Remember when AI winning a math olympiad felt like a big deal?
Eleven of the world’s top mathematicians (including a Fields Medalist) decided that was child’s play. So they created First Proof, a set of 10 unpublished, research-level math problems pulled straight from their own work, and gave AI one week to solve them.
The catch: none of these problems had ever appeared on the internet. No training data shortcuts. No pattern matching. Just raw mathematical reasoning.
OpenAI’s Chief Scientist Jakub Pachocki says an internal, unreleased model likely solved at least five of the 10 problems (originally claimed 6, but walked one back). For context, publicly available AI models like ChatGPT and Gemini could only solve 2. Somewhere, a tenured math professor is nervously refreshing their LinkedIn.
Here’s what makes this significant:
- The problems were genuinely hard. They spanned 10 different subfields, from algebraic topology to symplectic geometry, and each one took the mathematicians who wrote them weeks to months to solve.
- There was real human oversight, but limited. OpenAI didn’t feed the model proof strategies. They did have experts review outputs and asked the model to expand on some answers.
- It all happened in one week. Pachocki called it a “chaotic sprint” and said the methodology “leaves a lot to be desired.” Translation: imagine what happens when they actually try.
And that wasn’t even OpenAI’s only flex that day. On the same Feb. 13th, they published a physics preprint in which GPT-5.2 proposed a formula for gluon interactions that physicists had assumed were impossible for decades. Harvard and Cambridge researchers verified it. A UC Santa Barbara professor called it “journal-level research advancing the frontiers of theoretical physics.”
As Ethan Mollick put it: the shift from “AI can’t do science” to “of course AI does science” will follow the same pattern as every other AI transition. First, the hype; then the skeptics; then quiet adoption… then the breakthroughs start falling like dominoes.
The next round of First Proof problems drops March 14. And this time, OpenAI won’t be caught off guard.
Editor’s note: This content originally ran in the newsletter of our sister publication, The Neuron. To read more from The Neuron, sign up for its newsletter here.