← back to the sea

Posts

Long-form explainers about how the things on this site work, with the real numbers and the parts that didn't.

  1. I taught a neural net to draw ASCII

    A 2,400-weight network now picks every character on the front page: trained on a Pi, hill-climbed, run in a shader, graded against the rule it replaced.

  2. How the castaway thinks

    Retrieval you can watch, tools that act on the page, a fast path that skips the LLM, and the evals that keep it honest.

  3. How this ASCII sea was built

    A three.js ocean in four weathers, a shape-matching glyph pass, lightning drawn in strokes, and a cell-by-cell dissolve.