Post #1223848

@cthos Can anyone explain to me how exactly a system prompt or any other context is expected to be privileged in an LLM execution? Is the prompt just more context but with a weighting multiplier so it has a disproportionate effect on the token prediction? Or is it somehow used as a metalinguistic wrapper around a quoted form of the user prompt? Surely it can't be as simple as "Play the part of a character with the role described by and respond in character to someone who has come to you and asked ''." But, if the LLM is just responding in any way at all to some concatenation of tokens from its context, there is no real semantic structure to what is input. Or does the LLM first process the system prompt from natural language to something closer to a vector of control parameters so that it can transform the effective matrix of predictive weights for token sequences? Regardless, I'm completely at a loss to understand why this seems to anyone like a properly tunable way to control token prediction.