How useful have you found ChatGPT and other similar websites to be for coding?

G

gourav

Skilled
Over the past few days, I've been learning Vuejs and I've been asking ChatGPT to generate code for it. It does a really good job of it.

Today I took a small project I'd done in Python and asked ChatGPT to translate it to Javascript. It gave me working code for all the functions. I don't know Javascript well and this made it really easy to get things working. I can literally write any code in python and get it translated to Javascript as long as it is not using any libraries.
 
I was watching a Chris Titus live stream yesterday, and he had a basic script in bash, and he asked GPT to convert it to Zig, and it did, but the code didn't work(understandable since there isn't that much Zig code out there compared to rust) after a bit of tinkering he asked GPT to convert the broken code to Rust and after a bit of debugging the code worked.

And as for me, YES I find them very useful for understanding something and being able to ask follow-up questions or when I try to explain something to GPT to see if I'm on the right track. Whereas if it was a real person, I would hesitate, thinking it's a stupid question. Also, line by line explanation of algorithms and data structure is very helpful for me. Obviously, I crosscheck most of the stuff but still very helpful.
Coding related questions and letters/email are probably my top use case.

Can't speak on using GPT as a coding assistant because I try not to use it for auto completing a whole function. But I have tried this VS Code extension called Codeium just to test it out, it's like GPT but just built into VS Code, and it isn't as obstructing as copilot.
 
Last edited:
I would say very useful, My needs were basic so I cannot comment on advanced programming.

I knew C but had zero knowledge about Windows programming (esp the GUI) & unicode C programming.

Within a day I developed a pretty decent Windows unicode C program with a GUI and all.
Program to change drive letters, BTW.

I got ChatGPT to show me, step by step, code for each function like combo-box for the drives, then how to change a drives letter, etc.
Without ChatGPT it would have taken me ages.

I don't have visual studio, I use w64devkit. I'm a great fan of portable software.

it's also v good with AutoIt & AutoHotkey.

Capture_2024-05-19_011843.jpg
 
ChatGPT, Claude and Bing AI all 3 of them have been very helpful. I have zero knowledge of coding so while asking questions related to excel/google spreadsheets ChatGPT would give an alternate solution based on Python so i tried Python and now almost all the heavy work is done using codes Thanks to these websites. I found Claude to be better than ChatGPT when it comes to coding (I am using free ChatGPT and not the paid 4.0)
 
ChatGPT is good as a starting place. If you want to modify or change the data in anyway, you can be verbose with it and ask what you want the end result to be. Usually it will use the appropriate function / module and give a script template.
Of course, sometimes it gets output objects and stuff wrong, at that time just ask it to modify the script or provide an alternative. Most of the tidyverse and dataframe related stuff is present in the model.
However, if you have no coding background, ChatGPT will lull you into a false sense of security, making you believe you can get things done in Python/R or whatever - just do your due diligence before using the code and you should be fine.
 
It's good, definitely a step up over stackoverflow, going through documentation, and googling. Especially for languages/frameworks you are not familiar with.

But even slightly more complicated stuff, and they make a simple mistake in spite of asking it to avoid that mistake. That mistake often makes their output very useless, and I had to write a lot of the code myself. I admit I was getting greedy in asking the AI to do somewhat more complicated stuff.

This includes ChatGPT, Bard, and GitHub copilot plugin for vscode.
 
Just an aside : I noticed ChatGPT consistently making a mistake with the swprintf function. I asked ChatGPT if I could point it out. It said certainly. I did, and from then on I got swprintf correctly.
 
ChatGPT phenomenally speeds up RAD -- provided you know what you are doing. Saves one from opening 10 Stackoverflow pages, and get the answer quick. Example: I needed to figure out to read top ten updated files on an S3 bucket. It gave the loop version, which I asked to convert using recursion. Worked with a bit of modifications. I easily added to my boto3 class object as a method. 2 minutes....tops.
But on the flip side. I was using StandfordNER NLP which is a Java code, via the NLKT library in Python. It randomly times out post token. Even increasing memory heap did not solve it. ChatGPT had no working answer. Same with an issue I had with a deep learning model in R. Kept erroring out. No viable ChatGPT solution.
 
Last edited:
I (non-coding background) was doing basic python coding, df mainly, and it used to take several days to get something to work. Major headache was datetime.
Cgpt reduced it to a few hours but there are times when cgpt goes on a tangent, stackoverflow / excel forum rescues.
 
So, so far i have been staying away from AI stuff, continuing to use old way of google/stack exchange when needed.
My projects are already mature so did not feel the need yet.

But maybe i should explore this, esp for new things. Where exactly do you guys ask stuff? Is there any tutorial you recommend to get started ?
-------
I do hope that sites like stack exchange do not die because of AI. Else it would be net loss for all, might be ok for old stuff but anything new and ai wont have any inputs.
 
it works pretty fine for basic cases but can't really handle anything complicated, we are using gpt 4 turbo in our org for semantic labelling of data and honestly its incredible just how accurate it is. As for learning programming using it, I'm much more ambivalent, its pretty nifty for learning stuff quickly and like others said, it saves time opening articles/documentation/stackoverflow, but I grew up usingthese resources and most of my learning came from them, specially StackOverflow comments, half the time I didnt even know what I really wanted to ask and you would only realize that once you actually went through those resources but for those who just want a working solution asap, its a godsend
 
It's funny how people usually overrate AI/LLM abilities in many, many areas but then get it wrong when it comes to the one thing they're actually insanely good at, which is absorbing and disseminating large amounts of information very quickly. It's just a information database with search, but a lot smarter than the ones you're used to. It won't replace most people and it certainly won't cause forums/stackoverflow to die. You're on one right now, aren't you? You still prefer to talk to humans when it matters.

