I’m adamantly _not_ a fan of LLMs trained on scraped publicly-available data, and that goes double when used for coding. And so many of the examples people put forward for using it to code are along the lines of “I asked it for an app to do X, and was up and running super-quick!”
That seems like a templating issue to me. My current day job is mostly #Rails, and its generators (built-in and extensible) allow for automating a lot of the boilerplate building models, migrations, etc. Sure you still need to do some customization, but even with the “LLM success stories” there’s always a bit where they say “I had to tweak a few things, but it worked!”
So would it be a crazy idea to have a collection of templates and or code snippets to do some task in some framework/environment? Tag it for easy search, etc. i.e., if I needed OAuth in my Rails app I could search “oauth rails” - maybe “oauth rails_8” if I know I want the latest major framework version. Integrate the result with my app, and move on.
I can see plenty of issues. For one thing it’d be basically stackoverflow, but with gists that are a bit more organized. So someone would still have to understand what was going on. But if they’re copying answers from SO or anywhere into their code without first looking at the _why_ of how it supposedly fixes the issue, that’s the first problem to address.
More generally, it’d be great if framework and other teams would put out these kind of things on their own, so someone finding them could be reasonably sure best practices had been observed. But having done dev and doc writing for decades, I know how overwhelmed everyone gets.
#SoftwareDev
#FuckAI