Comment by svantana

That's mentioned in the article, but is the lock-in really that big? In some cases, it's as easy as changing the backend of your high-level ML library.