It's not about that we should drop LLM completely from the mix, but something like AI -> LLM control -> old-school classifier control -> script / human oversight is the way. If something has potential to cause millions in damages, it should be subjected to human oversight (likelihood / impact analysis needs to happen early in the system design).
Comments
It's not about that we should drop LLM completely from the mix, but something like AI -> LLM control -> old-school classifier control -> script / human oversight is the way. If something has potential to cause millions in damages, it should be subjected to human oversight (likelihood / impact analysis needs to happen early in the system design).