This 24-Year-Old Raised $64M to Build an AI Smarter Than the World's Best Mathematicians

Overview

Karina Hong, founder and CEO of Axiom Math, is building an AI system she calls “mathematical super intelligence.” The objective is not only to solve known math problems, but to:

• Generate new conjectures
• Produce formal proofs in Lean
• Improve over iterations

The intent is to narrow the gap between today’s best human mathematicians and the hardest bottleneck in advanced math: verifying complex technical lemmas.

“Super Intelligence” as a Multi-Stage Goal

The discussion frames “super intelligence” in stages:

1) Matching elite performance

Hong argues the biggest gap between top mathematicians (e.g., Terrence Tao) and others is often the time/effort required to verify hard intermediate steps. If an AI can reliably generate correct intermediate proof steps, it can close that gap.

2) Superhuman discovery and scale

Beyond matching humans, the AI should help mathematicians:

• Think out of the box
• Continuously generate high-quality new problems
• Then prove them

This is framed as an iterative loop of conjecture → prove → learn from what succeeds or fails.

System Stack

Axiom’s system is described as a combined stack:

• Conjecturer
• Prover
• Knowledge base (including a human/proven math library)
• Mechanisms to verify everything

Neuroscience Connections

Hong connects her background in neuroscience to math/AI, describing mathematical concepts she studied that appear in neuroscience, including:

• Firing patterns in hippocampal grid cells related to mathematical tilings/optimization
• The Chinese remainder theorem linked to neural “capacity” ideas
• Persistent homology used in topological data analysis

She suggests the brain–transformer analogy is one reason AI for math feels like a natural fit, though she hasn’t seen much work where neuroscience directly drives machine learning research in this specific way.

Why the Market Is Opening Now

A major theme is that math-focused AI is bottlenecked less by raw “ideas” and more by infrastructure challenges, especially in formal proof languages like Lean. Two key bottlenecks are emphasized:

1. Data scarcity Despite enormous amounts of code online, there is far less Lean-formalized math data available to train models.

2. Auto-formalization difficulty Translating proofs written in natural language into formal Lean code is harder than typical translation—compared to differences in abstraction levels between human language and programming.

This produces a chicken-and-egg problem:

• Without formal data, auto-formalization is hard.
• Without auto-formalization, formal data can’t be expanded much.

Strategy vs. “Frontier Labs”

Hong contrasts Axiom with “frontier labs” and broad general scaling efforts:

• Frontier efforts can brute-force more broadly, but may move slower on the hardest verification work because pushing toward the hardest problems reduces breadth—“losing Shakespeare” while focusing on tasks like curing cancer or proving famous conjectures.
• Axiom emphasizes a focused, fast-moving small-team strategy centered on a core capability: verified reasoning (including auto-formalization), as a foundation for downstream applications.

Impact Beyond Pure Math

Verified reasoning is positioned as potentially useful in other domains:

• Formal verification of hardware/software, particularly where verification is expensive and critical
• Similar “bridging abstraction levels” challenges in chip design and chip verification
• A claim that strong math + coding ability correlates with performance on reasoning benchmarks (including legal-oriented tasks mentioned during the conversation)

Tooling for Discovery and Proof

The interview describes Axiom’s tooling beyond only formal proving:

Mathematical discovery vs. formal proving

Mathematical discovery is framed as generating:

• Constructions
• Examples

These guide intuition and are distinct from the later step of formal proving.

Highlighted toolkits

Two example toolkits are described as complementary to the formal proof pipeline:

• PatternBoost A generative method that creates novel examples to build intuition and guide proof strategies.

• Into (translative methods) Learns from problem–solution pairs to transfer between similar structures and reconstruct solutions for new tasks.

Preventing Millions of Trivial Theorems

When asked about preventing the system from producing huge volumes of trivial theorems, the emphasis is on quality control using:

• Judging interestingness (e.g., language-model ranking or other scoring)
• Proof structure signals (e.g., whether a proof bridges previously disconnected domains)
• The idea that novelty often comes from diffusion of ideas across fields

Verification Rigor and Failure Modes

Axiom’s verification approach is described as rigorous:

• If the system outputs a formal proof in Lean, the Lean “checker” effectively validates correctness—i.e., proofs can be verified through the system.
• There remains additional concern about the possibility that formalization itself could be wrong, meaning correctness depends on both:
  • the formalization step, and
  • subsequent compilation/checking steps.

Funding and Timeline

Axiom reports it has:

• Raised $64M
• Is working on solving ~130-year-old math problems

Examples mentioned include:

• A graph theory conjecture
• Longstanding problems involving Lyapunov functions in dynamical systems

Lean-based formal math is positioned as the differentiated foundation others can’t easily replicate.

Product Vision

Within years, Axiom aims to provide a default reasoning platform for professionals doing complex quantitative work—described as an IDE-like interface where users specify problems/specs and the model:

• proposes intermediate steps
• generates formal Lean artifacts
• allows checking to ensure outputs are sound

Over time, the goal is to require less human supervision, especially as high-quality training data accumulates.

Data Strategy

Hong describes a bold synthetic-data approach, focusing on:

• Converting informal math into formal Lean data
• Generating new synthetic examples from existing formal proofs

Rather than relying entirely on crowdsourcing Lean proof labor.

Core Claim: The Bottleneck Isn’t Just Compute

Hong argues that the biggest “industry bottleneck” for AI-for-science isn’t necessarily compute or robotics for math progress. Instead, it’s the:

Algorithmic / formal verification stack — providing coherent theoretical guidance before expensive real-world trial-and-error.

Presenters or Contributors

• Karina Hong (Founder & CEO, Axiom Math)
• Corey Nolles (Editor, The Neuron podcast)
• Grant Harvey (Writer, The Neuron podcast)