This got me thinking about the future of ML. Right now ML solutions are like large blocks (centerpiece), which we feed with data and extract "value" from.<p>Perhaps rather than making ever more powerful ML "monoliths," maybe the future is with "ML modules" that can be glued together meaningfully (i.e. beyond data proc. pipelines). It would make sense that probabilistic programming languages will be useful for gluing such modules together.<p>Link to church in js book from the article: <a href="http://dippl.org/chapters/02-webppl.html#and-inference" rel="nofollow">http://dippl.org/chapters/02-webppl.html#and-inference</a>