I get and have built Probability monads often enough. They seem really cool since keeping normalization around is theoretically nice, but they seem broken in "real" applications.

Everything I do with probability ends up looking mathematically dissimilar to probability. Most algorithms go far out of their way to avoid on-the-fly normalization because it's so expensive. An EM update monad is just the State monad.

I love Haskell (never used Scala) but I find the very things you're marketing it with abhorrent to use.