I like what Facebook had done on tensor comprehension. Basically generalizing from pure mul+add to other operations and allowing specifying what to reduce, as

I like what Facebook had done on tensor comprehension. Basically generalizing from pure mul+add to other operations and allowing specifying what to reduce, as needed for e.g. expressing convolutional kernels. Sadly abandoned; I still love the conciseness.