@freyfogle @icing@chaos.social are we even certain @icing is not an LLM ? I mean someone came to curl-up but that could have been a very well trained crisis actor ...
I use them regularly for tasks that are adjacent to my main coding activity, never for that main activity. I think it's because I know and remember very well the apis I use regularly. But for say, sending an email using the gsuite python API, AI will do that for me.
I find GitHub Copilot to actually be quite helpful. To me, it was adjusting my mental model of it being some magical tool to it being better autocomplete. It's garbage if you're relying on it to code for you. But if you're using it to write out that obvious for loop coming up, it just saves keystrokes.
When doing boring things, I've found a small amount of benefit by using as fancy auto complete. Like, in my case, we're talking boilerplate code, boring patterns. Basically, I only trust it to write code that I know exactly line by line what I wanted to write. Which, because writing code is such a small part of programming compared against thinking about what you want to write, it's only a marginal improvement.
I tested that CoPilot thing for a little while and found it unimpressive. It works for mundane tasks like generating snippets, but less effective than using macros.
The one area were it showed any promise was for giving a summary/explanation of code - with the caveat that it has no context awareness and would also just make up stuff at times.
I don't use them either. an intern was doing some JS coding the other day. he relies on copilot. heavily.
we had a hug and he was unsure how to fix it. I don't do JS myself normally, so I was reading the docs on the side. he used copilot. we arrived at the same solution at the same time, except I now knew how that particular framework intended for this to be done. I had learned something. pretty sure our intern had not yet truly understood the code at that point.
@flowchainsenseisocial which then implies you think so, so I'm asking what exactly you think I'm missing out that you think would improve my developer life? I am genuinely curious.
I also don't test new Linux distros, I don't test new text editors, I don't test new compilers or debuggers etc. For a reason.
Days have limited time and I need to decide how to spend it. By assessing "is it worth it?" "does it look like it would improve my life?" "will it make me more effective?"
If I don't think any answer would be YES, then I don't try it.
I've used LLM (I can't bring myself to use the term AI) in coding. It can help speed up some aspects. More often than not though, I've worked with junior programs who have used LLM and didn't have the experience necessary to understand why the code it was generating wasn't ready for prime time-made too many assumptions of inputs, didn't handle all cases, used poor data constructs, only partially handled errors. Hey it worked once-let's ship the code! See what I just accomplished!
Same here! What I have seen of others using things like copilot only makes me feel sad. It looks so disruptive to have snippets of code that most of the time is either completely irrelevant or broken pop into view while trying to formulate ideas.
for development i found it makes me slower, i prefer fixing my own bugs 😀. I asked it to review some of my code, and this can be useful if you’ve been sloppy. Perhaps there already are ai linters, that could be useful. But really, you need so many iterations it just slows me down..
like, I've tried. It's far worse than my junior dev ass, so I'm not sure what benefit I'm supposed to get from it. I'm fully capable of finding outdated libraries and stack overflow posts from 2013, thanks?
can AI tell us what the client will want after they're done changing their mind three times? Writing code is the fun part of the process, automating production of mediocre code doesn't make anything easier or better.
I recently gave it a try. The output was quite good for what I needed: a backstory for a random NPC my players met on their adventure in dungeons and dragons 🧌
I successfully used ChatGPT 4o (3.5 was awful) to help me develop a simple-ish Android app in an extremely time-constrained setting. I can 100% tell you that had I been experienced in Kotlin and Android development, I would have done it in a fraction of the time. AI is not taking away dev jobs, just natural learning, as I learned nothing in comparison to me having built it by myself.
my son is learning to code and he uses it as a teaching tool. The LLM gets him in over his head all the time, but he patiently researches the pieces he doesnt understand (which is a LOT), nearly always rewriting for cleaner code or better interoperability or sometimes because he doesnt have any clue what the code does...I just dont see "AI" as a dev tool, but a tool to learn to read code? Oh yeah.
not sure if predictive text and tooling around karaoke counts as AI, but yesterday I sang a live set of Karaoke where the song Somebody That I Used To Know was syllable matched to negative reviews of the board game Candyland from board game geek dot com.
They feed those tools answers from SO don't they? Do they program it somehow to recognize when the answers are wrong? Do they take age into account? When I ask it to hack me out some C++ metaprogram is it going to give me a constexpr based one or some C++03 TMP monstrosity? I used to write those.
