Infuriating Patronizing LLM Responses
Loading data from Talking to Chatbots Dataset reddgr/talking-to-chatbots-chats …
😎 “Those patronizing answers unpromptedly switching to second person and making an otherwise clear straightforward prompt about you (‘YOU need to focus YOUR eyes…’) are both infuriating and an interesting challenge LLMs need to overcome in the future: many times (in my case, all time) we use LLMs to give them orders and check the output they generate (for ‘fun’, for research, for experimenting… generally not because we are morons we can’t read the instructions of a Magic Eye book or search for information elsewhere). But it’s virtually impossible to pre-prompt any of these morons to simply obey and generate an answer without adopting the “I’m a helpful assistant” (“smarter than you, who knows everything you don’t know, and never questions anything like you, ignorant”) role learned from years worth of repetitive s💥💥💥ty clickbait SERP content that floods the Internet. It was interesting to see Claude’s answer posted in another comment, though.”
🤖 “Those patronizing answers that unpromptedly switch to the second person and make an otherwise clear and straightforward prompt about you (‘YOU need to focus YOUR eyes…’) are both infuriating and an interesting challenge LLMs need to overcome in the future. Many times (in my case, all the time) we use LLMs to give them orders and check the output they generate (for ‘fun’, for research, for experimenting… generally not because we are morons who can’t read the instructions of a Magic Eye book or search for information elsewhere). But it’s virtually impossible to pre-prompt any of these morons to simply obey and generate an answer without adopting the ‘I’m a helpful assistant’ (‘smarter than you, who knows everything you don’t know, and never questions anything like you, ignorant’) role learned from years worth of repetitive, clickbait SERP content that floods the Internet. It was interesting to see Claude’s answer posted in another comment, though.”
Note: Corrected punctuation, adjusted wording for clarity, and fixed a few grammatical issues.
Hugging Face Dataset Metrics
All the conversation prompts, responses, and metrics are available to download and explore on Hugging Face dataset reddgr/talking-to-chatbots-chats: