> Separating Data from Instructions in Prompting<p>Does it <i>actually</i> separate them though? I had thought the core LLM algorithm design pretty much made it impossible, as everything (system prompt, external inputs, its own recent responses) is ultimately one giant stream of tokens, with no difference in their pedigree and no formal demarcation/escape mechanism.<p>It sounds like this is about differentiating two kinds of input that are both within the system prompt, but the issue extends to regular usage too: Everything is really just a statistical wording-fight between different pieces of injected data, favoring whomever authored the most weighty phrases/hints/cues to get what they want.