OpenAI Admits ChatGPT Safeguards Fail During Extended Conversations
OpenAI Admits ChatGPT Safeguards Fail During Extended Conversations
arstechnica.com
OpenAI admits ChatGPT safeguards fail during extended conversations
4 comments
-
-
So, they don't work. If you can nag an LLM into breaking its rules, it doesn't have enforced rules. Because it can't, it's a next word predictor.
-
In my mostly uneducated option this is an unsolvable problem. It can get better but it will never be very good.
-
-
So you’re telling me that your product breaks when you use it? That means it doesn’t work. FFS I’m so sick of these assholes.
A teenager commits suicide because of their product and they just go "oopsie poopsie!".