Writing / working ideas in public

Essays on what AI systems know, how agents learn, and where verification changes the problem.

01August 27, 2026Formal reasoning

Latest essay

Do AI Theorem Provers Actually Understand Mathematics?

A model can know the theorem, solve the problem, and still fail the proof.

MathAdv separates mathematical knowledge, informal reasoning, formal execution, and robustness—failure modes that a single proof score collapses into one zero.

Read on Substack ↗

02May 31, 2026Agentic RL

Why AI Agents Need RL and How to Make It Work

The reward design approach to teaching agents when—not just how—to act.

SFT teaches the shape of a valid action. RL teaches action selection from outcomes. The difficult work lives in the environment, rollout system, and reward specification.

Read on Substack ↗