I'm a software engineer and I don't find much value in ChatGPT. It provides a bit of help when writing very specific and short pieces of code for languages I know but don't use often. I'd be curious to have actual data on how other SWEs use it.
I have privacy concerns with both ChatGPT and Copilot, but I also don't get the desire to use these tools on a daily basis. I am very adept at talking to a computer, and figuring out what it's doing/wants (though it may take a while). Trying to convince a language model to act as an inbetween for me just seems like a massive hassle.
It's like delegating work to a junior that is completely untrustworthy, but instead of you working to level up your junior and gain a useful coworker, you're forever stuck with the kind of dumb, needs simple things explained person.
I read posts like this and I wonder if I'm really missing out by not writing English paragraphs a lot of the time instead of code.
The type of question I used to type into google, now I ask ChatGPT. For a language I know well it tends to bring a 5 minute task down to 2 minutes. For a language I don't know well it takes a 30 minute task down to 2 minutes (One weird day the site for a framework was down and I couldn't read the docs, but ChatGPT was there with the right answer).
Yeah, and while it's true it can't really solve something you have to think about specific to your problem domain, I find it really good for common implementations. If I'm making a music app and I'm implementing shuffle, I just type
function fisherYates(arr: IPlaylistItem[]) { tab
and it's done.
Similarly, I just type
class SortedSet:
// api has to have add, rank, cardinality, itemsinrange(make this a fluent api with min max fns returning self each time, and a get to resolve it, and items() to start it)
// use redis
And it does the zrank, zcard zadd etc figuring out how to call the py-redis API.
Similarly I cba when learning an ORM to find out how specifically it does something by scouring it's docs, I just type SQL or natural language there or the beginnings of it atleast, and let copilot fill out the orm api for me. After which I can just hover over the functions it chained/nested together and the docs pop up and I verify that it did it right, and read any caveats they mention. I pretty much learnt how to use prisma just by seeing what copilot produced.
I plug it directly into my editor (via https://github.com/gsuuon/llm.nvim) and have it fill out code for me. I write what I want with comments and ask it to fill the rest - if it's straightforward enough it basically always works. I also get it to write commit messages (based on git diff) - though I need to improve my prompt a bit as it gets verbose and I end up rewriting it most of the time. I was working on trying to feed it things like hover and tree-sitter information before I got distracted, but that'd be another power boost as well whenever I get around to it.
