137. 对洪乐潼的4小时访谈：AI for Math、把数学变成Lean、数学天书中的证明、直觉、被创造与被发现的
Summary
This 张小珺Jùn｜商业访谈录 episode interviews Hong Letong / 洪乐潼 about mathematics, AI For Math, and the deep-tech company Axiom. The technical core is that AI for mathematics is not only about solving contest problems; it is about turning mathematical claims, definitions, and proofs into verifiable formal artifacts through Lean Theorem Prover, Mathlib, Interactive Theorem Proving, and Auto-Formalization. The business core is that Axiom is presented as a systems company around Axiom Prover, with early commercialization more likely in Formal Verification than in a consumer math product.
Key Claims
- Hong Letong / 洪乐潼 frames mathematics as a civilization-like structure built from accepted axioms, definitions, intuition, proof, and communal judgment rather than as only contest problem solving.
- She treats AI For Math as a stack: prover, conjecturer, knowledge base, and auto-formalization all matter, so theorem proving alone is not the full AI Mathematician.
- Lean Theorem Prover and Mathlib matter because formal proof turns mathematical output into executable, checkable, error-reporting code-like artifacts.
- Axiom Prover is presented as proving all 12 Putnam Competition problems from the 2025 exam after human conversion into formal statements, but the source treats this as a system-and-team event, not merely a model benchmark.
- Axiom is described as a deep-tech systems company rather than a model company: the bottlenecks include sparse Lean data, brittle formal language, slow verification tools, auxiliary tooling, and orchestration choices.
- The episode contrasts elegant human proofs with AI proofs that may be thousands of lines of Lean; useful AI mathematical reasoning need not imitate human taste at first.
- Auto-Formalization is framed as underappreciated because turning informal papers, definitions, theorems, and proofs into Lean can be as hard as proving a statement.
- Formal Verification is presented as Axiom’s most plausible first market, especially for chips, software, and complex systems where proofs can replace or strengthen finite test cases.
- The source links mathematical proof and code generation through the claim that “math is code, code is math”: proof systems can improve code verification, and software can become more like formally checked mathematics.
- Human mathematicians are not simply displaced in the source’s view; they move toward higher-level abstraction, problem choice, conjecture, benchmark design, and research taste as AI proof capacity improves.
- The long-run vision is Mathematical Abundance: a math-rich world where AI can expand mathematical discovery and feed theory back into science, engineering, and verification.
Key Quotes
“Math is code，code is math” — the episode’s bridge between mathematical proof and software verification.
“bet system，不 bet model” — Hong’s description of Axiom’s technical bet.
“学徒” — her proposed epitaph, framing the founder role as continued apprenticeship rather than status completion.
Connections
- Hong Letong / 洪乐潼 — guest and source of the episode’s mathematical, technical, and founder claims.
- Axiom, Axiom Prover, Shubo, and Ken Ono — company, prover system, cofounder, and senior mathematician central to the startup story.
- AI For Math, AI Mathematician, Interactive Theorem Proving, Auto-Formalization, Lean Theorem Prover, and Mathlib — main technical layer introduced by the episode.
- Formal Verification, Formal Specification, AI Verification, and AI Coding Verification — verification branch connecting mathematical proof to software and chip markets.
- Google DeepMind, OpenAI, Anthropic, ByteDance, and Doubao — AI organizations or products discussed as peers, precedents, or competitors in AI reasoning and AI for math.
- AlphaGeometry, AlphaProof, and Putnam Competition — benchmark and system references used to position the field’s progress.
- AI For Science, Recursive Self-Improvement, Discovery Model, and Research Taste — broader research and self-improving AI themes extended by the source.
- Subagent Workflow, AI Skills, and AI Organization Design — system and organization mechanisms the source says Axiom uses or needs.
Contradictions
- No direct contradiction with prior wiki content.
- The source sharpens existing AI Verification and AI Coding Verification themes by arguing that formal mathematical proof is a stronger verifier than ordinary tests, while still depending on correct Formal Specification.
- It qualifies broad AI For Science optimism by treating mathematics as a cleaner digital sandbox than wet labs, robotics, or physical experimentation, where feedback can be faster and more exact.