I guess LLMs can learn directly from documentation websites. But there is no human feedback on correct solutions or better ways to do things that SO provides.

Another nice thing is that SO is, ironically, often relatively hostile to asking questions. It seems like every post there's some group of people racing, often recklessly, to try to mark things as duplicates when they're not, asking for code samples even when the question clearly does not require it, etc. And then there's the passive aggressive stuff. It seems like 95% of the SO community is awesome, but that 5% sure is a turn-off from the site. GPT cuts through that layer of nonsense.

For me, ChatGPT has unlocked an entire class of questions and investigations that would be almost impossible to handle with just search.

Most people I know have done exactly that.

I did to a large extent. If verification is easy(which in most of cases it is), I first ask the question to GPT 4 before searching SO. And tbh I have found SO top answer to be more probability of being wrong than GPT 4. In most of the code, if the code doesn't look wrong and could pass the testcases, which I use GPT 4 to generate and verify manually, it is likely fine.

Also there is a huge difference between GPT-3.5 and GPT-4 so if you have only tried GPT-3.5, you experience will differ wildly.

Man, maybe it's just me, but I have switched almost exclusively from Google searches and Stack Overflow to just asking ChatGPT. I've been doing dev for 20+ years and GPT4 is like switching from regular tools to power tools. Granted, it's only useful for things that have been done before and are well documented, but most professional software development is exactly that.

For the simpler questions it is quite ok. As in write me x in language y where those are inputs

SO is left only for really hard or niche one.