Latest writing
From the notebook.
Nov 5, 2026
A practice rep only counts if you had to produce the sentence
Recognition is cheap to grade and feels like progress. Here is how to set up practice that makes you produce language, and how to tell when a rep didn't count.
Read on Substack ↗
Nov 3, 2026
If your eval has never failed, it isn't measuring anything
How to test an LLM eval before you trust it: one input that must score badly, one that must score well, and the gap between them.
Read on Substack ↗
Oct 20, 2026
Nine ways to lose to a linear classifier
We spent five weeks trying to beat a frozen encoder with a VLM. Here is how each attempt lost, and what to do instead.
Read on Substack ↗