Interesting! Any reason why bindings for torch or tensorflow weren't created instead?

It essentially is a binding for tensorflow if I understand it right.

Would this article read any differently if it was just titled "Kotlin DSL for Deep Learning"? I'm not sure where the "idiomatic" comes in to play really.

Nice!

How's Kotlin in general for machine learning/deep learning?

Almost as good if not on par with the JVM ;)

That depends how much it will expose from Valhalla, Vector 2.0 and cudaGraal.

Then there is the engineering effort to provide the same elsewhere.