Daniel Litt: The Mathematician's Guide to AI

In this A16Z Infra podcast episode, Daniel Litt provides a nuanced analysis of AI's current and future role in mathematics. He begins by identifying his favorite fully autonomous AI result: the solution to the Erdős unit distance problem, which he considers creative because it brought techniques from the 1960s into a new domain and subsequently enabled solutions to other open problems.

Litt establishes a taxonomy of AI mathematical capabilities, noting that models excel at applying known techniques across multiple domains and handling long computations, but struggle with fuzzy activities like intuition development, theory building, and determining which questions are worth asking. He describes the models' output as "very human" in character—recognizable chain-of-thought reasoning that doesn't appear inhuman or require unexplainable leaps. He notes that both Claude and ChatGPT solve similar collections of problems, representing a relatively narrow band of mathematical activity.

A critical distinction Litt makes is between solving a problem and understanding it. He explains that his own work involves developing non-rigorous philosophies, analogies, and fuzzy reasoning to guide precise mathematical development. When models struggle with a problem, he's found that working through examples and discovering a better statement of the lemma—something models initially couldn't prove—led to more elegant proofs than the brute-force grinding approach the models could have taken.

Regarding his own mathematical practice, Litt describes himself as a problem solver motivated by understanding rather than aesthetics. He works on problems like the Gross-Sacks p-curvature conjecture, viewing them as benchmarks of ignorance. His approach involves finding small situations where understanding breaks down, developing theory to address them, and building analogies across mathematical domains. He notes that problems he's worked on for 3-5 years remain largely inaccessible to current models, as they likely require developing new techniques rather than applying existing ones.

Litt discusses how AI is actually affecting his daily work. Projects predating AI remain largely unaffected—he uses models primarily as a substitute for Google when learning related topics. However, AI has enabled him to pursue coding-intensive projects he previously avoided, demonstrating how the technology shifts which projects become feasible rather than accelerating existing ones.

On the comparative development of Claude and ChatGPT models, Litt notes they appear roughly equivalent in capabilities, with ChatGPT having reached useful performance for research mathematics earlier. He observes that ChatGPT 5.6 provides clearer exposition and better theory-of-mind regarding what the user knows, though both models are "pretty bad" at theory of mind generally.

A significant portion of the discussion addresses institutional and community-level concerns. Litt expresses worry about emerging incentive misalignment in mathematics: postdocs on the job market can now generate multiple papers by having models attempt proofs of recent conjectures (what he calls "playing the slot machine"), without developing genuine understanding. He's documented instances of three to five identical papers proving the same theorem being posted within days, indicating mode collapse in the models' reasoning. This threatens what he sees as mathematics' core value: letting "a thousand different flowers bloom" as people pursue diverse curiosities, which collectively expands knowledge boundaries.

Litt articulates why this diversity matters: mathematical progress depends on humans bringing varied intuitions, asking different questions, and developing understanding that can cross-pollinate unexpected domains. If mathematical research becomes dominated by what models naturally pursue—technically strong application of known techniques—the field may lose the generative diversity that enables breakthroughs.

Regarding the future, Litt remains confident that AI capabilities will continue advancing and may eventually develop theory-building abilities through improved reinforcement learning environments and curricula based on varying-difficulty conjectures. However, he emphasizes that optimal AI performance doesn't guarantee optimal societal outcomes. Institutions must actively design incentive structures to preserve human engagement with mathematics and maintain the pipeline of people developing mathematical thinking skills.

Litt also discusses the issue of proof verification and length. Models produce short, clever proofs because those are what can be checked; longer proofs may contain undetected errors. The recent 800-page AI-generated claimed proof of resolution of singularities exemplifies this—while impossible to verify, it clearly exceeds current model capabilities for rigorous arguments. This suggests published short results may overrepresent models' true capabilities.

On a personal note, Litt describes his approach to his three-year-old daughter's mathematics education, emphasizing conceptual understanding over grinding. He expresses hope that core values of mathematical thinking—clear reasoning and better understanding the world—remain relevant regardless of AI advancement, while acknowledging that educational institutions may require substantial adaptation.