Backbone, at least as I practiced it, has a model that very closely corresponds to what goes over the wire. You observe stuff on that, you react to events, the event observers react and update the model and as a result the model is sent over ajax.
Amber, again as I practice it, has a model that corresponds to the complete application state of a particular component of the page. If you have a checkbox that doesn't 1-1 match with something you'd send to your server, you create an attribute on your model purely to track the state of that checkbox and wire up observers to update the appropriate "real" attributes you're sending to the server. This is not immediately apparent when looking at the two frameworks but I find the difference to be crucial.
I started using backbone the day you open sourced it since I had a personal version of more or less the same thing that was twice the code and less elegant. I found that it was fantastic as long as you could do idempotent DOM updates (i.e. innerHTML) but got lost as soon as I wanted to start adding sub components that needed to be initialized and I was back to dealing with a lot of event twiddling and dom manipulation for updating the view. I believe the split in state between the DOM and application model causes this complexity which is avoided in Amber by having all client side application state in one place. I think I could apply the same ideas in Backbone but I haven't tried since switching over.