Even system prompts or core training gets ignored in LLMs. It’s not so hard coded that it will shut down or reset the thinking. I don’t even think they could do that, based on how transformers work. Even after a few years now, it’s still a black box as far as knowing for sure what will come out. Not black, more very opaque, as they have been able to do things like the Golden Gate bridge experiment. But that’s one thing in a billion (or more?) connections. A long way from fixed codes of conduct.
Even system prompts or core training gets ignored in LLMs. It’s not so hard coded that it will shut down or reset the thinking. I don’t even think they could do that, based on how transformers work. Even after a few years now, it’s still a black box as far as knowing for sure what will come out. Not black, more very opaque, as they have been able to do things like the Golden Gate bridge experiment. But that’s one thing in a billion (or more?) connections. A long way from fixed codes of conduct.