So what separates the two is the error metric, in other words, how much humans would bother if you replaced y for x. But this involves humans, not general intelligence -- you'd need a lot of testing with actual humans to get good results (jpeg has extensive testing to find optimal quantizers for chroma values based on human perception, for example).
Not that the latter would not be interesting too; I just think it's less practical.
Relevant: http://www.stanford.edu/class/ee368b/Handouts/04-RateDistort...
Imagine, for instance, how many articles there are about how A has relation R with C, and B also (e.g., George Washington was a president of the USA, Abraham Lincoln also). For some definition of "understanding", you wouldn't need to encode data about the text of both articles A and B, you could use NLP to somehow store just the ontology. The problem would change from encoding text to encoding a graph. So it would be a lossy process (because you wouldn't be able to decompress into the original Wikipedia), but in my opinion seems more relevant to AI than bothering with the original corpus.
Exactly, I see the same problem. With the current rules, this is just a challenge of text compression.
I believe compressing the knowledge in the Wikipedia would be more akin to extracting meaning with NLP into a highly-normalized dataset, and then compressing that. From an AI perspective, this is more close to how our memory actually works. Unless you're savant, you don't memorize the actual text of a topic, but the inferred associations, which compress much better, although lossy.