[abstractbill]

The problem with using a learning model in an interface is making it smart enough to do its job, but dumb enough that the user can easily have a mental job of how it will act.

Yes, exactly. At the same time as the switcher program is trying to learn the user's common patterns, the user is also trying to learn how the switcher works. I can see that not ending very well!

It would be particularly frustrating during the learning phase, when the switcher doesn't know the user's habits and so switches poorly, and at the same time the user is struggling to figure out what order the switcher will put things in.