newsfilter
← Interviews
Sequoia Capital40 min

Why Vlad Tenev and Tudor Achim of Harmonic Think AI Is About to Change Math—and Why It Matters

Vlad Tenev, Tudor Achim, Sonya Huang, Pat Grady

Interview
  • Harmonic's Core Thesis: Vlad Todorovic (CEO of Robinhood) and Tudor (co-founder) aim to create mathematical superintelligence by explicitly teaching AI to reason through math, rather than treating math as an emergent property of scale.

    • They view math as the foundational language of the universe, applicable to modeling phenomena from black holes to atoms and engineering designs.
    • Their hypothesis is that proficiency in rigorous mathematical reasoning transfers to superior performance in broader scientific and engineering domains.
  • Data Strategy & The "Data Wall": Harmonic relies heavily on synthetic data generated through internal self-play and reinforcement learning, contrasting with traditional foundation models that rely on static internet data.

    • Existing open-source repositories like mathlib (written in the Lean programming language) provide human-written proofs, but the volume is insufficient for training, necessitating AI-generated data.
    • The company argues that the "data wall" encountered by general-purpose AI (exhaustion of high-quality internet content) is a primary reason for their approach.
    • Synthetic data generation allows for the creation of problems with increasing complexity, mirroring how human mathematicians break large problems into smaller, solvable chunks.
  • Technology Stack (Lean & Formal Verification): The system utilizes Lean, a programming language created by Leonardo de Moura (Amazon), where mathematical statements are encoded in the type system.

    • Compilation in Lean acts as a formal verification step, providing an objective, binary reward function (correct/incorrect) essential for reinforcement learning.
    • This eliminates the ambiguity of human feedback loops, enabling rapid, recursive self-improvement cycles similar to AlphaGo Zero.
    • The mathlib library has grown organically into a massive open-source repository of formalized mathematics, surpassing usage for its original purpose of software verification.
  • Timeline & Predictions for Major Milestones:

    • IMO (International Mathematical Olympiad): The speakers predict winning the IMO by 2024 or 2025 with AI assistance.
    • Millennium Prize Problems:
      • Hybrid human-AI solutions for easier Millennium problems are estimated by 2026.
      • A fully AI-aided or AI-solved Riemann Hypothesis is estimated by 2029.
      • The current speaker estimates a 43% chance the next Millennium Prize is solved by AI or significant AI assistance, believing this figure is an underestimate.
    • Superhuman Reasoning: The team anticipates achieving reasoning capabilities exceeding top human mathematicians (e.g., Terry Tao) within two years, with models reaching the 99.99th percentile in the next year.
  • Future of Mathematics & Human Roles:

    • The role of the mathematician will shift from calculation to strategic direction: defining problems, selecting which areas to attack, and interpreting AI-generated results.
    • The field is expected to evolve past the "human + AI assist" phase toward fully autonomous systems, though human guidance will remain in the loop for problem selection.
    • Unlike zero-sum games like chess where optimal strategies eventually cap progress, math has no upper bound, allowing for continuous, unbounded improvement via increased compute.
  • Broader Applications Beyond Pure Math:

    • Software Engineering: The speakers predict a future where the vast majority of software is provably correct and verified by AI, shifting the industry focus from code review to rigorous specification.
    • Physics: Harmonic aims to accelerate theoretical physics research, potentially helping to deduce fundamental laws of the universe and design experiments to test them.
    • General Reasoning: If the core premise that "math is reasoning" holds, the technology could eventually apply to any domain requiring logical deduction, including humanities (e.g., history essays), though this is a secondary focus.
  • Historical & Strategic Context:

    • Why Now: The convergence of recent advancements in autoregressive language models (capable of infinite action spaces) and the maturity of the Lean formal verification tool (released officially in September 2023) created a viable inflection point.
    • Community Sentiment: The math community is split; younger mathematicians are generally pro-AI and pro-formal methods, while older practitioners remain skeptical.
    • Inspirational Figures: Vlad Todorovic cites John von Neumann (for his transition from math to computer science and game theory) and Gottfried Leibniz (for his 17th-century prediction of a "universal characteristic" or formal deductive language) as key intellectual influences.