Raw LLM Responses

It's definitely overhyped. I've tried Cursor, Copilot with it's various models, ChatGPT (various). The one thing any of them do reliably is generate boilerplate. And to be fair they're very good at that in most cases. However if you exclude very standardized MVC API frameworks it's essentially incapable of writing anything production worthy. Small one off scripts that you would forget the design of in a week anyway? It's good at. I used to be a pretty big user kf software called Kite which was copilot before copilot. I've been using things like this for a while and honestly the boilerplate is most of what it's good at. That's why it can add on to an existing MVC simple REST API framework so easily developing DTOs, Services, Controllers, etc is largely boilerplate. If you're proficient at something like Vim/Emacs/VSCodeExts/etc then you can edit at a pretty rapid pace. If you haven't gotten there yet that's okay, but it's definitely worth a shot imo. If you've got that the main thing it's useful at is: if you've forgotten various SQL features, forgotten the use of your libraries, aren't using a DSL/metaprogramming system, and sometimes finding the most idiomatic way to write a small code snippet to solve a well understood problem. Most of which is due to lack of familiarity with an existing codebase. Despite people constantly saying this or that new model is so much better I haven't noticed much since Chat-GPT4. Some do get things right where others get things wrong, but not in a consistent way. The main thing that I think drives Cursor hype is just that it's much better at dealing with context. Using context with copilot is a pita. For generating CI/CD workflows, throwaway shell/python scripts, IaC with well known providers it's fairly alright. But it essentially just provides a rapid increase in typing and googling speed. In niche areas like embedded programming it's alright, but you often have to give it a lot of context. For ML code it's alright, but they will