🔮 Exponential View #575: AI’s math breakthrough and its creative limits

Discrete geometry and algebraic number theory don't usually talk to each other. They're separate cultures within math, each with its own specialists, its own conferences, its own blind spots about what the other field even cares about. So when an OpenAI reasoning model solved an open problem that had sat unsolved for 80 years by drawing an unexpected bridge between the two, the mathematicians who checked the proof called the connection both surprising and non-obvious. That's not a small thing. It's the kind of leap that echoes AlphaGo's famous Move 37 — a move no human would have played, but one that turned out to be exactly right.

What makes this interesting isn't just that an AI solved a hard problem. Plenty of hard problems get solved eventually by someone grinding through the math. What's notable is the shape of the solution: it required knowledge that lives in two different specialist silos, and the model apparently didn't respect the wall between them. Modern science is organized into narrow fields partly because no single human can hold everything in their head at once. AI systems don't have that constraint in the same way, and that suggests one of the biggest paybacks from these tools won't be raw speed — it'll be stitching together disciplines that have drifted apart simply because nobody had the bandwidth to work across both.

Speed matters too, though. A separate example makes that case. Robin, a multi-agent AI system, ran an entire scientific cycle on its own: forming a hypothesis, picking which experiments to run, analyzing the results, then refining its thinking based on what it found. Human researchers handled the physical wet-lab work, but the intellectual scaffolding — the loop of guess, test, learn, guess again — was driven by the system. The result was identifying an existing drug that could potentially be repurposed to treat macular degeneration, a discovery that came out of compressing a process that normally takes scientists months or years of iteration.

Taken together, these two stories point at something bigger than either result on its own. AI's contribution to research might end up being less about grinding out answers faster and more about noticing connections and running loops that the human system of specialized, siloed science structurally struggles to produce. That's a genuinely different kind of leverage than "AI writes code faster" or "AI summarizes papers." It's leverage on the shape of discovery itself.