If we only used the Array API internally to work on our .data, we would get support for other kinds of arrays than Numpy. This would be a huge task, but would bring high rewards (transparent dask or GPU support).

This would make #530 obsolete I suppose.

If we provided/exposed an Array API, which is a different task (also huge), we would also get high rewards: our Array objects would be usable as-is in some downstream libraries such as scikit-learn.

A much smaller/easier task would be to implement only the Data Interchange protocol (aka DLPack), which would make la.Array objects consumable directly by downstream libraries like scikit-learn (but would loose labels in the process).

I think this would only be a matter of implementing __dlpack__ and __dlpack_device__ methods both forwarding their calls to the respective methods on .data