I think we can have the best of both worlds. I would love Arc to have an intelligent compiler that figures out the best representation for your data, but I would also like access to the low-level data types.
Why? Because Lisp is supposed to be a reprogrammable programming language, where the user can do anything the compiler can do. If the user has access to the underlying bits and bytes, then they can implement new representations of high-level structures, rather than relying on the compiler writer. In fact, I would argue that Arc should only provide low-level structures in the axioms, and then high-level ones can be implemented on top.
