Instruction-files: A way of expanding GPT instructions capabilities
"Instructions-files" could be a way of naming an alternative use to 'Knowledge Files' on ChatGPT-4. These may be particularly useful in cases where achieving consistent obedience to Instructions is challenging.
Most GPT-creators may have noticed by now a poor performance of their custom models when attempting to run complex AI Instructions, either because they are not followed, only partially followed, or followed inconsistently across new interactions. In all of these cases resulting in a really time consuming iteration and frustrating experience for the creator.
Truth be told, I am not familiar with the exact technical concept surrounding this specific limitation, yet it appears clear from just brief experimentation that there is some sort of context limitation linked to the complexity of the instruction per se regardless of its length. This seems to go beyond the "token-limitation" from an instruction's length perspective to some sort of instruction’s “complexity” dimension.
What I can vouch for are my repeated efforts in rephrasing Instructions through many trial and error attempts on a quest to figure out how to get a consistent obedience of the AI agent through multiple independent conversations. This becomes remarkably accentuated when 1) trying to chain different actions to get within a single AI output, 2) expecting a specific sequence of outputs, or 3) combining different IFs and ELSEs based on the user's input, let alone all those combined. From my perspective, These conditions seem to overload the system's ability to process instructions, making it unreliable.
Using Knowledge files to offload complexity from GPT Instructions
After a long struggle about thinking how to cater for a specific experiment, I started trying to remove some of the AI Instructions that were most complex (those that have been more prone to failure during my several trial and errors) and added them as literally plain text Instructions on a .txt file. This was my attempt to use those .txt as if they were some sort of “configuration files” that I would load as Knowledge documents and reference their file names within the Instructions field to replace the “offloaded complex AI instructions” . The results of this little experiment were promising as I was able to execute the same degree of complexity (and more) in a stable manner that before was just not working.
From then on, I started taking a different approach and orchestrating the expected "instruction flow" through the Original Instruction but defining the steps in each step of the flow through text files uploaded as Knowledge files, or what I would call as “Instruction files”.
Let's take a look at the example:
The trade-off
Plainly, the AI response comes out way slower. Once again flagging my lack of technical expertise, yet what I can tell based on several observations is that the chain of responses becomes less agile since it is also requiring the GPT “Code interpreter” functionality to be enabled as well as an “unpacking” of those Instructions files when “summoned”.
Opinion: I may be overcomplicating things
By overdictating the output of this great generative AI product it seems like I am trying to use something that conquered the world for being surprisingly “autonomous and adaptive” as something rather “limited and process-based”. I may be using this wrongly and some people may be right to think that I am underusing technology and the resources behind it just to run some simple chain of commands that a basic program with simple code would make. Therefore, it would make more sense to learn how to program in one of the many codes out there than using this method.
I totally agree, but for me this is some atomic evidence that even with no programming background this technology understood what I wanted to achieve with my own cognitive code. If I can find some value here I believe some other people may too. Will we be able to keep producing videos, pictures, audio faster than learning to express our requirements to great generative AI? Could this someday happen in programming?
"Instructions-files" could be a way of naming an alternative use to 'Knowledge Files' on ChatGPT-4. These may be particularly useful in cases where achieving consistent obedience to Instructions is challenging.