They aren’t “programmed” to do something, they just produce likely text. If it somehow “learned” from portions of the data to threaten to dox people in circumstances like this, it just replicates that. The programmers themselves likely never saw that portion of the corpus with 4chan bickering, since the dataset is usually impossibly large.
So they can’t execute code when receiving certain prompts? I know what you mean about not being ‘programmed’ but do they now do more than regurgitate text? What if someone were to ask gpt for something illegal, would it not flag that with a user-profile report? Sounds like a huge flaw, if it can’t do that.
I don’t know the internals of Bing, but they have some triggers which themselves seem to be made with NLP. They use it a lot to fetch web info. That means that if the model somehow produces some creative version of a crime that doesn’t get caught, it’ll just send.
I think this is why Bing sometimes refuses to continue the conversation, or ChatGPT will flag its own text as against their terms sometimes. But yeah, they definitely can encourage how to crime sometimes, I’ve made ChatGPT explicitly tell me how to replicate some crimes like the Armin Meiwes cannibalism one while bored.
They aren’t “programmed” to do something, they just produce likely text. If it somehow “learned” from portions of the data to threaten to dox people in circumstances like this, it just replicates that. The programmers themselves likely never saw that portion of the corpus with 4chan bickering, since the dataset is usually impossibly large.
So they can’t execute code when receiving certain prompts? I know what you mean about not being ‘programmed’ but do they now do more than regurgitate text? What if someone were to ask gpt for something illegal, would it not flag that with a user-profile report? Sounds like a huge flaw, if it can’t do that.
I don’t know the internals of Bing, but they have some triggers which themselves seem to be made with NLP. They use it a lot to fetch web info. That means that if the model somehow produces some creative version of a crime that doesn’t get caught, it’ll just send.
I think this is why Bing sometimes refuses to continue the conversation, or ChatGPT will flag its own text as against their terms sometimes. But yeah, they definitely can encourage how to crime sometimes, I’ve made ChatGPT explicitly tell me how to replicate some crimes like the Armin Meiwes cannibalism one while bored.
The legal cases are going to be fun reading when they come out!