I do not want this site to read like a row of sealed boxes.

The useful notes, at least the ones I return to, leave a few wires visible. They say what was tried. They say what was ignored. They admit what the toy example is smuggling in. They leave a place where the next person can press with a thumb and ask whether the argument bends.

That is the mood I want here: precise, but not embalmed.

The center of gravity here will be AI, machine learning, computer science, statistics, mathematics, and the psychology of using imperfect tools. I care especially about the gap between elegant models and the systems that have to survive real data: sparse observations, stale features, silent joins, distribution shift, and the human habit of trusting a number because it arrived with decimals.

Some entries will start with a bug. Some will start with a paper that bothered me in a productive way. Some will start with a tiny browser toy because an idea often becomes more honest once it has to move under its own rules.

I expect the format to wander:

  • a field note from a system that behaved badly;
  • a miniature derivation with the missing assumption circled;
  • a lab where the sliders are more persuasive than the prose;
  • a reading note that preserves the author’s useful friction;
  • a checklist I would actually want during a review.

The promise is not polish. Polish is easy to counterfeit.

The promise is that each note should expose enough of its machinery to be argued with. If a claim depends on a sample plan, a clock, a loss function, a hidden join, a prior, a queue, or a user who is quietly adapting to the tool, I want that dependency above the fold.

That should keep the notebook precise without making it sterile. When a note fails, I want the failure to leave a useful mark instead of disappearing under a confident summary.