Many text model system prompts begin with "You are an AI assistant. You are helpful. You …. [user input] [model output]", but looking at what might have been in the training material, I would expect other perspectives to be more useful, such as
- "I am an expert. Ask me your questions. [user input] [model output]"
- "The following is a transcript of an expert explaining a topic to an interested user: [user input] [model output]"
- "Top list of the best answers to [user input]: 1) [model output]"
Or similar prompts that (possibly) match high quality answers in a large uncurated corpus of crawled text.
The "You are ... You do ..." prompts seem to imply that they are not processed in the way one would expect a transformer text generation model to work, but would require some sort of self-awareness. Why and how do they work?