I doubt that a lot of people will use this, although the obvious place to deploy it is in mobile communications, where per-word prediction is still very slow and text entry is very inefficient. This seems 'good enough' for most functional communications. But I would imagine that it is or will be running in the background soon on all Google pages featuring text entry, like docs or mail, because that will provide a huge flow of data to refine their models against.
I don't have a really strong opinion, though I lean a bit towards the behaviorist approach. There have been some suggestions from ethnomusicologists that small children everywhere sing the same sort of melody on the same scale, which inclined me to the opposite view for a while, but I think that was more of a theory than a finding and still needs evidence.
I think the main issue here is that logic behind scribe is based on statistical frequency of co-occurrence of strings rather than any semantic logic. Granted, automating the semantics here could be tough and very expensive with the amounts of data available to Google's. However, there are some ways to do that if the index can be structured more efficiently and by using low-level semantic logic like taxonomies - I wrote a white paper about this if anyone is interested: http://www.exorbyte.com/index.php/White-Papers/ajax-incremen...