There is no intent, nor is there a mechanism for intent. They don't do long term planning nor do they alter themselves due to things they go through during inference. Therefore there cannot be intentional deception they partake in. The system may generate a body of text that a human reader may attribute to deceptiveness but there is no intent.
I'm not an ML engineer - is there an accepted definition of "intent" that you're using here? To me, it seems as though these GPT models show something akin to intent, even if it's just their chain of thought about how they will go about answering a question.
> nor is there a mechanism for intent
Does there have to be a dedicated mechanism for intent for it to exist? I don't see how one could conclusively say that it can't be an emergent trait.
> They don't do long term planning nor do they alter themselves due to things they go through during inference.
I don't understand why either of these would be required. These models do some amount of short-to-medium term planning even it is in the context of their responses, no?
To be clear, I don't think the current-gen models are at a level to intentionally deceive without being instructed to. But I could see us getting there within my lifetime.
