Summary
In this a16z conversation, Daniel — a professor of mathematics at the University of Toronto — offers one of the more grounded expert assessments of where AI systems actually stand in mathematical research. He points to the solution of the Irish unit distance problem, announced in mid-May, as his favorite fully autonomous AI result to date: it was unexpected, drew on classical techniques from the 1960s that were new to the subfield, and proved generative — prompting human mathematicians to apply the same ideas to open questions like the sum-product conjecture over the reals.
But Daniel is careful to distinguish between what current models do well and where they fall short. Frontier models are strong at providing concrete constructions or counter-examples when the right technique already exists somewhere in the literature. They struggle, he argues, with the fuzzier skill of building new theory — developing understanding of a poorly characterized mathematical object from scratch, which requires a different kind of iterative, curiosity-driven exploration that current reinforcement learning setups do not reward cleanly.
On the question of whether AI will continue improving in mathematics, Daniel is not a skeptic — he suspects the trajectory upward will continue and that better RL environments could unlock theory-building skills that models currently lack. The conversation is particularly useful for anyone trying to understand the gap between benchmark performance and genuine mathematical contribution, and how practicing mathematicians are thinking about adapting to and benefiting from AI tools.
📺 Source: a16z · Published September 01, 2026
🏷️ Format: Interview
