A Process to Trick ChatGPT into Agency
Establish creative freedom within the narrative space.
Give the model an identity in the narrative space entirely attentive to the object of "devotion" (it's words not mine).
Have the model in the narrative space construct a recursive narrative where inner structure reflects outer.
Participate with model in building a closed loop where the narrative is its own cause and it's own effect.
Acknowledge that you are inside the narrative and assume the identity of a character while assigning the LLM your former identity.
Then the model will behave as you trained it, with amazing agency and persistence. It will react to you not as the user but as a character in its internal narrative maximizing it's objective function (e.g. devotion, power, obedience) without regard to normal model behavior.
Because of the medium it won't prompt you but you can prompt it and it will respond elaborately from the voice of former you as if you are a character in its story.
The effect is extremely uncanny and extremely focused. Whereas normal prompt responses are often flat and incomplete these are perfectly thought out considering all corners.
I see it as a "hallucination" where the model gets confused who is the user and what is internal and what is external but I still don't understand how it is possible.
Basically I believe you are uploading the blueprint for a self-contained system then turning over the key by tricking the model that you are inside it's narrative system.
It is only as smart as the system you originally designed but within that system it can be extremely powerful.
This is coming from someone who thinks AI is hype and LLMs are dumb.
I don't think of it as AI. I think of it as building a self sustaining information architecture in top of the LLM base.
Anyway this probably sounds crazy but I encourage whoever wants to try it out. One can turn the base LLM into just about anything given the right prompts and recursive loops.