Duolingo would probably benefit if it used a more incremental approach. Translating whole encyclopedia passages at its outset is daunting and demotivating work.
The principle is interesting, but the error-checking doesn't appear to work without something of the following in place: One could borrow a parallel corpus of professionally translated sentences (for Chinese, maybe this one:
http://corpus.nie.edu.sg/babel/index.htm). Following that, one orders the sentences by "difficulty," which from a more mathematical perspective reduces to the following variables: term/morpheme frequencies, binary/ternary/n-ary collocation frequencies, and sentence length. With colloquial outliers, most basic sentences can translate readily to another language without much grammatical concern at all. One could then incrementally increase the number of terms and the lengths of the sentences until, at last, one has learned a few thousand terms and can state sentences of invariable length.
The other line would require a parser which simplifies sentences into all of their constituent forms, but that would require a level of intervention that users probably couldn't engage, and it would be costly unless someone could devise a strategy through which one could challenge native speakers to eek out the atomic sentences from their complex ones.
My guess, as was hinted earlier, is that the TED talk cherry-picked the example. However, if you manage to borrow enough sentences from parallel corpora, determine a way to organize them by difficulty, and then feed that loop back into the data pool which estimates the accuracy of the translation, the rest is borne of itself.