I think millstone is exaggerating a bit as most programs people write normally can be rewritten without too much pain in a static language with a rich type system. That said, there are some examples of times where your program asks for clever type system features (that are not always enabled by default or add to the learning curve) or where different type system features don't play nice together (so its hard to have both features in a single static language, but in a dynamic language you can use one style of programming at a time and never notice the problem).

In the first category, one example is programs requiring higher-rank polymorphism. They are always allowed in untyped languages but in Haskell this feature is hidden behind a language pragma to keep type inference simpler by default.

On the second category, an example I like is parametric polymorphism vs subtyping. The type system gets really tricky if you try to use both at the same time so most static languages either limit the polymorphism or don't allow subtyping. ON the other hand, in a dynamic language you can use both styles if you want, although you have to do so at your own risk.