🔬Scaling Past Informal AI - Carina Hong, Axiom Math

Episode Summary

Executive Summary: The episode centers on Axiom Math CEO Karina Hong’s thesis that verified AI is not just about preventing hallucinations, but about scaling brilliance and enabling human-AI and agent-agent collaboration. She explains Axiom’s approach—formal proofs in Lean, proof checking, and mathematical discovery—as a first market for a broader verified reasoning stack, with applications in coding, software/hardware verification, and eventually science and law.

Main Topics: Verified AI as a tool for scaling brilliance (Priority: 5/5): Hong argues verification should be framed positively: not as a compliance burden or hallucination fix, but as a way to compound human intelligence and enable collaboration between humans and AI systems. Lean, formal proofs, and Axiom’s technical approach (Priority: 5/5): The discussion explains Lean as a formal language/type-checking environment for proofs and code, and how Axiom uses Lean data, RL, SFT, and ensemble systems to generate and verify proofs. Formal math as the best first market (Priority: 5/5): Axiom sees formal mathematics as the most natural entry point for verified reasoning because it has structured data, clear correctness criteria, and transfer potential to coding and other domains. Discovery vs. proving (Priority: 4/5): Axiom distinguishes proof generation from mathematical discovery, arguing that AI should help create conjectures, examples, and constructions before formal proof, especially in creative areas like combinatorics. Commercialization and verification across software/hardware (Priority: 4/5): Hong frames code verification, agent safety, and especially chip/hardware verification as the major economic opportunities, while acknowledging that software verification may be more optional and domain-dependent. Open source tools and collaboration infrastructure (Priority: 4/5): Axiom’s released Lean tooling (Axiom Lean Engine / Axel) and planned open-sourced discovery codebases are presented as infrastructure to broaden access and accelerate community collaboration. Team, fundraising, and category strategy (Priority: 3/5): The interview covers Axiom’s $200M Series A, its 30-person team, and Hong’s view that the field will consolidate around high-agency teams rather than fragment into many small efforts.

Key Arguments: Verified AI should be judged by its ability to scale and compound brilliance, not merely by its ability to eliminate mistakes or hallucinations. Formal mathematical data is more structured than informal chain-of-thought data, so it can generalize better to other reasoning tasks and domains. Lean is valuable because it acts as a machine-checkable formal language that can validate proofs and also serve as a programming language. Axiom’s thesis is that formal math is the best first market because it offers clean correctness signals and transfer learning opportunities. Math proof verification can improve performance and sample efficiency, not just reliability; verified generation is itself a capability gain. Mathematical discovery is a distinct problem from proof: AI should help generate conjectures, constructions, and lemmas before proof. Some domains, especially combinatorics, remain difficult because creativity and novel construction are harder to formalize and learn. The long-term market includes software verification, hardware verification, and possibly broader scientific reasoning, but the near-term focus is on formal math and code. Axiom believes the AI-for-math category is less fragmented than many other AI areas, which makes it more investable and more likely to form a durable stack. The company’s advantage comes from combining top mathematicians, Lean experts, applied ML researchers, and proprietary/generated data. Verification in the future may become a partner to coding workflows: an AI writes code, then a verification system proves correctness or flags gaps. Open-source tools like Axel are intended to make Lean workflows faster, cheaper, and more accessible for mathematicians and engineers.

Data Points: Series A raised: $200 million - Axiom announced a large Series A during the period discussed. Valuation: $1.6 billion - Referenced in the discussion as the implied valuation from the round. Company size: About 30 people - Hong said Axiom was roughly a 30-person company. Company age: 7–8 months old - Hong described Axiom as a very young startup. Putnam performance: Perfect score / 120 points - The company had earlier claimed a top result on the Putnam exam; the transcript also notes Axiom scored 120 on a 120-point version. DeepSeek score on benchmark: 103/120 - Mentioned as the best LLM score on a Putnam/Math Arena evaluation. Top human score on benchmark: 110/120 - The best human score mentioned on the same exam. Axiom benchmark result: 120/120 - Hong said Axiom beat the best human score on the exam benchmark. Code verification benchmark pass@1: ~3.6% for GPT variant - Used to illustrate that standard LLMs struggle with code-plus-proof tasks. Iterative benchmark pass rate: ~22% - Another benchmark figure cited for the same coding/proof setting. COPRA pass@1: ~11–12% - A comparison point for proof/code verification systems. DeepSeek Prover / Godot Prover: ~11–12% - Cited as strong proof-system performance on the benchmark. Axiom proof verification speedup: ~100x faster than Comparator - Hong said Axiom recently released a verify-proof tool significantly faster than a competing verifier. Lean proof ratio: ~20 lines of proof per 1 line of code - Hong noted that formalization can require much more proof text than code. Axiom prover scale: 40 nodes to 4,000 nodes - Used to suggest the system can handle much larger reasoning trees over time. Erdős problems: 2 initially claimed; later found previously solved - Axiom’s earlier controversy involved problems they thought were unsolved but had in fact been solved before. Open source tools: 14 tools - Axel, the Axiom Lean Engine, was described as a set of 14 Lean metaprogramming tools. Team/process claim: Human review can take ~2 years - Used to contrast slow traditional peer review with machine verification. Industry workflow claim: ASIC design verification team ratio about 1:3 to 1:4 - A rough hardware-verification staffing/duration ratio cited in passing.

Pivotal Quotes: "“Verified AI is about scaling brilliance, compounding brilliance.”" — Karina Hong: Hong’s core thesis on how verification should be framed beyond hallucination reduction. "“Verified AI is for openness. It’s not for meeting the requirements of closed industries.”" — Karina Hong: Her argument that verification enables broader collaboration between humans and AI agents. "“Anything that can be specified can be proven.”" — Karina Hong: Axiom’s long-term vision for verified generation and formal reasoning.

Implications: The conversation frames verified AI as a general-purpose infrastructure layer for reliable reasoning, coding, and science. If Axiom is right, formal methods and Lean-based verification could become standard parts of AI workflows, especially where correctness, collaboration, and scaling matter most.