As it happens, I'm working on something like this right now...
My current best solution is to use mouse/pen movement as cspline sample points, even when not drawing. The motion as you swing into a curve, even before the pen touches down, often predicts the trajectory of the pen stroke. (This also lets you "swing in" a circle without getting ugly linear segments).
This may be of limited utility on a touch-based device, however.
Your mention of Kanji is particularly salient, as the blog post catches these skeuomorphic icons in the act of evolving from pictograms to ideograms...
> I think it's disingenuous to just throw out "take that energy and try to build a company" as an alternative.
As someone who's just read your comment, I'd like to say: calm down, and re-read the parent comment with an open mind. It's dripping with life experience.
Interesting. The OP got me thinking along the lines of manually tagging salient features of each model (as well as ranking models by salience, either manually or automatically based on criteria related to the object that the model represents).
Please don't promote sectarian strife here.