When i started, I tried instructing the AI to engage a concept in a single shot.
providing an algorithm to guide it.
That works in many ways.
However - it fails in novel aspects.
I have found that you have to have a discussion with the AI, fully explaining an idea, until it understands it.
It seems there is an iterative process of understanding - which ironically, makes sense when you consider… thats what humans do.
Im not suggestion it thinks like a human, no…
I am suggesting that it talks to you, like someone else youve talked to.
Heres my measure of understanding.
: IF… it can paraphrase, and create an accurate visualization of the concept you are trying to convey, then… you can be confident it “understands” enough to proceed with any more advanced thought.
But of course, this is my perspective and my understanding.
This is just a curiosity, not a peer reviewed paper.
BTW.
I recommend using 4k, 120 camera, uncompressed to screenshot anything important these days.
Yes, this is an understood technique, “warming up” the AI to orient it to the task at hand, just as much as multi-shot to demonstrate answering questions with the desired quality.
Showing the AI the “assistant” acting as an expert is better than just telling AI that it is an expert.
If you need to program in Go libraries or things at the edge of the AI knowledge, having the AI write documentation answers about the subject with increasing specialization is just as transformative as finding and pasting some documentation manually, and less effort.
BTW, I recommend (ALT) + (Print Scrn) to screenshot single application windows.
Rather that I remember everything I wrote and don’t want to type again.
Indeed if I could just download into others brains, or make a chatbot with all the answers that are somewhat relevant to something that will be asked again, then I wouldn’t have to continually demonstrate “here’s what you could have found on the forum”.
Except I can’t fault someone for not being able to find anything on this forum or its search function. Even a search (where a search for just me is "search terms @_j") where I try to find what I know I already posted just weeks ago is frustrating.
Each thing you say is building up a longer and longer “context” that will ultimately be responsible for the answer quality. When asking an extremely detailed question, I’ll normally start by asking if it understands the topic, and then narrow down a little at a time to get closer to the actual question, to be sure it’s “following me” the entire time. So by the time I ask the actual hard question I know it’s thinking in the right context.