[superasn]

Yes I've switched to chatgpt for all my programming questions nowadays because of many reasons:

1. It understands the versions more clearly, it won't generate the code for Bootstrap 2 when my question is about Bootstrap 5

2. Asking follow up questions is easy. So the code didn't work. I tell it that and instantly it tells me what I need to do to fix it (leave missing config file, etc) or gives me an alternate.

3. Its answers instantaneously except when their site is down.

4. Its more customised as the answer is not a generic question posted an year ago.

5. It spans multiple domains of knowledge, so if lets say I get an error curl.so not found, it can tell me what I must run on the command line to fix that on my system too.

6. Its so much faster because there are no stupid questions. It's just like typing on google vs SO where you need to proof read and make sure its not a duplicate.

[keyme]

Chat gpt will never get smarter though (learn anything new), unless actual people use stack overflow to provide the training data.

The most important comment on SO is the one where the OP says "I did this and that, and can confirm it's working". There is no such data when you only use the chat bot. It's a major problem. Perhaps the solution is a more tailored "stack gpt", where the "conversations" are published along with the human responses.

[scarface74]

ChatGPT is already “smarter”. I can ask ChatGPT for the answer to my very specific problem and it can answer it snd give me code.

For instance, I needed to write a simple script that gave me all of the roles containing a list of policies where you can specify multiple policies from the command line using “-p” multiple times.

If I look it up on Google, I get this response which is close enough

https://stackoverflow.com/questions/66127551/list-of-all-rol...

And I still have to make slight modifications.

Of course I could write the entire thing myself just by looking at the boto 3 docs.

But instead I told ChatGPT

“ Write a Python script that returns a comma separated list of arns of all AWS roles that contain policies I specify with the “-p” parameter using argparse”

Then I told it, “that won’t work with more than 50 roles”. It then corrected the script and used a paginator.

Yes, I had to know enough to recognize the bug. But I count have written or modified a Python script that fast and I write boto3 based Python scripts all of the time.

I definitely could have told it that the first version wasn’t returning all of the results and it would have corrected itself like it did and added a paginator

[helsontaveras18]

They know this is a problem, so they’re hiring contractors to write code and label it to teach the AI.

[superasn]

I think the thumbs up and down button it shows on each response can be good for that. I think it can already classify similar questions so if it consistently gets a thumbs up or thumbs down it can become smarter too.

[ryanackley]

There is a chicken-egg problem with using ChatGPT for domain specific technical knowledge. ChatGPT is trained on data from sites like SO. If everyone starts using ChatGPT instead and technical communities die off, it seems like ChatGPT would quickly become ineffective.

You would still have documentation published by originators of the technology but a lot of programming is figuring out the quirks of things that aren't in the documentation or getting past bugs.

[devjab]

I genuinely find ChatGPT very useful for a lot of tedious tasks, but it's also very obvious that it's data cutoff point was in 2021. You can get it to write you an abstract generic Odata controller class in C# better than most searching on google, but it'll be with IActionResult and not ActionResult, sometimes it'll even be with Task<Iqueryable>, which is even older and it'll do this extremely confidently. Which is sort of fine, it'll work after all. It's also not fine, because there is a reason it's moved on to become ActionResult performance and ease of maintenance being some of the reasons.

Though to be fair, it's not like Stack Overflow, or even some of the less updated parts of the official documentation, will do any better than ChatGPT, but with ChatGPT you don't get a date on the knowledge and you don't get the comments from other people telling you that it's wrong or outdated. For me personally, there is also some thing about its confidence that makes me "trust" it more than the internet that I've spent 20+ years not trusting. I'm not sure if that's just me or even why that is exactly. I'm fully aware that the language model is basically just the internet, and still I believe it? I'm happy the first thing I asked it was on a subject I knew a lot about so that I could see straight away that its answer was very outdated. Because if it had been on a subject I didn't know much about, I'm not sure I would have even found out its cutoff point was in 2021. I only learned that fact because the answer it suggested was with a library that I knew was abandoned, to which I asked if it knew that, and it told me when it had stopped "learning".

[roncesvalles]

I find it hard to believe that ChatGPT is useful beyond trivial things or giving you a lead that you can follow on Google. It's very frequently just plain wrong.

[primax]

If you invest time in improving your skill at prompt engineering you’ll find it quite worthwhile.

And yes, often I do follow up google searches for some things - but it has already saved me days of research time compared to Google

[abigail95]

i'm wrong about most things most of the time. i'm pretty sure i'm still useful.

[roncesvalles]

Well even wrong answers have utility in a forum. When someone posts a wrong answer on Stack Overflow, Reddit or in real life(?), it baits a bunch of replies describing in detail just how wrong that answer is.

ChatGPT has the user interface of an oracle sitting alone in a cave in the mountains, but it isn't as accurate as an oracle is expected to be. That makes it useless to me.

[ricardo81]

I may be wrong but isn't chatGPT essentially trained on a web dataset that includes SO? I remember reading something about Common Crawl being used.

[bruce343434]

I would be interested in knowing why you are getting downvoted. Much of your experience rings true, except for me chatgpt tends to generate bogus code (though pointing me in the right direction).

[bitcharmer]

7. It doesn't ask pretentiously why would you even want to do that and opine on this being an anti pattern.

I'm so sick of these people on SO.

[ComplexSystems]

I was literally about to comment this as well. Actually, that's the biggest plus for ChatGPT right now.