It will actually result in a higher quality of information on, say, stackoverflow, because all the low-hanging fruit and obvious answers will be available to everyone, instantly. So humans can hone in on what they're actually good at. Solving real problems, not just ones borne from a lack of information or a reluctance to read 200 pages of docs.

Also, if you've ever tried to troubleshoot a problem, you know that half of the time the human answers will misdirect you. Because they're outdated or just plain wrong. LLMs have a better hitrate than that.
 
variablevector said:
It's funny how people usually overrate AI/LLM abilities in many, many areas but then get it wrong when it comes to the one thing they're actually insanely good at, which is absorbing and disseminating large amounts of information very quickly. It's just a information database with search, but a lot smarter than the ones you're used to. It won't replace most people and it certainly won't cause forums/stackoverflow to die. You're on one right now, aren't you? You still prefer to talk to humans when it matters.

It will actually result in a higher quality of information on, say, stackoverflow, because all the low-hanging fruit and obvious answers will be available to everyone, instantly. So humans can hone in on what they're actually good at. Solving real problems, not just ones borne from a lack of information or a reluctance to read 200 pages of docs.

Also, if you've ever tried to troubleshoot a problem, you know that half of the time the human answers will misdirect you. Because they're outdated or just plain wrong. LLMs have a better hitrate than that.
Click to expand...
that might be so in your experience but definitely not in mine, I have had many issues over the years where the solution wasn't obvious and it was only after reading stackoverflow/forum comments that I found out that my question and approach in itself was wrong and I could better structure in some other way and while you are right that it can disseminate a large info dump pretty quickly, you havent found any weird case specifically in your scenario that you need info to even frame your question?

almost 90% of the questions I need help with aren't even present in Stackoverflow, yet some old guy in an old comment would have mentioned something which is exactly what you were looking for, or the discussion that springs in the comments when someone posts an answer to the question, to me that is the good stuff, Chatgpt works perfectly when you just want a quick solution to, "How do I loop through this weirdo object structure", 'How do I filter out documents in Mongo which are only 10 days old or less', for anything thats not majorly a syntax solution its crap and thats what I need to deal with in my job day-to-day or atleast thats what I've seen over the years with me and my colleagues/friends
 
altair21 said:
that might be so in your experience but definitely not in mine, I have had many issues over the years where the solution wasn't obvious and it was only after reading stackoverflow/forum comments that I found out that my question and approach in itself was wrong and I could better structure in some other way and while you are right that it can disseminate a large info dump pretty quickly, you havent found any weird case specifically in your scenario that you need info to even frame your question?

almost 90% of the questions I need help with aren't even present in Stackoverflow, yet some old guy in an old comment would have mentioned something which is exactly what you were looking for, or the discussion that springs in the comments when someone posts an answer to the question, to me that is the good stuff, Chatgpt works perfectly when you just want a quick solution to, "How do I loop through this weirdo object structure", 'How do I filter out documents in Mongo which are only 10 days old or less', for anything thats not majorly a syntax solution its crap and thats what I need to deal with in my job day-to-day or atleast thats what I've seen over the years with me and my colleagues/friends
Click to expand...
I see, you mean the kind of situation where you're asking how do you do X with Y? And someone says, why don't you just use Z which is designed for that? The framing of the very question is misguided or wrong. I think that's a great case where it can fail compared to a human.

I wasn't saying LLMs can solve every problem. Or that it makes stackoverflow or experts in general redundant. I said it's an efficient way of searching a vast database of information. It can't solve all your problems for you. It is a very useful tool with limitations. And that's your job security. If LLMs could do what you're suggesting, almost everyone in a technical field would be a lot closer to losing their job.
 
variablevector said:
I see, you mean the kind of situation where you're asking how do you do X with Y? And someone says, why don't you just use Z which is designed for that? The framing of the very question is misguided or wrong. I think that's a great case where it can fail compared to a human.
Click to expand...
yeah, this is what I mean by LLMs being a god send for basic questions, LLMs shine when you can structure your prompt perfectly and when the question you ask is direct to the point but its rarely that simple, with just a single word change in the prompt, the answers can differ wildly, this is why I'm really not in support of using it for learning purposes. we have been experimenting with prompt engineering in our org and you should see some of the prompts we are using, its literally holding a child by its hand and walking it through step-by-step but the benefit to this is, once your prompt is good enough, LLM's can happily handle any semantic reasoning that you might need.

for e.g. you want to know the country name from a location, where the location can be a state/city or even some landmark, this would be really hard to do with traditional approaches but trivial for an LLM, specially if you can feed it more contextual data which is usually the case.
 
Back
Top