Mathematicians vs. o4-mini: 10 unsolved problems, $75k prize

Mathematicians vs. AI: Ten problems resist while the rest fall

In May 2025, the University of California, Berkeley campus hosted a meeting that didn't make mainstream headlines until later: thirty of the world's most prestigious mathematicians sat opposite o4-mini, an OpenAI language model, for two days. The format was brutal. Each problem the machine failed to solve earned its author $7,500 (approx. €6,389). In the end, ten questions stood firm. The rest fell. The gathering was organized by non-profit Epoch AI under the FrontierMath project, and that detail—who pays for flights, who covers hotels—is the first thing anyone familiar with similar fairs asks.

What exactly peine in Berkeley

The headline circulating globally spoke of humiliation. The actual result was more nuanced and, therefore, more uncomfortable. There was no human defeat, but there was an explicit warning: the group managed to formulate ten problems that o4-mini could not solve, and everyone present understood that this advantage will become increasingly difficult to maintain. The possibility of reaching "level five"—questions insoluble even for the best humans—no longer seemed like science fiction.

The pogre is what scares. In just one year, the model went from solving level three problems to solving level four ones. If the curve holds, level five enters the horizon. That is the data. Everything else is noise.

The headline trap: neither humiliation nor oracle

The journalistic treatment of the matter deserves its own section. Some argue that reducing serious technical pogre to clickbait headlines is a form of disinformation as effective as any hoax. The question "Are mathematicians becoming obsolete?" wasn't asked by any mathematician. It was posed by those needing the click.

At the same time, the defensive reaction from the other side doesn't help either. Part of the public dismisses the issue with "it's a parrot that doesn't understand what it says." The nuance matters: a language model doesn't perform internet searches to solve a level four problem. It reasons, chains steps, and provides justification for each one. In some cases, it used schemes that hadn't occurred to the mathematicians themselves, although they knew another solution path. This is a qualitative leap, not a Google query.

The multiplier: for the expert and the novice

The most useful distinction emerging from all this separates two profiles. For those who master their craft, AI is a brutal force multiplier: extracting information from hundreds-page documents in seconds, suggesting structures, accelerating mechanical parts. For those mastering nothing, it's a superficiality multiplier. A specific case circulating involves programmers who don't know what they've been asked, copy the model's output verbatim, and integrate it without review. That isn't the machine's fault.

The underlying economic problem isn't that AI replaces anyone tomorrow. It's that work previously done by five people is now done by one with the right tool. That adjustment is already happening in specific sectors, without headlines.

What AI hasn't solved yet

There is a contrast worth keeping in mind. The same technology capable of reasoning about elite mathematical problems has failed to produce decent customer service. Not even a bot that resolves an incident without passing through three menus and a human agent. The market exists, is huge, and remains unresolved. This should temper enthusiasm about the imminence of certain things.

The other limit is deeper. Solving known problems suspiciously resembles trinc a recipe. True mathematical creation—introducing a new concept, like Galois groups, which were conceived over a couple of intuitive afternoons—doesn't depend on computing power. It depends on something we still don't know how to program.

---

Level five remains unreached. The ten problems that held up in Berkeley are still there, like a narrowing frontier. No one knows how long it will take for the next one to fall. No one knows if it will fall.