Measuring the impact of AI on experienced open-source developer productivity

I actually think that pasting questions into chatGPT etc. and then getting general answers to put into your code is the way.

“One shotting” apps, or even cursor and so forth seem like a waste of time. It feels like if you prompt it just right it might help but then it never really does.

I've done okay with copilot as a very smart autocomplete on: a) very typical codebase, with b) lots of boilerplate, where c) I'm not terribly familiar with the languages and frameworks, which are d) very, very popular but e) I don't really like, so I'm not particularly motivated to become familiar with them. I'm not a frontend developer, I don't like it, but I'm in a position now where I need to do frontend things with a verbose Typescript/React application which is not interesting from a technical point of view (good product, it's just not good because it has an interesting or demanding front end). Copilot (I use Emacs, so cursor is a non-starter, but copilot-mode works very well for Typescript) has been pretty invaluable to just sort of slogging through stuff.

For everything else, I think you're right, and actually the dialog-oriented method is way better. If I learn an approach and apply some general example from ChatGPT, but I do the typing and implementation myself so I need to understand what I'm doing, I'm actually leveling up and I know what I'm finished with. If I weren't "experienced", I'd worry about what it was doing to my critical thinking skills, but I know enough about learning on my own at this point to know I'm doing something.

I'm not interested in vibe coding at all--it seems like a one-way process to automate what was already not the hard part of software engineering; generating tutorial-level initial implementations. Just more scaffolding that eventually needs to be cleared away.