agreed, and this always been the driving philosophy of pytorch, and perhaps why it kind of won so much brainshare against tensorflow despite _long_ odds when torch was ported from lua.
Soumith Chintala had a keynote talk in juliacon where he focused on these points;
https://www.youtube.com/watch?v=6V6jk_OdH-w
