Just learn how to code

bnew

Veteran
Joined
Nov 1, 2015
Messages
63,322
Reputation
9,672
Daps
173,076
I’ve personally noticed the code start to drift in implementation as it never tends to keep the same styling/formatting or “tone”/vocabulary over time. It does some subtle things that are different from month to month. Like why did it all of a sudden start changing naming formats? There are some “black swans” it tends to pull out of nowhere despite telling it to not make changes. It’ll change something, then you got to tell it to not change that from now on.

That’s only for personal projects. Now imagine multiple teams working on different parts of the codebase on projects with 5-10+ million lines of code meant to keep infrastructures running and working with each other. We’re not even talking the mega codebases. You want humans to be monitoring stuff like that (it’s fine for A.I to write it, but human professionals need to verify and have test cases). Blindingly trusting an A.I for that is asking for a lot of death. I’ve seen these things mess up a fair amount if nested else if and modular logic to be very thorough double checking the output.

I like it as a tool right now for small to moderate complex coding, but I’ve had a good amount of mixed experiences as well.

your complaints sound like my experience with chatgpt from over year ago possibly longer which is one of the reasons i don't use chatgpt for coding as often because it did as you described even after the second message. it seemed to have gotten better but i still noticed some issues. when i started using different models I noticed which ones would actually follow instructions and retain info better. a lot of of the code i ended up using in the past two years was generated by llama3 and 3,1 models and I sometimes tweak/fix some complex functions with claude or chatgpt.

if AI researchers tackle the generate the same image X amount of times test like they did the number of R's in strawberry than i think the fix could translate to better models for coding.

the closest example of something i've done to that which isn't about coding is formatting a transcript with punctuation, paragraphs and sometimes multiple speakers but it's been a few months since i've done that. I've seen sentences completely rephrased with some models when given that task.
 
Last edited:
Top