Over the past one and a half years, Stack Overflow has lost around 50% of its traffic. This decline is similarly reflected in site usage, with approximately a 50% decrease in the number of questions and answers, as well as the number of votes these posts receive.
The charts below show the usage represented by a moving average of 49 days.
What happened?
Yeah it gives you the answers you ask it to give you. It doesn’t matter if they are true or not, only if they look like the thing you’re looking for.
An incorrect answer can still be valuable. It can give some hint of where to look next.
@magic_lobster_party I can’t believe someone wrote that. Incorrect answers do more harm than being useful. If the person asks and don’t know, how should he or she know it’s incorrect and look for a hint?
In the context of coding it can be valuable. I produced two tables in a database and asked it to write a query and it did 90% of the job. It was using an incorrect column for a join. If you are doing it for coding you should notice very quickly what is wrong at least if you have experience.
I don’t know about others’ experiences, but I’ve been completely stuck on problems I only figured out how to solve with chatGPT. It’s very forgiving when I don’t know the name of something I’m trying to do or don’t know how to phrase it well, so even if the actual answer is wrong it gives me somewhere to start and clues me in to the terminology to use.
Google the provided solution for additional sources. Often when I search for solutions to problems I don’t get the right answer directly. Often the provided solution may not even work for me.
But I might find other clues of the problem which can aid me in further research. In the end I finally have all the clues I need to find the answer to my question.
How do you Google anything when all the results are AI generated crap for generating ad revenue?
Well then I guess I have to survive with ChatGPT if the internet is so riddled with search engine optimized garbage. We’re thankfully not there yet, at least not with computer tech questions.
In my experience, with both coding and natural sciences, a slightly incorrect answer that you attempt to apply, realize is wrong in some way during initial testing/analysis, then you tweak until it’s correct, is very useful, especially compared to not receiving any answer or being ridiculed by internet randos.
Well if they refer to coding solution they’re right : sometimes non-working code can lead to a working solution. if you know what you’re doing ofc
Even if you don’t know what you’re doing ChatGPT can still do well if you tell it what went wrong with the suggestion it gave you. It can debug its code or realize that it made wrong assumptions about what you were asking from further context.
How is that practically different from a user perspective than answers on SO? Either way, I still have to try the suggested solutions to see if they work in my particular situation.
At least with those, you can be reasonably confident that a single person at some point believed in their answer as a coherent solution
That doesn’t exactly inspire confidence.
Better than knowing there’s some possibility that the answer was generated purely because the sequence of characters had the highest probability of convincing the reader that it seems correct based on the sequence of characters it was given as input (+/- a decent amount of RNG)
Still debatable, IMO. Human belief is stubborn and self-justifying whereas an RNG can be rerolled as many times as needed.
Yeah but if you keep rerolling the RNG, how do you know when a right answer gets randomly generated?
Also, my point above was that if a human believed the solution was true, it probably was true at some point. With generative language models, there’s no guarantee that there’s any logic to what it tells you.
You know when the code compiles and does what you want it to do. What’s the point in asking for code if you’re not going to run it? You’d be doing that with anything you got off of Stack Overflow too, presumably.
What point are you trying to make? LLMs are incredibly useful tools
Yeah for generating prose, not for solving technical problems.
You’ve never actually used them properly then.
One example is writing complex regex. A simple well written prompt can get you 90% the way there. It’s a huge time saver.
It’s great a writing boilerplate code so I can spend more of my time architecturing solutions instead of typing.
the good thing if it gives you the answer in a programming language is that its quite simple tontestvif the output is what you expect, also a lot of humans hive wrong answers…