When I looked at implementing some of that on Symbian OS in trying to develop an intelligent mobile context sensitive recommender system, I discovered that the problem was _not_ the ‘AI’ (or whatever you want to call the magic) but the fact that you couldn’t source the context and user actions easily from the frameworks and apps in order to feed them into the ‘machine learning black box’ ; at least not without full access and major enhancements inside the system’s code. This seems to be true for most if not all mainstream environments desktop or otherwise (with notable exceptions Apple’s Open Scripting Architecture).

For mobile’s to fulfil the “personal digital assistants” vision and infact Alan Kay’s Dynabook vision, in my opinion mobile platforms must be tuned/redesigned to allow proactive agents to take action. To achieve that they must be engineered to reflect what is going on in the system (apps, UI etc). This is the only way for agents to have valid info on what is going on and therefore to drive machine learning, until then application will be “narrow”. Nevertheless such “narrow” applicability can be very useful perhaps (like Lieberman’s Letizia, Bradley Rhodes’ Remembrance Agent, Balabanovic’ Fab/Slider or the Daily Learner by billsus and Pazzani).

FYI:
http://web.media.mit.edu/~lieber/Publications/Publications.html
http://www.bradleyrhodes.com/Papers/physical-context-ieee-toc.pdf
http://www.ics.uci.edu/~pazzani/Publications/billsuspazzanichen.pdf
http://www.balabanovic.pwp.blueyonder.co.uk/marko/Marko-Balabanovic-Research.html

I’ve often said (or complained): “There are two types of application: those that do what they are told and those that do what they think I wanted them to do; I know which I prefer”. Part of this is that applications aren’t yet very good at anticipation, but part of it is genuine personal preference.

Consider human personal assistants. Some people prefer assistants who effectively and efficiently do what they are told, others prefer assistants who try and anticipate and act upon their needs. It really is a matter of personal preference.

There are two complementary problems for digital assistants. One is that of intrusiveness (the Microsoft paperclip announcing: “You seem to be writing a letter, do you want to…”). The other is silently doing the wrong thing (my word process or silently correcting a spelling error when in fact the word it was correcting was part of some Swedish text included in my document – rather than correcting an error it introduced one).

The examples you give avoid these problems by providing a selection from which the user can easily reject the wrong items. It is neither intrusive, nor does it introduce errors.

There are two challenges for digital assistants – anticipating what the user wants and providing that information in a non-intrusive and non-error-prone way. I don’t think we’ve seen much advance in either of these problems, other than in the narrow area of assisted search. So no, I don’t think we’ll see mobiles ‘fulfilling, at last, the vision of “personal digital assistants”’ within the next decade.