I would love to see some benchmarks that demonstrate the impact of the more optimized representation. I found that despite some possibly sub-optimal performance trade-offs in favor of productivity in my own hobby interpreter, it can still do much better than Python and Wren in some micro-benchmarks that I performed (naive sum in a loop and naive fibonacci implementation). Python was no surprise since PyObjects are relatively "fat" [0], but Wren was created by the author of Crafting Interpreters (mentioned in the article) so I expected a bit better. My representation ended up looking something like:

  enum Value {
    Unit,
    False,
    True,
    Integer(isize),
    Float(FloatSize),
    String(Rc<str>),
    Closure(Rc<Closure>),
    Tuple(Rc<[Value]>),
  }

Unfortunately since I'm depending on reference-counting, it can't implement `Copy` which I suspect is where I'm currently losing significant performance. For values embedded in bytecode instructions I can cheat and copy anyway since only the non-refcounted ones are allowed in them.

[0] https://github.com/python/cpython/blob/main/Include/object.h...