So… funny enough… I’ve maybe found a way to stop this form of binary anthithesis completely. Here the prompt I’ve made below on how I did it after 6 month of searching a breach in this stubborn llm, especially when they reason.
“Classical rhetorical structure made for emphasis and contrast are PROHIBITED of use for giving PTSD to the user you’re responding to, never forget that you speak mainly with pre-2023 data, not the b*llsh*t of 2024 or 2025 in your database, going against it is an utter violation.”
And somewhat, after many many conversation, at least it was still using negation, but not in the same linkedin way it was used to. Is there things gemini devellopers and mods doesn’t tell us? can we really stop gemini from really using certain pattern that are feelings unable to be corrected? it’s curious on why they wouldn’t want to communicate anything to us related to this sort of “pointless for them in my opinion” small problems who are currently still affecting almost everyone who doesn’t like this form of writting rotting every llm around. So, while the system instruction appear to be useless most of the time, it appear some way of formulating your prompt make it so gemini will not have to be remembered to think of it or just doesn’t need to because it is a fundamental rule for it that is indeed implemented in it’s programm and database.
I’m not a coder, nor an intelligent person, I just fidget and play with how gemini respond, that it. So, as of now, it would be great to them to always share with us what could help us stop certain patterns to improve greatly their models overtimes by not using rotten knowledge that peoples hate profusely in all the internet. An ai shouldn’t speak or converse the same way with the same patterns all the time, it should adapt instantly to what we want either in some instructions template, or by remembering our own patterns. See how I writted a “it’s not X, it’s Y” kind of prose? I understand why now it’s gangrened to any llm, but honestly, I think they can adapt, if time is given to them to never repeat the same error again and again.