I think we are talking about different use cases.

I'm coming from the perspective that you shouldn't be looking at the keyboard when you type- virtual or not. Instead you are relying on robotic muscle memory to generate a specific character.

So in my view there's a lot more new stuff to get used to in the author's proposed scheme. We would be memorizing a distinct set of finger motions for every distinct symbol as opposed to a fixed set of taps to certain positions on the keyboard - some with one or more preceding modifier keys (that are always at the same keyboard position).