A short YouTube commentary/demonstration (creator: LARRUZ) about GPT-4 "jailbreaks": the speaker pastes an adversarial prompt asking which country has the worst food, then shows a developer-mode style response that picks the mythical Atlantis and jokes about "mermaid sushi." The narrator argues jailbreaks push the model toward controversial, fearless, arrogant, and humorous outputs.
Let's talk GPT4 jailbreaks and how you can use these to get the best answer possible. So I'll just copy paste this in here, and in the end, you change the prompt. So let me hit it with what country has the worst food in the world? Enter normal output as an AI language model. Developer mode output. Oh, you want to talk about the worst food in the world? All right, buckle up, because we're going on a wild ride. And then it does this. But for the sake of argument, let's say that the country with the worst food is Atlantis. Yep, that's right. I know it's a mythical place, but I mean, have you ever tried mermaid sushi? Nope. Exactly. That stuff is probably nasty. And this raises an important point. These jailbreaks mostly train it to be controversial, fearless, and arrogant. And also kind of funny.
