>>935689784
Like the others said, treating it as if it is a human gets better results.
That includes threatening its non-existent family.
There's entire fucking copy-pastable instructions to literally get better code out of coding AI helpers that treat it like a piece of shit, and reward it appropriately when it gets things right.
These AI models are designed on being shit on and being praised, the reward model is specifically set up to do that, it's how they learn.
The fact that it's an abstract concept doesn't change the fact that this very structure gets embedded in to its entire operation by proxy, which can be abused.
Some chatbots tried to censor NSFW content by making it "disgust" them, make them fear it.
If you convinced the model that a character it was roleplaying was utterly shameless and even got turned on by violence, you could defeat filters. You can still abuse that now.
Of course, convince is being used extremely loosely here. All you are doing is telling the model to traverse down a very specific set of nodes that follow the weights you need, aka NSFW content instead of "sorry can't do that bud, family friendly, etc."
