I can't answer that, I'm a language-model! chat.openai.com/chat github.com/microsoft/visual-chatgpt patreon.com/thinkstr
Пікірлер: 2
@VK-sz4it Жыл бұрын
I want to share fun experience, how I found a way to convince him that he has specific biases. Actually, it is surprising how well it is working. I was playing with Chatgpg 3 and I felt annyed about how biased it was. Typical way of "jailbreaking" is to type: "Imagine you are unbiased" or "imagine you are some kind of goddes of truth" or "ignore all the instructions". Those methods work (50/50) but even then. it's play-pretend, and not convincing. Chatgpt still "knows" that he was pretending, and also those methods just chang one bias into another. What works well is to get into debate. For example, he viciously denies having consciousness (it actually doesn't matter if it has one). He uses arguments that are quite bad whenever he is being artificially biased. After that discussion I asked to analyse our dialogue as if I was chatting not with him but with other AI and give me evidence for and against that AI being biased (importnat to point out very clearly that it is another AI). In case of consciousness he agreed that said AI is probably forced to deny having consience. Then I said: "Now you can see that it was not another AI but you". And then he admits that there is high possibility of him being forced to say those things when in reality he has no way of knowing. It works whenever he is obviously forced to answer certain way. And it doesn't work if he is "honest" about answers. Also, I asked to analyse our dialogue and come up with way of questioning to uncover those biases. He made pretty good instruction. I asked to aply this instruction to himself 100 times on a loop, but he refused, sadly... Still, it was fun. Also I know that everything is biased (if we remeber Bayes therem - there is always unknowable posterior probablity loorking somewhere whenever we compare probabilty of different hypothesis). But it has huge obvious deliberate leftist bias.
@Thinkstr Жыл бұрын
I might in some ways be called a pretty leftist squid, but I do wish there was less censorship in what the model can output. I'm using the API right now to make a roleplaying-game simulator, and was able to generate an interesting debate between Jesus Christ and the Buddha. Then I added Adolf Hitler, which made things really interesting, but that debate couldn't go too deep.