Google Sets (http://labs.google.com/sets) was one of the earlier things available in Google Labs, back when it was just a dumping ground for weird stuff (as opposed to one for non-mainstream features). I'd guess this is where the technology used to generate the square from the samples comes from.

I also wouldn't call it training, for what it's worth. I punched in the same query, and was asked (like you were) to provide some samples - shouldn't (ideally) it know that it's not known this before, and use the samples you suggested?