I'd say that Knockout can be suitable for large applications, but that you need to handle having an unusual computational model pervade your code. If you're interacting with existing code, it can be tricky to get inputs and outputs mapped correctly, but the knockback library mentioned in the article probably solves that for backbone users.
One rule of thumb that might be useful is that any (non-event-handler) function you give to Knockout must handle being called an arbitrary number of times at arbitrary times, with some inputs being potentially outdated. If it cannot, you're not using the computational model the library essentially requires.
Yes, I have a lot of custom bindings, but most are very reusable, and many have saved my butt (like wrapping bindings to include JQM's markup enhancements).
That said, I wasn't new to MVVM when I first came across Knockoutjs, so I used it efficiently from the get-go.
If you don't like MVVM, you won't like Knockoutjs.
I was casting around for a simple way to implement a responsive UI in our latest application and Knockout appeared to be a good option (I'd not come across Backbone at that point).
Now it just feels like more work... I have written so many custom bindings it is unreal!
(I realise this is, as the author notes, mostly my failing in writing JS :))
I actually find writing custom bindings pretty enjoyable. That's where the nice separation and reusability come from.