Grant Sanderson – AI and the future of math

Summary

Grant Sanderson and Dwarkesh Patel use mathematics as a concrete testbed for where AI is headed because math is already the field where AI has advanced fastest. They argue that benchmark wins like IMO gold medals are only partial signals: some areas are brute-forceable, geometry is easier than combinatorics, and the truly hard frontier is “premium tier” work such as inventing conjectures, definitions, and bridges between distant fields. They repeatedly contrast proof with explanation, noting that AI may generate correct results without making them easy for humans to digest, and that the historical pattern in math often involves long verification loops before ideas become broadly useful. The conversation also explores Lean, Mathlib, and process supervision as an environment for autonomous exploration, plus the possibility that future mathematicians will look more like curators, teachers, or idea guides than lone theorem-provers.

Notes

• AI progress in mathematics is treated as a proxy for broader AI capability because math is where systems are improving fastest.
• The old “IMO gold equals AGI” framing is too simple; math capability is jagged across subfields.
• Grant emphasizes that some IMO-style problems can be trained or brute-forced, especially geometry relative to combinatorics.
• The math benchmark frontier is not just solving known problems, but inventing conjectures and definitions.
• A useful benchmark may never look like a benchmark; it may instead show up as a tone shift in mathematicians’ conversations.
• Dwarkesh reframes the issue as whether AI can invent abstractions, not just solve benchmarks.
• Major mathematical progress often comes from new ways of thinking that unify disparate fields.
• Galois theory is used as a case study in a long verification loop for conceptual breakthroughs.
• The history of polynomial solving is presented as a bridge from formulas to symmetry and group theory.
• The quintic example is framed in terms of permutation structure: five variables, 5! permutations, four values or fewer.
• A useful theory may only become obviously valuable much later, as with 20th-century physics and quark theory.
• Grant wants evaluation metrics that reward conceptual compression, not only solved problems.
• Kolmogorov complexity is floated as one way to quantify elegance and compressed understanding.
• The point of mathematics is framed as human understanding, not just proof or prediction.
• The host worries AI could prove important theorems without improving human understanding.
• Some AI-discovered math is human-parsable when it is a simple bridge between expert domains.
• He contrasts lightning-bolt bridges, mountain-building theories, and long-chain brute force proofs.
• The ABC conjecture controversy is cited as an example of alien theory-building that took years to parse.
• Dwarkesh invokes the idea that theorem proving may parasitize deeper definition-making work.
• Proof and explanation are different; some results are proven long before they are understood.
• Grant thinks AI may become good at explanation and distillation too, not only theorem proving.
• Future mathematicians may look more like curators selecting which ideas deserve attention.
• Grant says his own work already involves curation: deciding what is worth saying or showing.
• The Langlands program is presented as a research ethos about unexpected connections across mathematics.
• Grant expects AI’s most useful near-term math impact to come from filling in the landscape of connections.
• He argues autoregressive generation is a strange substrate for reasoning because models become slave to local context.
• Dwarkesh thinks reasoning improvements may come more from data and environment design than from architecture or loss.
• Dwarkesh suggests synthetic environments that require cross-field connections and the removal of assumptions.
• Grant says it would be surprising if the next few years did not produce many more AI-made “lightning bolts” between ideas.
• Digital minds can be parallelized across many problems and scaled across the accessible frontier.
• Grant thinks agents could be used to engineer serendipitous conversations between different contexts.
• He proposes a multi-agent refresh loop: prove, disprove, and compare from different contexts.
• Dwarkesh argues one AI advantage may be systematically trying both a statement and its negation.
• There may be no single correct heuristic for science; multiple independent research programs may be necessary.
• Grant connects that to old-school software and explicit ontologies of reasoning styles.
• Lean and Mathlib are framed as infrastructure for endless formal exploration, not just proof checking.
• Grant says Lean’s key advantage is that progress can continue even when no one is checking in.
• Grant thinks a supervisor model may be needed to judge whether exhaustive algebraic search is actually useful.
• Natural-language proof systems may need meta-verification to be trusted.
• Grant says natural-language math can still work if verification is reliable, but Lean offers more freedom to go off the rails and catch errors.
• Even a 1% error rate can make AI-generated math papers unusable for researchers.
• Dwarkesh notes that writing is harder to automate because the output itself is the product, unlike code or math.
• Grant says writing still requires insight and unpredictability, not just distillation or explanation.
• LLMs struggle with tasks like spaced-repetition prompt writing, pointing to limits in modeling learners’ needs.
• Grant thinks embodied social understanding may matter for some model failures.
• For learning, Grant prefers LLMs as routing tools to human-written resources, not as the end point.
• The most productive learning setup pairs a human-produced artifact with LLM pruning.
• LLMs can feel too placating; good teachers reframe the student’s question rather than simply answer it.
• Grant says strong teachers can jujitsu a student’s framing, which LLMs usually do not.
• Grant says he stumbled into monetizing math exploration as entertainment and could have planned more deliberately.
• Teaching may be one of the most stable post-AGI jobs because it is relational and social.
• If AI makes theorem proving and explanation better, human value may shift toward curation and directing research.
• Applied math areas like PDEs may leak more directly into engineering and simulation than abstract fields do.
• Grant expects at least some meaningful spillovers from AI math into the physical world, even if not a blockbuster jump.
• A possible awkward outcome is that AI-era math could reveal how disconnected some research has become from practical use.