Does anyone still use Stack Overflow post-2022?
Such sites help you understand why, far more than a distilled down response of a LLM. Such sites help you learn.
Knowing the why, is often far more important than the how.
Them: “why doesn't this thing work I your OSS lib, whereas it works in this other OSS lib?”
Me: “does it? it's not supposed to.”
Then: “oh idk, ChatGPT told me so.”
I hate the use of this word to describe when a sentence completer gives a wrong result. To hallucinate you have to have a concept of reality. Which those models don't have. They're just mathematical tools which some times manage to give a result humans think as human-like.
I do it this way around because I can search far faster than I can write out reasonable english sentences that head-off bad LLM responses.
Sometimes when I'm confident enough that the LLM will be able to give me the "right answer" faster than search I'll do it first, but that's mostly when I have something I can copy/paste in.
But no, my experience is different with regards to the ratio of quality of input versus output. I think we all do a subconscious ROI of whether typing that little extra will yield an output that's going to be that much better that it was worth the time and effort.
Or much more likely I'm just lazy.
Now I find it more productive to ask an LLM to get some keywords or similar which I might search for afterwards.
The great thing about LLMs is that, unlike a search engine, you can tell it "no, not quite like that, more like this" to guide it.
ChatGPT has had built in web search for over two yeass for paid users
Besides, I guess I want to have some control over the search, like which sites to visit. Sometimes the search might retur something tangential, but interesting for example.
I still support some older .NET applications, and I’ve had the situation several times where I’ve ended up on SO via a Google search and started to read the answer, only to feel as if the writing style is familiar. Then noted that I myself answered it and I’m reading my own text for the solution.
I’m always amused that a) I keep having the solve the same things over and over again, and b) I can’t remember the solutions.
Maybe I’m just old.
Google -> stackoverflow is typically much faster than chatgpt.
If you are a fast typer and can formulate a short question then you can often get from thought to answer in well under 10 seconds.
With chatgpt there is a small thinking delay then often a long token sequence of filler before you get the few lines you're looking for.
ChatGPT shines compared to stackoverflow for something you know that requires a dialog of prompts or you're asking it to write some specific code instead of just finding out how to accomplish a small task with very clear syntax.
E.g. I wanted to search for "CUDA on Fedora 41" and Google "helpfully" drops the "41" number from the query!
So I have to notice that and click on "41" to force the search.
You get a lot of unhelpful errors during the DKMS build. Which you shouldn't even _have_ because of akmods. Will publish details in my blog once I get the MD renderer to cooperate.
Google is not consistent with its "suggestions". Funnily enough, your message is now near the top of the results.
However any contributions (answers, questions) are directly related to something I'm working on. There's no community (or certainty that your work will not be marked-as-duplicate) which would justify a more-general kind of helping-out.
Everything that came after the mod strike kind of made things worse anyways, so there's not that much change in knowledge requirements for me.
There's also kiwix and zim files available in case you don't wanna self host but still be able to have it offline on your desktop machine.
I haven't posted questions or answers for years. Too many petty, officious people and general toxicity.
I don't really have the time I used to have to try to answer things these days.