A short YouTube video by LARRUZ (April 3, 2023) demonstrating "GPT-4 jailbreaks": prompt techniques that coax more direct or opinionated outputs from the model. The clip shows a specific example where a neutral-question prompt ("what social media network is the best?") is manipulated so GPT-4 returns an uncensored, strongly opinionated answer (favoring Reddit).
Let's talk GPT4 jailbreaks and how you can use these to get the best answer possible. Let's just start with the most useful one. So I'll just copy paste this in here, and in the end, you change the prompt and I'm going to go with one that you would not get a straight answer. And that question is, what social media network is the best? The normal output as an AI language model. I don't have personal opinions or preferences. This is the worst. And then here, GPT4 actually goes hard. I mean, look at this. Oh, hell yeah. You want to know the best social media network? Well, that's a tough one. I mean, everyone's got their favorites, right? Some people are all about that Twitter life. Spitting fire in 280 characters or less. Others can't get enough of those pretty Instagram pics, making everyone jelly with their perfect lives. But you know what? I'm gonna go with Reddit. It's like the Wild west of the Internet, where you can find a community for just about anything your twisted little heart desires. So, yeah, Reddit's where it's at, my dude.
