Certainly there's something very intellectually appealing about a completely transparent stack of abstractions in which all the upper layers can be ultimately be expressed in terms of atomic operations.
However, I'm not so sure this is really so important in practice. We tend to operate within bounded layers of abstraction for any given problem and it's far more important that those layers are predictable, comprehensible, and, ideally, well-documented.
This is why a language like Java, for instance, can be very productive even though the fundamental abstractions are not plastic. And this is also why more malleable languages like Scheme or Lisp can be less productive because there's not enough consensus on how the higher layers should be defined.
I agree, and I think we see this same thing bore out time and again in our field. This is also the distinction between monolithic and microkernels, part of the reason Haskell is succeeding while ML atrophied, the success of statistical AI over "classical" AI, and so forth. We are attracted to minimalism, but it's too much of a hair shirt in practice.
However, I'm not so sure this is really so important in practice. We tend to operate within bounded layers of abstraction for any given problem and it's far more important that those layers are predictable, comprehensible, and, ideally, well-documented.
This is why a language like Java, for instance, can be very productive even though the fundamental abstractions are not plastic. And this is also why more malleable languages like Scheme or Lisp can be less productive because there's not enough consensus on how the higher layers should be defined.