In considering where Objective-C could go, it's worthwhile to start by understanding the work that's already been done.
Below are a selection of papers from Gilad Bracha's 1990s work on Strongtalk, an extension of Smalltalk (from which Objective-C's design derives) with (among other things), stronger type-safety tooling.
What's interesting (to me, anyway), is that the work done on the Self/Strongtalk VM in the early 90s was actually bought by Sun and became the modern Java VM. When Google got started on their V8 JavaScript runtime, guess who shows up again — Lars Bak, who was the technical lead for both the Strongtalk and HotSpot Java VMs.
If we're going to be talking about how to apply "modern" (1990s!) ideas to Objective-C, we'd be wise to review the considerable work done in considering those sorts of problems in a Smalltalk-derived universe, and lifting whatever good ideas we can, and discarding whatever ideas have been superceded in the intervening decades.
- Modules as Objects in Newspeak (2010)
- Mixins in Strongtalk (2002 - ECOOP02 Workshop on Inheritance)
- The Strongtalk Type System for Smalltalk (1996 - OOPSLA96 Workshop on Extending the Smalltalk Language.)
- Extending Smalltalk with Mixins (1996 - OOPSLA96 Workshop on Extending the Smalltalk Language.)
- Strongtalk: Typechecking Smalltalk in a Production Environment (1993 - ACM Conf. on Object-Oriented Programming, Systems, Languages and Applications)
- The Programming Language Jigsaw: Mixins, Modularity, and Multiple Inheritance (1992 - Gilad Bracha PhD thesis, University of Utah, 1992)
A lot has happened since Strongtalk, not least of all Objective-C, which showed that you don't actually need a VM, byte-coded instructions and corresponding JIT to get dynamic OO and good/excellent performance.
In fact, the Combined Lambda Architecture (COLA) work by Ian Piumarta (VPRI) removes more and more stuff until he's got something that's very close to Objective-C, just cooler. Even messaging is implemented by messaging the class object, with the infinite recursion cut off by pre-populating a few select method caches. See for example http://piumarta.com/papers/DLS-2006-slides.pdf
I think we also need to take into account the pain points that have been exposed in practice. Strongtalk was never used in anger.