From there though, you have 2 more problems: keeping all data from being public (maybe becoming "friends" with another user involves a key exchange?) and distributing updates in a reasonable way. It'd be kinda strange if Lady Gaga had to buy a ginormous server to serve all the requests to her Twitter feed. I'm not sure how IPFS is implemented but if it's similar to BitTorrent in the sense that data can be seeded from any node, then that problem sort-of takes care of itself. You definitely can't think about "dynamically creating content" in the sense that twitter.com/:username may render different things at different times. The concept of web search also becomes an (even more) interesting one.
On a somewhat-related note, one also has to wonder how this system would interact with the European "right to be forgotten". Any legislature ensuring such a "right" would make this protocol a bit of a non-starter (at least in those jurisdictions).
Ultimately I think IPFS is a pretty cool concept, one I've kicked around a bit but never fleshed out. Major kudos to Juan and everyone else for taking it on - I look forward to seeing where this goes.