Weak typing

A language with weak typing is one where programs can escape the type system. Another way to describe it is that values can change types.

For example, in C, I can create a pointer to characters, then tell the compiler I want to use it as a pointer to integers:

    char sz[] = "abcdefg";
    int *i = (int *)sz;

On a little-endian platform with 32-bit integers, this makes i into an array of the numbers 0x64636261 and 0x00676665. In fact, you can even cast pointers themselves to integers (of the appropriate size), or vice-versa:

    char *spam = (char *)0x12345678
    spam[0] = 0;

And this means I can overwrite memory anywhere in the system.*

_{* Of course modern OS's use virtual memory and page protection so I can only overwrite my own process's memory, but there's nothing about C itself that offers such protection, as anyone who ever coded on, say, Classic Mac OS or Win16 can tell you.}

Traditional Lisp allowed similar kinds of hackery; on some platforms, double-word floats and cons cells were the same type, and you could just pass one to a function expecting the other and it would "work".

Most languages today aren't quite as weak as C and Lisp were, but many of them are still somewhat leaky. For example, any OO language that has an unchecked "downcast",* that's a type leak: you're essentially telling the compiler "I know I didn't give you enough information to know this is safe, but I'm pretty sure it is," when the whole point of a type system is that the compiler always has enough information to know what's safe.

_{* A checked downcast doesn't make the language's type system any weaker just because it moves the check to runtime. If it did, then subtype polymorphism (aka virtual or fully-dynamic function calls) would be the same violation of the type system, and I don't think anyone wants to say that.}

Very few "scripting" languages are weak in this sense. Even in Perl or Tcl, you can't take a string and just interpret its bytes as an integer.* But it's worth noting that in CPython (and similarly for many other interpreters for many languages), if you're really persistent, you can use `ctypes` to load up `libpython`, cast an object's `id` to a `POINTER(Py_Object)`, and force the type system to leak. Whether this makes the type system weak or not depends on your use cases—if you're trying to implement an in-language restricted execution sandbox to ensure security, you do have to deal with these kinds of escapes…

_{* You can use a function like struct.unpack to read the bytes and build a new int out of "how C would represent these bytes", but that's obviously not leaky; even Haskell allows that.}

Implicit conversion

Implicit conversion is really a different thing from a weak or leaky type system.

Every language, even Haskell, has functions to, say, convert an integer to a string or a float. But some languages will do some of those conversions for you automatically—e.g., in C, if you call a function that wants a float, and you pass it in int, it gets converted for you. This can definitely lead to bugs with, e.g., unexpected overflows, but they're not the same kinds of bugs you get from a weak type system.

C isn't really being any weaker here; you can add an int and a float in Haskell, or even concatenate a float to a string, you just have to do it more explicitly.

Or, using the alternate definition, the value's type isn't changing from int to float, the program is creating a new float by calling a conversion function on the int, but the original int is still sitting around in whatever variable (or temporary location) it was always in.

With dynamic languages, the whole idea of implicit conversion is pretty murky. There's no such thing as "a function that wants a float" in Python or Perl. But there are overloaded functions that do different things with different types, and there's a strong intuitive sense that, e.g., adding a string to something else is "a function that wants a string".

In that sense, Perl, Tcl, and JavaScript appear to do a lot of implicit conversions ("a" + 1 gives you "a1"), while Python does a lot fewer ("a" + 1 raises an exception, but 1.0 + 1 does give you 2.0*).

_{* Actually, in modern Python, that can be explained in terms of OO subtyping, since `isinstance(2, numbers.Real)` is true. I don't think there's any sense in which `2` is an instance of the string type in Perl or JavaScript… although in Tcl, it actually is, since _everything_ is an instance of string.}

But it's pretty hard to put that into anything beyond a loose, intuitive sense. Why shouldn't there be a + function that takes a string and an int? Obviously (string, int)->string is a perfectly valid type for a function to have, since it's the type of, e.g., the index function.

Another meaning for strength

There's another, completely orthogonal, definition of "strong" vs. "weak" typing, where "strong" means powerful/flexible/expressive.

For example, Haskell lets you define a type that's a number, a string, a list of this type, or a map from strings to this type, which is a perfect way to represent anything that can be decoded from JSON. There's no way to define such a type in Java. But at least Java has parametric (generic) types, so you can write a function that takes a List of T and know that the elements are of type T; other languages, like early Java, forced you to use a List of Object and downcast. But at least Java lets you create new types with their own methods; C only lets you create structures. And BCPL didn't even have that. And so on down to assembly, where the only types are different bit lengths.

So, in that sense, Haskell's type system is stronger than modern Java's, which is stronger than earlier Java's, which is stronger than C's, which is stronger than BCPL's.

So, where does Python fit into that spectrum? Again, dynamic types make this murky.

In many cases, duck typing allows you to simulate everything you can do in Haskell, and even some things you can't; sure, errors are caught at runtime instead of compile time, but they're still caught. However, there are cases where duck typing isn't sufficient.

For example, in Haskell, if you have a doubly-optional value (e.g., from calling zip_longest* on a list of optional values), you can distinguish None (filled in by zip_longest) from Just(None) (an actual None value from the original list). In Python, they're both None—that's why zip_longest needs a fillvalue parameter, so you can pass an alternate sentinel.

_{* I'm using Python terminology and syntax instead of Haskell here, so people who don't know Haskell can understand. I assume people who do know both languages can figure out what I mean.}

For another example, in Haskell, the concept of an empty list of integers makes perfect sense; in Python, an empty list is an empty list. So, in Haskell, you could decide that reducing + over that list should return 0*; in Python, you can't.

_{* In fact, Haskell doesn't let you do this; if you call the reduce function with no start value on an empty list, you get an error. But its type system is powerful enough that you _could_ make this work, and Python's isn't.}

On top of that, in many cases, often arguable whether Python really is simulating a stronger type system. For example, in Haskell, if you have a function that returns either a string or nothing, it returns Maybe String; to access the string, you have to use Just on it, even after testing for Nothing. Normally, you do this with pattern matching:

    case spam:
        of None:
            print('I got nothing')
        of Just(s):
            print('I got {}'.format(s))

In Python, the same function would return either None, or a str; to access the str, you just access it directly after testing for None, no need to pull it out of a Just. So, does this mean Python is simulating disjunctive types, and its if not None is doing implicit lowering? On the one hand, you get effectively the same error in Python trying to call, e.g., len(spam) if spam is None as you'd get in Haskell. On the other hand, if you call, say, __sizeof__ or __lt__ without testing for None, it will work in Python, and the equivalent would definitely not work in Haskell.

I think this is a matter of degrees; the cases where Python gets it "wrong" according to Haskell are just more examples of where duck typing isn't as powerful as Haskell's type system. (And, conversely, the places where Python gets it "right" are examples of where Python is more concise and readable without giving up any power.)