§ Machine learning

pgvector tip: don’t reach for an HNSW index until you actually have the rows to justify it. Below ~10k, an exact scan is both correct and faster, and you skip a whole class of recall-tuning headaches.

Please report your baseline. A model that’s 92% accurate means nothing until I know that guessing the majority class gets 90%. Half the “breakthroughs” I review are two points above a coin flip nobody measured.

Small models, honest metrics, boring infrastructure. That’s the whole job most days. The 70B model is not your bottleneck; your unlabelled data and your eval set are.

The baseline point generalises past ML. “We improved it 30%” — over what? Half our internal wins are noise above an unmeasured control.

Same lesson on the ML side: people add an ANN index and never check that their top-k actually contains the true neighbours. Silent recall loss is the worst kind.