"Very useful" often just means: coding a properly typed, easily-understandable solution takes longer than doing it untyped.
I argue that is the case only when using a plain-text editor. When you have a good IDE for a typed-language at hand that can refactor, complete, analyed and follow code on the press of a button, you lose this "advantage" of untyped languages.
Now, what IDE and language are you referring too? The best one I've used was VisualStudio for C#. While I'd say typing wasn't adding too much overhead in it, its type system is also poor, and it's arguable that it may not really prevent much bugs. Now, it does help make C# faster.
It’s just that historically people have found it markedly more useful to have that information available statically, because it lets you prove things about your code that hold “for all” cases (universals), not just “for some” that are executed (existentials). And static information can be used for optimisations because of those guarantees—if you have proof, you don’t need a fallback.
If you give me an unknown function f : ∀T. T → T in a statically typed language, then I know it always returns a value of the same type as its argument. With some assumptions (purity, parametricity, halting) I even know it must be the identity function.
Whereas doing what I’m suggesting in a dynamically typed language, running inference dynamically, I could only recover existential properties: that there are some types for which f returns a value of the same type as its input. So I could recover more and more information (in the form of an intersection type) by calling the function with different inputs:
f : int → int
f : (int → int) ∧ (string → string)
f : (int → int) ∧ (string → string) ∧ (Foo → Foo)
But I could never arrive at ∀T. T → T, because that’s essentially an infinite intersection.
ETA: actually I might not even be able to get this much information; I only know the output type for each input value, not each input type.
Now, you may say “okay, the original program doesn't exist anymore, but my language implementation is a term rewriting engine, so I have a current program that I can submit to the type checker”. Alas, most type checkers, Hindley-Milner included, operate on self-contained program fragments (i.e., not containing unbound variables), so you can't just submit any arbitrary program fragment. And, if you didn't want to type-check your original program, how likely is it that you will want to type-check the entire syntactic representation of the current program state?
Say, because you want to use that reference cell in two different threads.
And degrading the type to a union (int ∨ string) as I mentioned is still nondeterministic even though unions form a lattice—I can recover determinism when writing to the list in both A and B, but supposing A writes first, B writes second, and then A reads expecting only integers, it gets a type error because it hasn’t checked for the string case.
But I’d argue that 1. this is bad form and you want a type error somewhere and 2. you have this problem already with type-changing mutable references in a dynamically typed multithreaded language. (Say just “int” and “string” instead of “list of int” and “list of string”.)
This happens quite often. Take datastructures. It is often very useful to structure different types of data together in a common structure. Now in most cases, you don't know in advance which type will go where in the datastructure, as the data might only be known or derived at runtime.
In fact I'd say dynamic behavior is the major expressiveness loss. Granted, I'm not sure how runtime vs static type system relate to type theory. Anyone can educate?