Do y'all really use ChatGPT for coding questions? Every time I have tried, the answer I get is entirely wrong, and I end up going back to either first party documentation, or Stack Exchange.
... that it doesn't work exactly "as printed" when copy-pasted into an IDE? Or that it fundamentally misses the intent of the question? Or that it got some important key details right and messed-up other things?
Usually, in my experience, it's the last one. Often all you need is a clue to get you thinking in the right direction. Interactively asking chat-gpt questions and critically evaluating it's answers has been far more helpful than I could have ever imagined.
Does it ALWAYS help? No.
Is it an exciting and emergent alternative to what we've had before? Hell yes.
The first party docs are, of course, the gold standard. But they're often far more detailed and canonical than folks are ready to absorb.
SO can be helpful IF you can find your question already there and the answers are still relevant, or if you manage to ask your question in exactly the right way and the duplicate-police haven't shut it down (regardless of whether it's actually dupe). Or if a kind soul has provided a clue for you in a comment as your question was downvoted into oblivion.
Another option, IMHO, is github issues/discussion. The library developers are often there and generous about helping people out (without the perverse and infantile incentives of gamification that afflicts SO). But this option should really only be reserved for very carefully asked questions. In a way it has a higher bar than SO (but without the negative reinforcement of downvotes).
Of course I do. I use it (GPT4) every day and it's a million times better than Stack Overflow ever was. I find it extremely strange that this is not the standard now.
And no, the answers are correct 99.999% of the time, so I don't know what the hell you're doing.
If it counts, I have a SO account with >15k "reputation."
It really depends on what you are doing, in my experience. I often ask for code when I know there will be a decent amount of difficult to remember boilerplate. It will do a decent enough job to save some typing and get me a framework to work within.
It is pretty good at writing little DTOs based upon a sample JSON too.
But harder questions for poorly documented libraries will exhibit a lot of the same challenges as Googling.
Same experience: if you're trying to do something unusual (an obscure library, or a task that has minimal documentation and few usage examples in the wild), the results are usually bad, and it will very frequently try to call functions which don't exist.
My main use case for GT4 is simple/concise questions for common tooling: namely Docker, Git, Vim (chatting with a robot about Vim during long builds is kinda embarassing now that I think about it), Bash scripting and common Linux utils / config.
I am able to give a lot of information about what I need, and I'm experienced enough to know precisely what to ask for. There's also a huge amount of training data about these tools. I'm also experienced enough to be able to evaluate the output for correctness. I would never ask anything like "write a program in Python for <blank>, at this point.
Sure. Currently using it to help me figure out the data format liuliu[0] is using for their "Draw Things" SQL database, because I want to bulk export all my previous images and my entire knowledge of SQL can be written on half a postcard.
I use it as a starting point. I don't think I've ever actually used its answer directly, but I use it most days now to give me the skeleton of something I can then get working the way I want it from there. I've always been better at editing than filling in a blank page, so this has turned out to be a good fit for me.
There are some strategies to get it to work better. It’s great if you tell it to create a function that accepts (input) and produced (output). For example I recently had it come up with some date formatters that saved me a lot of time, eg, transform a date from MM-YYYY to “X months ago”, stuff like that.
>some date formatters that saved me a lot of time, eg, transform a date from MM-YYYY to “X months ago”
Obvs your use case may be wildly different, but I detest ms office for emails: 'received three days ago...Two days ago...'. Great, now i have to remember what today is and then do some mathematical gymnastics (i dont work m-f).
... that it doesn't work exactly "as printed" when copy-pasted into an IDE? Or that it fundamentally misses the intent of the question? Or that it got some important key details right and messed-up other things?
Usually, in my experience, it's the last one. Often all you need is a clue to get you thinking in the right direction. Interactively asking chat-gpt questions and critically evaluating it's answers has been far more helpful than I could have ever imagined.
Does it ALWAYS help? No.
Is it an exciting and emergent alternative to what we've had before? Hell yes.
The first party docs are, of course, the gold standard. But they're often far more detailed and canonical than folks are ready to absorb.
SO can be helpful IF you can find your question already there and the answers are still relevant, or if you manage to ask your question in exactly the right way and the duplicate-police haven't shut it down (regardless of whether it's actually dupe). Or if a kind soul has provided a clue for you in a comment as your question was downvoted into oblivion.
Another option, IMHO, is github issues/discussion. The library developers are often there and generous about helping people out (without the perverse and infantile incentives of gamification that afflicts SO). But this option should really only be reserved for very carefully asked questions. In a way it has a higher bar than SO (but without the negative reinforcement of downvotes).