The LLM is not a chatbot. We are using it to predict the text that would be produced by a chatbot, if one were to exist. In theory I guess this instruction mak

The LLM is not a chatbot. We are using it to predict the text that would be produced by a chatbot, if one were to exist. In theory I guess this instruction makes it more likely to output the kind of tokens that would be output by a chatbot that was quietly thinking to itself before responding. Does it work? Who knows! Prompt engineering is just licensed witchcraft at this point.