Nice! Certainly results in a much cleaner question. Do you think you could briefly sum up how it's formatting the question if not by NLP? ie. how does it turn: 'john smith was born in abc, usa' into 'what is the birthplace of john smith?'

With Linked Data you always have triples of the form (subject, predicate, object). E.g. (dbr:John_Smith dbo:birthPlace dbr:Abc) and (dbr:John_Smith rdfs:label "John Smith"). dbr:Abc and dbo:birthPlace again are the subject of other triples, including a triple with (dbo:birthPlace rdfs:label "birth place") (available in multiple languages) and one with (dbr:Abc rdfs:label "Abc, Germany"). By retreiving the labels for the properties (predicates) and objects, I can simply form a question and the answer options. Currently every question is built naively in the form "What is the <propertyLabel> of <entityLabel>?".

Is it really infinite, without any repetition?

No, to be precise it's not infinite and not guaranteed to be free of repetition. However, DBPedia holds ~15 million entities corresponding to Wikipedia articles. Although only a fraction of those can be used for quiz questions (many don't have labels or their potential quiz answers don't have labels etc.), the chance to get a duplicate is still pretty low.