This seems a bit overly optimistic, but it's nice to see this kind of positive attitude for a change instead of "Why invent yet another language?".

I am very excited about Swift and I have some respect for what Rust is trying to accomplish, but none of these languages have much foundation in category theory or the ambitious kind of type system that Haskell has been pioneering for a while. And the fact that Go is seeing some traction is a stark reminder that we still have ways to go.

Still, it's nice to see some of these concepts becoming mainstream (lambdas/closures, fold and functional iteration, phasing off inheritance based polymorphism, etc...).

This is even more applicable to Rust where the system that deals with lifetimes and ownership are based on ideas that evolved out of linear/affine type systems that can track resource consumption.

Funnily enough linear type systems have a direct connection to linear logic, and via the Curry-Howard-Lambek there is another correspondence to symmetric/closed monoidial categories. Even though sometimes it appears like there is no theory behind these languages, Rust has definitely been aggressively importing some cutting edge thinking from programming languages.

I think that Bob's (the OP's) is merely saying that the design of languages such as Rust and Swift has been strongly influenced by developers well-versed in type theory (the foundation of programming languages) as well as software development (the point of programming languages) -- this is true for Haskell as well, as he points out, but he is just noting that these well-designed, informed languages are becoming more mainstream.

Also, Rust's type system is more complex than you might expect, involving something called linear/affine types: http://en.wikipedia.org/wiki/Substructural_type_system#Linea...

The type system in Haskell is not complex and that's a reason to like it. Everything derives from 6 simple rules:

http://en.wikipedia.org/wiki/Hindley%E2%80%93Milner_type_sys...

Simplicity in a type system makes it very expressive - contrast with complex, concrete type systems which do not allow e.g. Church numerals or linked lists to be represented within it.

Hence importance of having Apple and Microsost bringing ML derived languages into the industry.

Is that a requirement? If so, why?

Having used a system with higher-level type constructs I find myself thinking in them. So many complex sequences of operations shake out as something much simpler once you realise the right monad, or arrow, or so on. And it becomes painful to manually expand it out; it would be like working in a language without generics. So I couldn't stand to work in a language that can't express these concepts; I need either higher-kinded types or something equivalent to them.

And for a type system that powerful, I'd want some assurance it was correct. Theoretical underpinnings are not the only way to do that - if it had been in use a while, and had enough tests and experience to give me a decent level of confidence, that would be enough. But for a new language I think the only way I could be confident enough to use it would be if it had this kind of theoretical backing.

FWIW it seems that many of the Rust developers want this for Rust.

However, the Rust team is trying to get 1.0 out the door, and higher-kinded types would be backwards compatible, so I don't think we'll be seeing them for a while. There's also questions about the overhead and syntax that would need to be resolved for them to make it in.

I have no idea if a similar problem is possible with Rust, but it'll be interesting to see how HKTs enter the scene if and when they do.

You can sometimes make up for this with lots of testing and detailed code coverage analysis, but that can be a huge amount of effort, and might be impossible in some cases (for example: how do you easily test that your program behaves well when it runs out of disk space?). Static types can fix this with a lot less effort.

/s

(On a more serious sarcastic note, searching over your impure state space is somewhat more challenging. It's unclear how very finite it is, but I'm sure some reasonable spherical user assumptions will trim it to feasible exhaustiveness)

Functions can opt-in to the possibility of receiving 'Nothing' as a potential value. The tooling provided to Haskell can then also tell you when your function has not exhaustively handled every possible type of input (including Nothing).

In this instance, the type system provides robust automated coverage of your function's input domain w/regard to null values. Obviously type systems can't do everything, but they are the most light-weight formal methods we have.

For the same reason people like unit tests. Only with mathematical rigor, checked by the compiler, and inspecting all aspects and code paths.

I actually don't understand why those who advocate for TDD don't rather advocate you write everything in something like ATS. Every single argument I've seen for TDD would apply even more strongly to doing that.

I tried doing TDD style development for a while, and I found these two things to combine to make it unfeasible"

1) For significant amounts of real-world code writing correct tests is actually harder than writing correct code.

2) In real life you end up discarding much of the code you write.

This means that you end up spending a lot of extra effort to write well tested code that will never end up being run in production.

I find quality integration tests combined with some form of analysis of the more critical parts to be far more productive than TDD.

The equivalent for type-safe would perhaps be type-optional languages like Qi/Shen. I haven't messed around with it much, but it seems to offer rapid prototyping with an enforced type-safe mode that can be turned-on later. Unfortunately its inference engine seems to be inferior to ML or Haskell right now, so transforming code from non-typesafe to typesafe is a bit of a slog (I last tried it a year or two ago, so it may have improved in the meantime).

