I suggested two approaches for representing Extensions, one explicit, one implicit. I decided on the implicit approach. One is because I haven't yet figured out how to make Elements typed by specification; the other is that FHIR has an odd mismatch between Extension allowed atomic values and Primitive allowed atomic values.
Extensions in FHIR provide a space to allow additional values and semantics against defined FHIR resources, without requiring implementors to hack into the design space of the FHIR resources themselves. They play the same role that x_ header values or, perhaps better, IANA registration of MIME and other types.
We keep changing our definition of Elements based on learning more about the structure of FHIR. As we implement id and extension, we'll have to change our definition yet again.