Matlab is a very flexible and productive environment for doing exploratory programming. In my experience it even beats python (ipython + (emacs+auto-complete))[1]. The ability to execute small blocks of code in sequence using the %% notation for marking blocks was[2] a real joy to work with.
I thought C++ (.mex or .oct files) are where algorithms go once they've been proven useful for 'normal people'.
[1] http://bfrsblog.blogspot.com/2011/10/making-emacs-first-clas...
[2] How I miss the free Matlab license at the Uni. Last I checked a commercial license costs about $2000. So, now I've been forced to shift to Octave. Heck, if Octave is good enough for Andrew Ng, I guess its good enough for me.
The first statement is true: statistical learning is considered to be the leading edge of current AI research, especially since it attacks combinatorial explosion head-on. However, linear algebra is the baseline of current work, rarely the driving force. While the 'pull eigenvectors out of shitloads of data" method has been very effective and profitable, it is a very shallow approach to AI (especially considering the problems attacked in the 60s). Luckily, more intelligent algorithms are being developed that combine symbolic and probabilistic reasoning in a more direct way: compare LSA to LDA for example; LDA has a more cohesive and explanatory structure. These kind of things demand flexible programming environments (such as Lisp): Matlab is where algorithms go once they've been proven useful for 'normal people;' it is the domain of the analyst who cares about the data instead of the algorithm itself :).