(1) Monadic programming gets really hard to understand if you don't have a strong type system correcting you when you screw up the difference between `x`, `m x`, and `m (m x)`. The tendency in weak type systems is to autopromote `x` to `m x` with `return`, but this can make it impossible to see if you need to deconstruct `m (m x)` to `m x` with `join` in any given application, unless you just auto-flatten everything, which sometimes (especially in the list monad!) is not what you want.

(2) Monadic programming gets hard to understand if you can't cleanly split out functions into having just one argument and one output.

Translation for people who don't know what a "monad" is: occasionally there are type-adjectives (I'll use the example "blue") where there is a canonical way to make any object into a blue object (called `return`), and a canonical way to make a blue blue object into a blue object (called `join`), and a canonical way to take some function from objects to objects and turn it into a function from blue objects to blue objects (called `fmap`).

One example is the adjective "either an X or a ____", which is written in Haskell as `Either x`. The canonical way to take a Y and `return` it into "either an X or a Y" is to say "it's a Y -- the type on the right." In Haskell we'd say that `data Either x y = Left x | Right y` and that this `return` function is `return y = Right y`.

Similarly the canonical way to map a function `y -> z` over such an `Either x y` to get an `Either x z` is "if the value is an X don't do anything, otherwise do the function to the value and collect the end result." This is written `fmap zy exy = case exy of Left x -> Left x; Right y -> Right (zy y)`. Finally, the canonical way to take an "either an X or an (either an X or a Y)" into "either an X or a Y" is just: if you've got an X, it's an X; if you've got a Y, it's a Y:

    join ex_exy = case ex_exy of 
      Left x -> Left x
      Right exy -> case exy of 
        Left x -> Left x
        Right y -> Right y

So it's adjectives which are (a) universal [we can apply them to anything], (b) collapsible [we can take an `m (m x)` and turn it into an `m x` and (c) mappable [we can take an `x -> y` and apply it to an `m x` to make an `m y`].

Haskell became obsessed with them because this structure describes computer programs really well. Suppose you build in a data type for "a program which produces a...". Then the basic "program composition" operator looks like this:

First, take two inputs: (1) a program which produces an X, (2) a function which takes X and produces a program which produces a Y. Then, yield a program which will (when run) produce a Y by first running program (1), then using the X generated as an input to the function (2) to compute a program which produces a Y (3), then runs (3) to produce the y.

The type signature here is `(IO x, x -> IO y) -> IO y`. This curious function is called `bind` and it can be more easily thought of as a combination of two separate steps: `fmap` the `x -> IO y` on the `IO x` to produce an `IO (IO y)`, then `join` the programs to produce an `IO y`. "A program which produces a ..." is thus a special case of a monad, an adjective which supports these three operations (`return` is "produce a program which does nothing and yields this Haskell value").

And you care about all of this because being all indirect -like about computer programs allows you to metaprogram up an impure program in a pure language, which is how Haskell does I/O.

So the translation of the above points is that (1) this whole process gets confusing if you don't know the difference between a wagon, a blue wagon, and a blue blue wagon at compile time, and the tendency in a Lisp is to say "I saw a wagon, I wanted a blue something, so let's just quietly use the `(paint-blue wagon)` subroutine to keep going," and "I saw a blue blue wagon, I wanted a blue something, so I am happy." You can force things to always use some `(coalesce-blue-paint ___)` function to always flatten blue-blue Xs into blue Xs, but this is usually counterintuitive. (2) You can't even clearly articulate what `fmap` does unless you can cleanly say "I have a function from wagons to wheels, and I want the corresponding function from blue wagons to blue wheels." If your functions are taking a bunch of extra parameters, like a wheel-index of "front-left, front-right, back-left, back-right" then it gets hard to say "here's the wagon input that I want to become a blue wagon input, please leave my wheel indexes alone!"