I think what you’re looking for is that languages are more language when they are interpreted by something that looks more like a VM, and interpreters are more VMy when they have constructs that look like “if” and “load from memory” (letting you choose how to act on observing current data, based on past experience).

If it’s easier to study the language as a what you term “direct,” it seems less languagey, or whatever.

I personally think this distinction is distracting because you can always create a (large) bijection between encodings and their output when interpreted and then it looks “direct”, without actually writing the VM (just a huge lookup table).

I feel like smart languages just look nicer to use because they compress better (length to output blowup is exponential), and we think that seems more like human communication.