I forget who said this, but programming language design is basically "applied psychology", and if you admit that, your language will go further. For example, the difference between Python and many similar languages is that it looks nicer and is "intuitive" to read, and that contributes a large part (perhaps the largest part) to its popularity.

He believes a few things that in concert arrive at the idea that typed languages are the only interesting point in the design space thus uncovered.

He begins by first rigorously defining what he means by "type". In particular, it's a piece of information computed statically (i.e. using the syntax of a program alone, nothing of its "runtime") which constrains the eventual values the dynamics must take. A "type-safe" program is thus one where the dynamics correspond to the predictions of the statics—in particular, a program respects its type as it operates and the program "does not get stuck" i.e. does not hit an undefined state.

From here he notes the power of union types, types which are "either" A or B. Given a sufficiently large union type, one which reads "either this or that or this or that" for every possible result value of running a program then you can create a "safe" semantics for any language using this type—every program has this giant union type and every choice of dynamics ends up in this giant union type. He then notes that while this genuinely is "type safe" it's kind of boring. Further, all "dynamically typed" languages inherit this boringness since they expose no "interesting" static structure.

(Even if you dislike the argument above, and there's plenty that might cause someone to dislike it, it's worth noting that this line of thought arrives very smoothly at the concept of "hybrid typing" being explored by things like Typescript and Facebook's Hack.)

One of the primary things you could argue with at this point is that having more sophisticated types is interesting and valuable. To this end Harper focuses on what he calls the "Holy Trinity" of logic, type theory, and category theory. The idea is backed by the Curry-Howard-Lambek correspondence which shows that each of these fields is but an image of the other two. In Harper's (paraphrased) words: you know that something is genuine and lasting if you can show it exists in every branch of the Holy Trinity.

So he declares basically by fiat that things which reflect the Holy Trinity are interesting and things which do not are not. Obviously you can levy many arguments to disagree with him here.

But it's hard to declare that the things discovered in the Holy Trinity are boring. This is where cutting edge research on things like Homotopy Type Theory and general Dependent Types live. This produces compelling arguments to back up why "null pointers are a billion dollar mistake" [0] or why if-branching and booleans lack meaningful providence [1]. These ideas, even if they are stripped of their Ivory Tower gleam, are what are slowly percolating into mainstream languages with names like "Enum"++ in Swift and "Optional" in Java.

---

Generally, to finally provide my personal take, I think that Harper is dead right so long as you consider programming languages to be "means of expression". I think that's an interesting POV on languages generally and it's not totally at odds with the more everyday POV that languages are "tools for creation". But they are sometimes at odds since they place the brunt of what they demand of a language at different places. Ultimately, if Harper's program is successful many years from now we'll all use shadows of whatever futuristic grade of dependently typed language gives us maximal expressiveness while nesting it in an ecosystem of tooling and libraries which aid construction of interesting things. I have no idea if Harper's "one true language" exists, but it does seem like increasingly the world is moving that way... thus Harper's enthusiasm about Swift and Rust.

[0] http://qconlondon.com/london-2009/presentation/Null+Referenc...

[1] http://existentialtype.wordpress.com/2011/03/15/boolean-blin...

This is likely because my first encounter with formal reasoning was in the field of classical philosophy, and many things I empirically knew to be wrong had been "proven" from first-principles in the literature.

That's a very concise way of putting it. My personal issue with a lot of what Harper writes is not that it's technically wrong per se, but it only makes sense if one is discussing programming languages purely from first principles. Harper seems really taken aback by the popularity of Haskell compared to SML in the functional programming space, and yet can't seem to wrap his head around the reason for the adoption is not so much about having ideal PL semantics but the fact Haskell has an industrial-strength cross-platform compiler with meaningful answers to modern problems ( FFI, concurrency, package management ). Haskell is rather unique in that it has a community that seems interested in building a language as both a "means of expression" and "tools for creation" and we've made some compromises between the two, no doubt. A lot of Harper's comments ( and many PL theorists ) seem to completely casually gloss over the gritty compiler engineering as if it's just beneath them, and that guys like SPJ and Simon Marlow are just wasting their time working on faster runtime systems and optimizations when he thinks they should be working on foundational "Holy Trinity" problems.

