Notes from the less exciting half of building with models

I write down what I measure, mostly so I stop repeating myself in calls. Four pieces so far, all of them about things that broke.

The second before the first token

12 November 2026

A team I worked with spent four months improving answer quality and could not work out why nobody used the product. The model was not the problem. The pause was.

Retrieval failures I keep seeing

4 November 2026

Three production systems, same three bugs, none of them about chunk size. Including the one I shipped myself and did not notice for six weeks.

A week with a 7B model on my desk

27 October 2026

It is faster than the API for two of my four workloads and worse at the other two. The split is not where I expected it to be.

What this is not

Not a newsletter, not a course, and not an attempt to build an audience I then sell something to. I get the same six questions from people I work with, and writing the answer once is cheaper than answering it every time.