When you work with LLMs, you can also take a different approach towards dependencies. Instead of seeing them as fixed, you can actually just completely integrate them and take ownership. https://chris.eidhof.nl/post/integrating-dependencies-into-llm-assistant-projects/
Integrating Dependencies into LLM-Assisted Projects — Chris Eidhof

@chris
I've been asking for a while:

If LLMs are so good at programming, why do they need to use libraries/frameworks at all?

@mikro2nd @chris That’s my take as well.
Instead of actually adding (vendoring) a dependency, you can just tell it to “write a target providing the same functionality as FOSS project X”, and it will happily inject the already stolen code for you.
@helge @mikro2nd @chris I mourn for what has become of us. I wish I’d stayed in nursing school all those years ago.
@causticmsngo @helge @chris
My intended implication is that LLMs are actually NOT any good for programming at all. To think that they are is a category error that arises only because the mathematical notations we use for specifying algorithms (i.e. "programming languages") superficially resemble human languages, and that's the only reason a Large LANGUAGE Model can even begin to approach "using" them.
@chris I think this is probably the end of distributed collaboration

@mattiem @chris I'm being optimistic here (rare for me), but with the help of LLMs, maybe there’s a way to maintain collaboration but move it to a higher level?

Instead of sharing a literal code base, maybe we find a way to share the algorithm / approach / design. Maybe we find a way to push the ways we changed and mutated the code upstream in a more descriptive way than just "here's a code patch”?

Maybe those who want to can mandate that their agents actively participate in this?