That's a tempting view to take regarding theory vs. practice but I don't think it holds up when you consider Harper's whole body of work and the substance of his criticisms of Haskell: He has published extensively on ML implementation and, as far as I know, has offered criticism of (GHC) Haskell's type system where it is unsound, never the (really cool) work being done on the RTS.

In fact, his post on exceptions in Haskell and his criticism of monadic IO led me to work on some alternatives (algebraic effects) which, oddly enough, helped me understand Haskell's approach at a deeper level. I wouldn't have known about the issue if it was only on the bug tracker.

However, there is research into untyped/gradually typed languages. See, for example, [1] for combining gradual typing with type inference, and [2] for a list of papers formalizing the runtime semantics of a language that allows both typed and untyped functions.

[1] Gradual Typing with Unification-based Inference [2] http://homepages.inf.ed.ac.uk/wadler/topics/blame.html

"Type-safe" means that the type of a value is always correct. For example, if I have a function of type Int -> String, a type-safe language will not let me pass it a Float argument, no matter what I do. A type-unsafe language might let me, if I can "cast" the value to a new type without changing its representation (ie. it's still a Float, but the type-checker now thinks it's an Int).

Type-safe languages are preferred over type-unsafe languages, since they avoid all kinds of problems like undefined behaviour.

"Statically-typed" means that every variable, function argument, return value, etc. is associated with (usually one, "principle") type. These may be written explicitly, inferred, etc.

"Dynamically-typed" means that all values use the same static type. This type is usually a recursive sum of lots of useful types, for example (in pseudo-Haskell):

    type UniType = I Int | L [UniType] | F (UniType -> UniType) | O (Object UniType) | E Error | ....

Many users of dynamically typed languages don't realise that static types are being used "behind the scenes", because:

* All of their code has the same type, so there's no need for any annotations, inference, type-checking or compilation. Many believe it's because there are no types, rather than the fact these have been made trivial.

* Sum types and recursive types aren't widely known, especially to those using dynamic languages who don't need to care about types at all. Hence, many make the assumption that "I don't know how to type this code" implies "this code is untypeable" and refer to such things as 'dynamic behaviour'.

* Many dynamic languages are implemented in languages with unsafe, anemic type systems, like C. Since 'dynamic behaviour' can't be represented safely in these languages, it's implemented unsafely, reinforcing the assumption that dynamic languages aren't type-safe.

* Other dynamic languages are implemented in themselves, which hides the UniType underlying it all.

* A few dynamic languages are implemented in languages with strong, safe type-systems, where the UniType is explicit and obvious, but most users stick to the dynamic language and never look at the implementation.

* Many languages confuse terminology by using the word "type" to refer to tags (the "I", "L", "F", "O", "E", ... in my UniType above); for example "if (get_type(x) === "int") {...}" instead of "if (get_tag(x) === I) {...}", but this doesn't make sense: types don't exist at runtime and even if they did the "get_type" function would always return the same thing (the UniType). The fact that many languages represent tags as strings doesn't help either!

* Dynamic languages allow any variable to contain any value, but most of the time only a small part of this domain is expected. This causes most functions to be partial, but whenever this partiality is encountered, one of two things happens: a) the program fails with a "runtime type error", which is incorrect and doesn't make sense, b) the language supplies an error value in response, and is regarded as doing something "unsafe" or "untyped", when in fact error values are perfectly valid (I've tagged them with "E" above) and it's the programmer's fault for not handling all of the possible cases they (implicitly, via UniType) asked for.

The work you've linked to is about selectively weakening the types in a program, so rather than having the whole thing be unityped or the whole thing use distinct types, instead we can write code both ways, then automatically strengthen/weaken the types and lower/lift the functions, so that they interoperate. The types are always safe, but the unityped part will have many more partial functions than total functions.

In contrast, we can't (yet) make a type system "gradually safe", since a single unsafe cast can break anything else ("ex falso quodlibet"). It would require some kind of para-consistent logic, which I've never seen in type theory. The difference is that 'gradual types' let us say "Any" instead of "Int", whereas unsafe types let us say "X" instead of "Int", for any "X".

The dynamic typing folks don't care to provide a definition of types and overload the term to mean many things, the static typing folks have a rigorous one.

Definition of type (n)

category or kind: a group made up of individuals or items that have strongly marked and readily defined similarities

person or thing: somebody or something regarded as belonging to a group or category by virtue of having the main qualities associated with it

particular kind of person: a person regarded as having a particular temperament or characteristics.

> tractable syntactic method for proving the absence of certain program behaviors by classifying phrases according to the kinds of values they compute.

