totally agree with you, reaching predictive performance is something that seems to be impossible...albeit Mathematica is a DSL and is there a need for predictability? Is Julia made for real-time systems? Do you have predictability using new/delete and/or exceptions in C++?
Did you expect the performance boost in what Tim wrote? The point for me is, that I expect from a scientific computing environment that I don't have to reuse old, dirty C-tricks etc. in order to gain speed. I think that is the strength of Wolfram and MathWorks, despite all their weaknesses they have and poor design decisions they've made along the line. Something that Tim wrote must come naturally in such a domain like Julia e.g. is in. If not, it is just wizardry and something you can gain as a power user of Mathematica/Matlab as well...somehow. If this is for wizards only, then it is nothing else but another programming language.
