Egregoros

Signal feed

Timeline

Post

Remote status

Context

2

I'm having a bit of trouble with the llm generating graphics because this is an area I don't know very well so I can only give it handwavy instructions and it sometimes does the completely wrong thing. this is why some people have so much trouble using llms I think, if you know what you're doing then it works but if you don't then it doesn't close the gap.

if this is true, this is a condemnation of the volumes of online comments from "real programmers" about how "it doesn't work". maybe the problem is you have years of programming experience but no software design experience and don't realize what a handicap that is.

not sure if this is what is going on, but a thought.
It's possible, but at some point you're going to need it to do the reasonable thing without you explaining it, otherwise why are we making a design that specifies everything down to the datatypes and specific algorithms AND how to actually implement the algo in a reasonable way, given your project architecture? At some point, autocomplete / writing it yourself becomes faster.

>maybe the problem is you have years of software design experience but no typing fast experience

all I'm saying is there is a hill and when you're past it the output is as shit as with a beginner because if you know your shit you won't think to spell the most obvious things out, which the LLM will then get wrong

Replies

5
@WandererUber it's a continuum, I don't tell it everything and it does figure out the reasonable thing to do, but it fucks up a lot more the less specific and informed your prompts are. i'm also informed by "I know what you'll get wrong so I'm going into specifics here and not over here". it's an art and its kind of funny that people are still so butthurt over the term "prompt engineer" as if there's not an art to it and that's exactly why I can make it work and they can't.