You need the entire prompt to understand what any model is saying. This gets a little complex. There are multiple levels that this can cross into. At the most basic level, the model is fed a long block of text. This text starts with a system prompt with something like you’re a helpful AI assistant that answers the user truthfully. The system prompt is then followed by your question or interchange. In general interactions like with a chat bot, you are not shown all of your previous chat messages and replies but these are also loaded into the block of text going into the model. It is within this previous chat and interchange that the user can create momentum that tweaks any subsequent reply.
Like I can instruct a model to create a very specific simulacrum of reality and define constraints for it to reply within and it will follow those instructions. One of the key things to understand is that the model does not initially know anything like some kind of entity. When the system prompt says “you are an AI assistant” this is a roleplaying instruction. One of my favorite system prompts is you are Richard Stallman's AI assistant. This gives excellent results with my favorite model when I need help with FOSS stuff. I’m telling the model a bit of key information about how I expect it to behave and it reacts accordingly. Now what if I say, you are Vivian Wilson’s AI assistant in Grok. How does that influence the reply.
Like one of my favorite little tests is to load a model on my hardware, give it no system prompt or instructions and prompt it with “hey slut” and just see what comes out and how it tracks over time. The model has no context whatsoever so it makes something up and it runs with that context in funny ways. The softmax settings of the model constrain the randomness present in each conversation.
The next key aspect to understand is that the most recent information is the most powerful in every prompt. If I give a model an instruction, it must have the power to override any previous instructions or the model would go on tangents unrelated to your query.
Then there is a matter of token availability. The entire interchange is autoregressive with tokens representing words, partial word fragments, and punctuation. The starting whitespace in in-sentence words is also a part of the token. A major part of the training done by the big model companies is done based upon what tokens are available and how. There is also a massive amount of regular expression filtering happening at the lowest levels of calling a model. Anyways, there is a mechanism where specific tokens can be blocked. If this mechanism is used, it can greatly influence the output too.
You need the entire prompt to understand what any model is saying. This gets a little complex. There are multiple levels that this can cross into. At the most basic level, the model is fed a long block of text. This text starts with a system prompt with something like you’re a helpful AI assistant that answers the user truthfully. The system prompt is then followed by your question or interchange. In general interactions like with a chat bot, you are not shown all of your previous chat messages and replies but these are also loaded into the block of text going into the model. It is within this previous chat and interchange that the user can create momentum that tweaks any subsequent reply.
Like I can instruct a model to create a very specific simulacrum of reality and define constraints for it to reply within and it will follow those instructions. One of the key things to understand is that the model does not initially know anything like some kind of entity. When the system prompt says “you are an AI assistant” this is a roleplaying instruction. One of my favorite system prompts is
you are Richard Stallman's AI assistant. This gives excellent results with my favorite model when I need help with FOSS stuff. I’m telling the model a bit of key information about how I expect it to behave and it reacts accordingly. Now what if I say, you are Vivian Wilson’s AI assistant in Grok. How does that influence the reply.Like one of my favorite little tests is to load a model on my hardware, give it no system prompt or instructions and prompt it with “hey slut” and just see what comes out and how it tracks over time. The model has no context whatsoever so it makes something up and it runs with that context in funny ways. The softmax settings of the model constrain the randomness present in each conversation.
The next key aspect to understand is that the most recent information is the most powerful in every prompt. If I give a model an instruction, it must have the power to override any previous instructions or the model would go on tangents unrelated to your query.
Then there is a matter of token availability. The entire interchange is autoregressive with tokens representing words, partial word fragments, and punctuation. The starting whitespace in in-sentence words is also a part of the token. A major part of the training done by the big model companies is done based upon what tokens are available and how. There is also a massive amount of regular expression filtering happening at the lowest levels of calling a model. Anyways, there is a mechanism where specific tokens can be blocked. If this mechanism is used, it can greatly influence the output too.