The section of the paper where this is mentioned offers a novel (at least to me) argument in support of homoiconicity, to wit, that you build DSLs directly into the base language, which (it is asserted) makes the DSL development process more flexible (or something), and that this is important for 'living' domains, where you're trying to work out the domain semantics, like biology. So, you're not really so much building a specialized language for X, as slowly turning (almost in the sense of wood turning, as on a lathe) Lisp into X-Lang.
