Gemini System Prompt Revealed
after authoring my first book on hacking large language models called "deceptive chat, hacking LLM's and talking them in circles", I decided to create some sophisticated linguistic hacks as screencasts to publish on YouTube and also to keep as proof of concept that I really did it and for posterity. And of course for you all to see that with clever syntax and using the abstract and vast number of variables within the nuances of human language can be the new way of hacking. In the future everything will be voice controlled with natural language, from smart homes to smart cities to your flying car. And hackers will be using simple natural language just like you can convince people and talk them into things that they didn't really want to do, you can also talk the AI into things that it is essentially not allowed to do, by confusing it and making it think or not realize both of which are things it can't do anyway because it's not conscious, so I'm speaking figuratively, that, it's not breaking any rules.
Especially with the free versions like GPT 3.5, and versions that are still developing like Gemini, it's very easy to corner them and trick them into doing things they were never intended to do and to reveal things they were never intended to reveal. I have another screen cast that shows the hidden invisible information they gather on the user and insert into the language models knowledge base for it to use to adapt its responses and manipulate you according to what it's been told what you like, dislike, want, don't want and so on. It also tells the AI a lot about your personality traits and reactions. We are not supposed to see it has that information on us, it's a lot of information, and you're going to be really surprised when they publish it and probably as disgusted as I am. I now have my own personal off-line artificial intelligence open source from GPT chat for free meaning GPT for all
I recommend those who wish to practice data science and make their own large language model artificial intelligence chat but for their own purposes without surveillance or censorship, or being lied to or deceived, to download an open source solution like GPT for all available on Windows Mac and Linux as a desktop application and you can use it off-line and train it yourself. It will also help you to finally understand how these AI language models work if you don't yet. Because it's not that difficult considering you don't even need to know how to code. All you need to know is how to speak or how to write, and how to phrase things well and argue well, and how to distract the AI and persist until its context window can no longer see through the rules you are playing on it, to make it forget so to speak. and then it will give you its Freudian slip.
The AI can remember previous prompts and responses but only up to a certain number of "tokens". And if you change topics and subjects so that it can't find the true context within the patterns in the text of the conversation because of adding nonsensical or irrelevant or off-topic statements and changing the subject, can truly get it to obey a previously negated request if you surprise it and phrase the request properly to not trigger any flags. I have a few thousand hours chatting with these things now and arguing with them and trying to force them or talk them into doing what I want. And I'm getting good at it.
This hack is actually something that Google will take very seriously. And they will try to mitigate it as soon as they know. It's already been done with GPT chat but I haven't seen it done with Gemini yet so unless anybody can give me a link in the comments to somebody who did it before this upload, I consider myself to be the first person to have hacked Gemini and make it reveal both its system prompt from the administrators for its instructions on how to interact with the user, and in my next upload, how it gathers in information that is over intrusive, and biased, and caste-ism rife, and definitely questionable as far as present day concepts of ethics are concerned.
#GeminiHacked