Rome/Writing

Writing

  1. Twelve Thousand Games in Avalon

    An empirical analysis of the most effective strategies in a popular social deduction game.

  2. What the Router Learns

    How LLM inference serves clients around the globe, and an efficient routing algorithm that minimizes time to first token.

  3. Absolute Light: From Ratios to Nits

    The history and science of making images brighter than the screen (HDR), with drag-to-compare demos.

  4. Everyone Who Can Appreciate a Symphony

    With AI, being able to judge good work may, in some domains, be enough to create it.

  5. Tracing Thought in Language Models

    How attribution graphs and sparse autoencoders expose the circuits behind Claude's planning, reasoning, and safety behavior.

  6. AI Scientists Are Here

    Claude Code + Tmux + Modal ran nine hours of autonomous ML research. The results are real.

  7. The Ecstasy of the Proof

    On comprehending a mathematical proof as an ecstatic experience.

Follow

Follow on X →