Login Page - Create Account

Support Board


Date/Time: Tue, 26 Nov 2024 08:56:47 +0000



Post From: ACSIL code generated by Chat GPT

[2023-12-22 08:00:52]
j4ytr4der_ - Posts: 938
I'm just using plain old free chatGPT. 3.5 I guess it is? Main thing is you can't expect it to spit out a fully complete study. This just won't work. It will make up imaginary functions and capabilities that sound very convincing but don't even exist.

What works well for me is having a pretty clear idea of what I need to do, and all the logic behind it worked out. If I can either write or find some code that is related or near to what I want, even better (I've used the included source codes from SC studies as starting points a bunch of times now). As you can see in the chat I pasted above, I fed it some code, then had to massage it from there. But you're right, it will start just giving you fragments, and if you don't realize that, you'll get really messed up. You can always say "Now give me the complete code" and it will give you everything, so that's worth considering.

Really you have to treat it as an assistant. It won't do the entire job for you, but it will bang out code faster than you probably can unless you're a highly skilled coder already, in which case you're probably faster just writing the thing anyway.

Speaking of documentation... I regularly give it the entire function definition from the docs when I'm doing something, since it often gets things wrong or tries to do something a very "non-SC" way. So it will do its own thing with some imaginary function, or sometimes just with stock C++ stuff that isn't how SC would do something. So I'll copy & paste the relevant section of the docs and just say "Here is the docs for the xxxxxxxxx function, try again using this" and it will take a new approach based on what it has now "learned".

Anyway, I love it but it's definitely more to get me a good part of the way there, and not to just spit out a working study. Everything has to be hand tweaked quite a lot, but even then I'll go back with my tweaks when I'm stuck on something and just tell it what I'm stuck on, and paste it the relevant code section I'm working on - giving it a little context about the goal - and it has been very helpful. I do find that on long conversations it starts to really bog down and get confused, so when that starts happening I'll often start a new conversation, explain just enough so it has context for what we're now doing, and just kinda pick up where we left off. Then it starts working well again (until it bogs down, again...)

But yeah, where this will be in a few years is going to be insane.