomplete.

This is the crucial ambiguity. The computational result — that the rhombic dodecahedron likely fails the Rupert test — may be correct. The path that got him there, mediated entirely through an AI coding assistant he did not fully comprehend, raises a question the mathematical establishment has never had to answer before: does a proof matter if the person who found it cannot explain why it works?

The Real Story Isn’t the Math

The Rupert property conjecture itself is a side plot in a larger story about who gets to do mathematics now.

For centuries, the gate was credential-based. You needed years of graduate training, mentorship from established researchers, and publication in peer-reviewed journals to contribute meaningfully to open problems. The system was slow, exclusionary, and expensive — but it also ensured that anyone claiming a proof understood it.

AI has cracked that gate open. A middle schooler with a subscription to Claude can now run computational experiments that would have required a research team a decade ago. The barrier is no longer access to advisors or lab space. It is access to the right prompt and the willingness to iterate.

This is not hypothetical. OpenAI recently claimed to have derived a solution to one of the seven Millennium Prize Problems — the Navier-Stokes existence and smoothness question — though the mathematical community has not accepted it. The pattern is emerging: AI produces results, humans decide whether to trust them.

The Verification Crisis

The deeper problem this story exposes is not about one teenager or one conjecture. It is about the erosion of a fundamental contract in mathematics: the author must own the argument.

Peer review exists because individual humans make errors. AI introduces a new category of error — the model is fluent but not truthful, confident but not necessarily correct. When Choeng ran 38 hours of computation through AI-generated code, he was not verifying his own reasoning. He was verifying that Claude’s code produced an output consistent with its own logic. That is a different epistemic act entirely.

Choeng’s own candid admission — that he cannot fully verify the paper’s details — is actually a sign of intellectual honesty. Most AI-generated results in any field never get this far into scrutiny. The fact that mathematicians are already debating its validity suggests the result has substance, but also that the community is trying to figure out how to evaluate work that falls outside traditional authorship models.

What Changes Next

The implications extend well beyond geometry or polyhedra. If a 15-year-old can generate a credible preprint on a 300-year problem using AI, the question is no longer who can participate in mathematics — it is what participation means.

Mathematics education in South Korea, already intense and competitive, will now face a new pressure: students who can wield AI tools effectively may outpace those who rely on traditional preparation. This advantage is not exclusive to Korea. It is global. Any student with internet access and a free or paid AI account now has access to computational tools that previously required institutional support.

The academic reward system is not ready for this. Journals still expect human-written, human-verified proofs. Funding bodies still tie grants to institutional affiliation. But the pipeline of discovery is changing faster than the institutions designed to evaluate it.

What is clear is that the era of AI-assisted mathematical exploration is here, and it does not care about your degree. The rhombic dodecahedron may or may not be Rupert. But the real shape of this moment is already visible — and it is wider, and more disruptive, than any single conjecture.