- 3 is nicer language to work in, transitions have been smooth
- I like different package managers for different purposes (pip for source or pure Python, conda for heavy precompiled binary packages, apt/yum for system packages)
- shallow copy makes memory management easier by default
- no built in stats structure: if there was, it would be the hammer looking for nails everywhere. Ever worked in MATLAB? Everything is a goddamn matrix (not true anymore but still).
A testament to the language and its extension APIs is that NumPy and Torch Tensors are very comfortable to work with yet aren’t built in data structures.