Writing

How I trained agent harnesses with a PyTorch-like loop over deterministic LLM inference, and how the gains transfer across models on Terminal Bench 2.0 and SWE-bench. [Read More]

Lessons from more than 1,000 agent-harness experiments: how experimental controls, evaluation noise, and promotion criteria shape what self-improvement can actually teach us. [Read More]

Trace a familiar training step through a from-scratch library to understand how computational graphs, backward passes, and optimizer updates make learning possible. [Read More]

Build a deep learning library from Python, NumPy, and math: tensors, an autograd engine, neural-network modules, and optimizers, with models trained from scratch. [Read More]

Video games were very attractive to me when I was kid. This is because the possibilities in the virtual world is vastly greater than the real world, as least during my teenage years where most of my life was filled with school work that can’t be related to in the... [Read More]