Proposed new feature or change:

On hacker news the following blog posts about NumPy have been trending a while ago:

• https://dynomight.net/numpy/ (disadvantages of broadcasting)
• https://dynomight.net/dumpy/ (proposed alternative)

It argues that while broadcasted notation has advantages over loops, it is not always obvious how it works on highly dimension arrays. It proposes an alternative api based on classic for-loops, which are still vectorised in the background. I was wondering if there is interest to incorporate something similar directly in numpy.

Alternatives:

• Pytorch has already added named tensors, although this features is still experimental. Since pytorch was inspired by numpy, this idea can be incorporated back, especially if this feature proves to be valuable.
• xarray builds on numpy and provides a similar datatype which has named dimensions.
• einx fully embraces the notation used by einsum. Since this function is already part of numpy, there might already be a precedent to add other similar functions to numpy.