But that's really not how it is - LLMs have their uses, obviously, but they really fall apart as soon as you're trying to do something even just slightly off th

But that's really not how it is - LLMs have their uses, obviously, but they really fall apart as soon as you're trying to do something even just slightly off the beaten path. As such, they're certainly still useful for answering questions for which StackOverflow has a dozen duplicates, and that may very well be a net positive for StackOverflow. Here's an example of something that came up in my work the other day - fiddling with prompts and checking the responses is ultimately more work than just writing the code from scratch. GPT4 and Sonnet 3.5 never came close to a correct response, even though the code always looks sort of "directionally correct". https://chatgpt.com/share/ba85ced9-63d6-4fb3-8dc1-b8efbaa58fa5 https://chatgpt.com/share/ba85ced9-63d6-4fb3-8dc1-b8efbaa58f... And that's despite this almost certainly being somewhere on Github in some form.