H.264 and JPEG are supposed to output random-looking bytes, by definitions. If you can compress those, something is very wrong.
Well, it seems to be applied per codec not a general compression algorithm like zip. And they probably say mobile-encoded for a reason, simple encoders have to work on low power and in real time, random JPGs from the Internet is probably the same. From what I can gather the algorithm basically take a global scan of the whole media and applies an optimized variable-length transformation making commonly used values shorter at the expense of making less commonly used values longer. Nothing you couldn't do with a proper two-pass encoding in the codec itself, the neat trick is doing it to someone else's already compressed media afterwards in a bit-reversible way. Very nice when you're a third party host, assuming the increase in CPU time is worth it but not so useful for everyone else.