The Web is intended to be an Internet-scale distributed hypermedia system, which means considerably more than just geographical dispersion. The Internet is about interconnecting information networks across organizational boundaries. Suppliers of information services must be able to cope with the demands of anarchic scalability and the independent deployment of software components. -- intro to Fielding's thesis
Centralized registries are evil. (or at least: suboptimal).
non-dns-based stuff like UriSchemes/freenet
It seems like the web has some kinds of anarchic scalability addressed, but not others.
In particular, defining new MIME content-types does not scale. The web works for text/html, image/jpeg, and a few others, ... and people can introduce new types in limited populations, ... but deploying a wholey new image format requires change on the scale of the web.
In general, scalability of formats/languages is acheived through a hub-and-spoke model, where instead of needing N*N converters between N formats, you can just write N converters to a common "hub" format.
The appeal of XML, and to some extent RDF, comes from its being closer to the hub. It can look at first like XML (or RDF) is itself the hub language, but it's not. XML is an extensible class of languages, not a language itself. For all pratical purposes, this is also true of RDF. You can say that rdf:type has rdf:type rdfs:Property in RDF/RDFS without extending it, but you can't talk about books, people, or the weather until you move into an RDF+extensions language.
RDF is stuck right now in not having a standard way to indicate which extensions/ontologies are being used.
This is closely related to SelfDescribingWeb and PullTheDefinition.