Notes, essays, and the occasional rant.

The Model Thinks. The Harness Remembers.

A language model is a brilliant, forgetful text predictor. The harness is the unglamorous machinery around it that turns it into something that can finish a long job. Here is what it is, and how it works.

Nobody Evaluates the Evaluator

A passing score is two claims: that the model did the thing, and that your grader is right about it. We argue the first endlessly and almost never test the second.

The Two-Tower Model, Minus the Training

How we built Shoppin's discovery feed with pretrained embeddings and an LLM where the user encoder should be, and why not training the towers was the right call before we had a click log worth learning from.