[solidsnack9000]

Setting aside for a moment the appropriateness of Swift for TensorFlow, this is a very impressive example of using an embedded DSL to work with a component that is a full programming system in its own right.

On the one hand, we do want full access to the programming model exposed by the component -- its control structures, abstractions, everything else. One the other hand, these are mostly duplicated by our host programming language: it's going to have variable bindings, operators, iteration, conditionals and everything else. Doing an embedding like this is a way to expose most of the component's facilities without introducing a ton of "new syntax" in the form of combinators or having programs where a lot of the critical code is escaped in strings.

This same problem shows up in programming interfaces to RDBMSes. LINQ is a good example of the same embedding technique.