Yeah, that’s a good parallel. Copying is cheap and automation is applied not only to a task, but to the automation of the task.

this 2 fold demultiplication hits my mind regularly.

  instance / class / metaclass / ()
  instance / model / metamodel / ()   -- very redundant.
  value    / type  / kind      / ()
  machine  / vm    / utm       / ()

() denotes the closed loop where you don't need another layer.

anybody has a theory that explains this trait ?

It’s just a linguistic phenomenon. In English, at least, the prefix meta- denotes not only a single level of metaness (if you will) but also two levels or more, e.g., meta-metaclass → metaclass. So the “closed loop” is there only for convenience, like saying “and so on”.

If we were being precise, we would explicitly denote the level at which we were working. However, we tend not to do so for two reasons: first, that we rarely work with n-meta things for high n; and second, that being explicit is just plain unwieldy.

It’s for the same reason that we follow line, square, and cube with 4-cube, 5-cube, &c.

i think it's deeper than that. the progression is from object (every one hand-crafted) to template-based object-maker (hand-crafted per kind of object, but can produce a bunch of objects of that kind), to metatemplate-based template maker.

however, once you hit that second level of abstraction, you have almost invariably added enough flexibility that the range of templates your template-maker can make includes templates for other template makers. that is, it's not just a quirk of linguistics that a meta-meta-template-maker is called a meta-template-maker, they really do tend to be objects on the same level of abstraction and flexibility.

We seem to be saying the same thing; I just don’t see any inherent meaning in it. A literal interpretation of meta- would have us define it like this:

    Instance ::= 0
    Meta(X)  ::= Succ(X)

But it’s much more convenient to define meta- recursively:

    Instance ::= 0
    Meta(X)  ::= Succ(X)
               | Meta(Succ(X))

In both cases, we have an arbitrary designation:

    Class    ::= Succ(Instance)

That kind of redundancy is a linguistic artifact, not a deeply meaningful one.

It’s not the case that reaching the second level of abstraction necessarily results in enough flexibility to make the inductive step. As a counterexample, consider C macros, which only add one meta- because they are only expanded once. Lisp macros, on the other hand, are expanded until a fixed point is found. That’s why Lisp macros are Turing-complete while the C preprocessor is only a pushdown automaton.

A bit of disagreement with your last paragraph: The (Common) Lisp macro processor is Turing-complete trivially because it invokes a Turing-complete language along the way, not because of power inherent in its expansion strategy.

Common Lisp macro bodies are written in Common Lisp, i.e. written in a Turing-complete language — they would still be Turing-complete even if we wrote them in the subset Lisp-without-macros. The C preprocessor, on the other hand, does not invoke a Turing-complete language to compute a macro's expansion.

Of course.. zero, one, infinity strikes again.