I generally agree, but I feel like there is less skepticism than you portray among the ML/Stats community about probabilistic modeling languages (hence the DARPA call for white papers). Universal inference for well-posed models is ambitious, but a lot of people have worked hard, with significant success, to make it happen.
The Geman and Geman paper from 1984 was directed at general procedures for inference using sampling, as was the work of the Brown group (Grenander et al.) throughout the 80s and 90s. BUGS (http://www.mrc-bsu.cam.ac.uk/software/bugs/) was specifically directed at that goal, and has been highly successful in real applications since the mid-1990s. Other more recent software, like Stan (http://mc-stan.org) is also targeted at this approach.
Looking at this from a different angle, the published work of many ML researchers has been directed at unifying common models with the clear view of producing software. I'm thinking about Michael Jordan's "plate" notation (http://en.wikipedia.org/wiki/Plate_notation) and the various unifying reviews of multi-component models and inference algorithms for time series (e.g., http://dl.acm.org/citation.cfm?id=309396), and the corresponding reviews for the EM algorithm done by Meng, van Dyk, and others (e.g., http://www.stat.harvard.edu/Faculty_Content/meng/JCGS01.pdf).
OK, fair points. Perhaps think that came across a bit more negative than I intended it to.
I think there's a lot of value (and a lot of buy-in) for probabilistic modelling languages, which impose some constraints in order to get efficient inference, which help you meet those constraints and don't provide a false promise of generality. And tonnes of value in research which looks for nice unifying formalisms to enable this sort of thing. Also in nice formalisms for inference in general, so perhaps that will be a useful side-effect of this work.
It's the idea of full-blown probabilistic programming, where you have unconstrained turing-complete non-deterministic programming language and inference just works, where I've seen a bit more in the way of healthy skepticism. Of the "this will be nice if it ever arrives, but in the meantime I'll be getting on with doing statistics, which it doesn't obviate the need for".
Similar to a computer scientist and the ultimate "sufficiently smart" compiler I suppose.
The Geman and Geman paper from 1984 was directed at general procedures for inference using sampling, as was the work of the Brown group (Grenander et al.) throughout the 80s and 90s. BUGS (http://www.mrc-bsu.cam.ac.uk/software/bugs/) was specifically directed at that goal, and has been highly successful in real applications since the mid-1990s. Other more recent software, like Stan (http://mc-stan.org) is also targeted at this approach.
Looking at this from a different angle, the published work of many ML researchers has been directed at unifying common models with the clear view of producing software. I'm thinking about Michael Jordan's "plate" notation (http://en.wikipedia.org/wiki/Plate_notation) and the various unifying reviews of multi-component models and inference algorithms for time series (e.g., http://dl.acm.org/citation.cfm?id=309396), and the corresponding reviews for the EM algorithm done by Meng, van Dyk, and others (e.g., http://www.stat.harvard.edu/Faculty_Content/meng/JCGS01.pdf).