If you can't make a language fast, make it parallel

This is not true at present. The core data structures, and the STM machinery are implemented in java, not as macros.

Isn't that a bit of a loaded question? In some academic sense, an ideal programming system might have a kernel language that is probably rather small and certainly clean, flexible and extensible in its feature set, and then almost everything else built on top of that kernel in libraries. In practical programming systems, the perfect kernel has proven rather elusive, though.

It's rather like the question of whether a standard library should be small, clean and extensible or comprehensive with a good-enough version of everything. Many of us might be inclined toward the former approach from an academic/theoretical/intuitive point of view, but in practice, just about every widely used programming language from the past two decades has come with a batteries-included library. While those libraries are often criticised for various technical weaknesses, and many of those weaknesses are never addressed because it's too much work, these systems are still good enough for most users out of the box.

Erlang has that, because it has its datatypes and a defined serialization for them and there is nothing in the language that is not those datatypes. It also doesn't permit you to layer any type-level assertions about those types into a user type. In fact, Erlang basically has no concept of user types. (Records are syntax sugar around the built-in tuple.) Since the Erlang data types are so weak, an automated serialization can be built that requires no work to use. But that doesn't come without cost

If you don't start with that, you have to use some sort of introspection to examine data types, and you probably don't have a good story for what to do if two ends have totally different ideas about those datatypes, or how to upgrade running processes where you literally want to change the datatype without shutting the process down. The stronger your type system, the harder that gets. The easier it is, the weaker your type system must be. Erlang's capabilities don't come for free, they exist because they wrote their (non-)type system so that their data structures never have any guarantees in them that don't trivially travel across the network. This has its own issues; Erlang has just about the weakest types you can have without actually having arbitrary casts, and that has consequences too.

There are a handful of characteristics in a language, like its type system, that have radical impacts throughout the language and no amount of library jiggery-pokery can completely paper over. Another example not entirely relevant to cross-process concurrency (but not entirely irrelevant) is immutability; no amount of library work can turn a mutable language into an immutable language.

On p.173 Joe writes:

  When we say things such as this:
    Pid = register(...),
    on_exit(Pid, fun(X) -> ..),
  there is a possibility the process dies in the gap between these two statements.

So in Spin you wouldn't be tempted to write something like above (probably not Erlang either?).