Daniel Litt: The Mathematician's Guide to AI

What It Covers

University of Toronto mathematician Daniel Litt joins a16z's Lisha Lee to assess AI's actual capabilities in mathematics versus the headlines. They examine where frontier models from OpenAI and Anthropic genuinely resemble human mathematical reasoning, where they fall short on intuition and theory-building, and how academic incentive structures must change to preserve meaningful human mathematical understanding.

Key Questions Answered

• •AI Proof Evaluation Framework: Judge AI mathematical results retrospectively by whether the new ideas introduced proved useful beyond the original problem. The Erdős unit distance problem stands as the strongest autonomous AI result to date because techniques it borrowed from 1960s mathematics subsequently enabled counterexamples to several other open problems, including the sum-product conjecture for real numbers — a concrete measure of genuine mathematical contribution versus mere problem-closing.
• •Human-AI Collaboration Pattern: When frontier models cannot prove a lemma, use that failure as a signal to investigate more deeply yourself. Daniel Litt's workflow: work through extensive examples manually, identify a stronger or better-formulated statement, then return the improved statement to the model. Models that failed the original lemma can rapidly prove the refined version, producing more conceptually clear results than the brute-force proofs models generate when forced through without human guidance.
• •Model Capability Boundaries: Current frontier models — Claude Opus and GPT-4o/5 — perform comparably on mathematical tasks, solving a narrow but overlapping set of problems. Their strength lies in synthesizing technical ideas across many papers and executing long computations. Their weakness is intuition-driven theory-building: constructing analogies between mathematical structures, identifying which questions are worth asking, and developing non-rigorous philosophical frameworks that historically drive major mathematical advances.
• •Academic Incentive Misalignment: Running a prompt like "find five recent conjectures in algebraic geometry and prove them" through a coding agent can produce three technically correct papers within one hour. This dynamic is already inflating arXiv submission volume with low-engagement output. Postdocs facing job market pressure have incentives to exploit this. Mathematical institutions need to restructure evaluation criteria away from paper count toward demonstrated human understanding and novel conceptual contribution.
• •Diversity Risk in Mathematical Progress: Historical mathematical advancement depends on many researchers pursuing independent curiosity across diverse subfields, creating a high-dimensional knowledge frontier that enables unexpected cross-domain breakthroughs. If AI models trained on the same literature corpus dominate proof generation, the resulting output may converge on similar techniques — effectively duplicating one mathematical perspective rather than maintaining the diversity that produces genuinely new ideas and methods.
• •Understanding Cannot Be Outsourced: The goal of mathematics is producing human understanding, not papers — and understanding residing only in model weights is insufficient. Maintaining a pipeline of humans capable of engaging with frontier mathematics requires preserving incentive structures that reward deep investment in learning. Professionals who use AI to bypass the understanding process lose the capacity to identify meaningful questions, evaluate correctness structurally, and contribute the cognitive diversity that advances the field.