I see your point about safety, but I think that needs to be be in the AI model than on the hardware side.