More or less, yeah. Most of the language-aware optimization in ghc happens in an intermediate form that still carries the types, and then ultimately it spits ou

More or less, yeah. Most of the language-aware optimization in ghc happens in an intermediate form that still carries the types, and then ultimately it spits out code that just has side effects as a result of evaluation like every other language. The code you sketch out above is basically the Writer Monad, which exists in the libraries, and does see use, but it's not IO -- as I mentioned above the semantics of IO don't entirely just fall out of the monadic structure. But I find the computed-imperative-program description lends itself to thinking about IO values as values in a way that thinking about them as type system goop to tame side effects doesn't; for example, the program: import Data.List (intersperse) import Control.Concurrent (threadDelay) main = sequence_ $ intersperse (threadDelay 1000000) (map print [1..]) will print out the natural numbers, one per second until you kill it. It works by constructing a (lazily evaluated) list of actions to perform, which alternately print out a number or wait for one second. sequence_ has type [IO a] -> IO (), and it just performs each action in the list in sequence (its type is actually a bit more general than this, but..). This kind of code is really weird to think about from the "tamed side-effects" model, but very natural from the "computing a script" model.