The alternative is when you run a script that you last used a few years ago and now need it again for some reason (very common in research) and you might end up spending way too much time making it work with your now upgraded stack.

Sure you can were you should have pinned dependencies but that's a lot of overhead for a random script...