I mean...I don't use the thing because I care if my code works and isn't ancient on birth. 😛 I also rarely actually GOT answers when I actually asked instead of answered. Yeah, about tha
I think you are correct at this point. Copilot offers me what I want 5% of the time. 45% I see where it’s going but no. 50% is just garbage, even going against intellisense. Overall that makes me slower. Claude is supposed to be better.
Ed Freyfogle
in reply to daniel:// stenberg:// • • •daniel:// stenberg://
in reply to Ed Freyfogle • • •Stefan Eissing
in reply to daniel:// stenberg:// • • •Jim Fuller
in reply to daniel:// stenberg:// • • •Cyrille Cailac
in reply to daniel:// stenberg:// • • •Pink
in reply to daniel:// stenberg:// • • •lachlan
in reply to daniel:// stenberg:// • • •🔥torstein🔥
in reply to daniel:// stenberg:// • • •I tested that CoPilot thing for a little while and found it unimpressive.
It works for mundane tasks like generating snippets, but less effective than using macros.
The one area were it showed any promise was for giving a summary/explanation of code - with the caveat that it has no context awareness and would also just make up stuff at times.
Marcus
in reply to daniel:// stenberg:// • • •I don't use them either. an intern was doing some JS coding the other day. he relies on copilot. heavily.
we had a hug and he was unsure how to fix it.
I don't do JS myself normally, so I was reading the docs on the side. he used copilot. we arrived at the same solution at the same time, except I now knew how that particular framework intended for this to be done. I had learned something. pretty sure our intern had not yet truly understood the code at that point.
Michael
in reply to daniel:// stenberg:// • • •Using AI tools for coding is just switching from "typing the code" to "reviewing the code" ...
You all review the AI generated code, don't you? 🙃🤪
FlowChainSensei
in reply to daniel:// stenberg:// • • •daniel:// stenberg://
in reply to FlowChainSensei • • •FlowChainSensei
in reply to FlowChainSensei • • •daniel:// stenberg://
in reply to FlowChainSensei • • •daniel:// stenberg://
in reply to FlowChainSensei • • •I also don't test new Linux distros, I don't test new text editors, I don't test new compilers or debuggers etc. For a reason.
Days have limited time and I need to decide how to spend it. By assessing "is it worth it?" "does it look like it would improve my life?" "will it make me more effective?"
If I don't think any answer would be YES, then I don't try it.
Lyricismist
in reply to daniel:// stenberg:// • • •Mihugo
in reply to daniel:// stenberg:// • • •Desert Plains
in reply to daniel:// stenberg:// • • •Kristoffer 🍄
in reply to daniel:// stenberg:// • • •n-gons
in reply to daniel:// stenberg:// • • •Thomas Svensson 🖖
in reply to daniel:// stenberg:// • • •Crafty
in reply to daniel:// stenberg:// • • •😀🚲
in reply to daniel:// stenberg:// • • •Janne
in reply to daniel:// stenberg:// • • •I google all day all the time...
Local model in the IDE saves so much typing. Most of code is boiler repetitive boiler plate anyways.
The important bits aside.
Benedikt
in reply to daniel:// stenberg:// • • •Alonely0 🦀
in reply to daniel:// stenberg:// • • •lemgandi
in reply to daniel:// stenberg:// • • •relogi
in reply to daniel:// stenberg:// • • •Epu
in reply to daniel:// stenberg:// • • •not sure if predictive text and tooling around karaoke counts as AI, but yesterday I sang a live set of Karaoke where the song Somebody That I Used To Know was syllable matched to negative reviews of the board game Candyland from board game geek dot com.
Game changer for developing comedy!
https://botnik.org/creations/
j5v
in reply to daniel:// stenberg:// • • •Artur Neumann
in reply to daniel:// stenberg:// • • •crazyeddie
in reply to daniel:// stenberg:// • • •They feed those tools answers from SO don't they? Do they program it somehow to recognize when the answers are wrong? Do they take age into account? When I ask it to hack me out some C++ metaprogram is it going to give me a constexpr based one or some C++03 TMP monstrosity? I used to write those.
I mean...I don't use the thing because I care if my code works and isn't ancient on birth. 😛 I also rarely actually GOT answers when I actually asked instead of answered. Yeah, about tha
Matt Fozard
in reply to daniel:// stenberg:// • • •