For the past several decades, mathematicians did roughly what they'd always done — they sat in rooms, filled chalkboards, and published papers. The day-to-day practice of research mathematics was stagnant from one generation to the next. That era is now ending. Since the beginning of this year, AI systems have disproved open conjectures, solved problems thought to be beyond reach, and synthesized results across fields that no single human could hold in their head.

Begin Proof is a new video series with the people at the frontier of this shift — the researchers building the models and the mathematicians figuring out what their work looks like next. It's hosted by Slater Stich, a partner at Bain Capital Ventures, where he backs early-stage founders building infrastructure and tools for data teams.

In the first episode, Slater sits down with Noam Brown, the OpenAI researcher who leads multi-agent reasoning and test-time compute scaling, and was a key architect behind the o1 reasoning model. Less than a year after OpenAI's models swept gold medals at the IMO, IOI, and ICPC, the company announced something mathematicians had not seen before: a general-purpose AI model had disproved a conjecture by Paul Erdős, a problem that had attracted serious attention for decades without resolution. The model found a counterexample construction so intricate that one Fields Medalist reportedly lost sleep trying to understand it. They discuss what the result means, what it doesn't, and what comes next.