I find working with LLMs much more fun and frictionless comprated to the drudgery of boring glue code or tracking down nongeneralizable version-specific workarounds in github issues etc. Coding LLMs let you focus on the domain of you actual problem instead of the low level stumbling blocks that just create annoyance without real learning.

Focusing on your problem means focusing on minimizing that stuff.. with LLMs it will fester.

Not in my experience so far. But if it becomes a problem, I'll reconsider.