Technical books often define terms in a way convenient for the material and approach they are presenting, not in a way that is general to the field; particularly, if terms are used in different ways within the field at the time they are written, they typically have to choose for consistency within the book -- that doesn't make that definition the right one for the field, it simply means that's what the term means in the context of the work.

And, ultimately, the semantic argument is meaningless. It doesn't really matter whether you have static and dynamic types in different languages or you call the former "types" and the latter "tags" (except that the former terminology actually explains the difference in a way that tells you what the difference is every time you use the terms.) The substance is the same, and getting worked up over line-drawing regarding the boundary of "type" is a distraction from any discussion of substance.

However, my main point was that "type-safe" and "type-unsafe" are not synonimous with "static" and "dynamic". The original question was probably about static vs dynamic, but I think it's important to point out the distinction; especially when reading material like Bob's that assumes some familiarity with the terms.

From a Curry-Howard point of view:

* Static types are logical formulas, values are their proofs.

* "Dynamic types" are the clauses of one big disjunctive formula: Int OR Bool OR Error OR Array OR String OR ....

* Safe type systems are sound logics; they don't let us prove "FALSE".

* Unsafe type systems are unsound logics; they let us prove "FALSE", and hence anything (ex falso quodlibet).

Importantly, the big disjunctive formula used by dynamic languages is trivial to prove. For example, I can prove it with a constant like "10", without having to perform any computation. Trivial formulas are equivalent to "TRUE" (the logical proposition, not the boolean value!), hence dynamic languages only allow us to prove "TRUE". This makes them type-safe, since they can't prove "FALSE", but it makes them uninformative: proving "TRUE" doesn't tell us anything we didn't already know!

The only languages which can be type-unsafe are static languages, since they're the only ones which can express types other than "TRUE" (eg. "FALSE"). If a static language is unsafe, then there's essentially no benefit to it having static types, since we can't trust any of the information it gives us.

That's why academics prefer type-safe languages, whether they're static or dynamic, whether or not you agree with Harper that one's a sub-set of the other. Unsafe languages specifically hinder our reasoning. Sure, we might be confident that our particular C program is safe, despite C being an unsafe language, but we can't generalise that to "for all C programs 'P'...", which makes it difficult to study languages-in-the-abstract. Of course, we lose nothing if we make the language safe, so we might as well do that (eg. define a safe sub-set of C like http://compcert.inria.fr/doc/ ).

Essentially, asking "Why do programming languages researchers seem so enamored with type-safe languages?" is like asking "Why do Mathematicians seem so enamored with sound logics?". Of course, there is research in non-monotonic logics, para-consistent logics, etc. and that's great, but if we're going to have an academia-vs-industry debate, then I think the unsound logics, and hence the type-unsafe languages, have the taller ivory towers.

Of course, there are many other ways that static and dynamic languages may be safe or unsafe (eg. memory safety, thread safety, non-total, etc.). (Safe) type systems can eliminate some of these problems (eg. memory safety with linear types), but just because it's possible with some type-system feature doesn't make a language lacking that feature type-unsafe; the worst we could say is that it's type-uninformative.

For example, Haskell's type system doesn't distinguish total functions (eg. via a data/codata separation). This makes Haskell unsafe with regards to termination/cotermination, but it doesn't make Haskell type-unsafe. Since dynamic types are completely uninformative, they can't solve any problems that type systems are suited to, but that still doesn't make them type-unsafe.

For example, it's not that dynamic languages allow type-unsafe operations like passing strings to "+ : Int -> Int -> Int", it's that dynamic languages don't allow us to restrict the type of "+" at all; we can't even specify that it's a function, let alone what that function's input and return types are! If it generates a run-time error that's fine, since "run-time error" is a perfectly valid value, as far as the type system's concerned.

Now, we may think this is an unsafe thing to do, but it doesn't make the language type-unsafe. It just makes it "semantically unsafe" or "unsafe without lots of tests" or somesuch.

Sure it works for the lonely coder or small shops, but as soon as the project scale increases anything goes.

Managers will push for features over unit tests, no one will refactor code without taking a few days with the caveat not all places will be touched and the team atrition will bring in new devolpers forced to read implementation code to be sure how to properly call the functions/methods.

Advanced static typed languages with HM type inference are a better option for large scale development.

Hence it may be that is harder to get funding for other types of research.

Other issue is that dynamic languages fo their very nature make tooling harder problem to solve, hence less interesting for many researchers, pressed with delivering quick results.

For example, consider the exponentiation function. When we raise an Integer to a positive Integer (2^2), the result will always be a positive Integer. You would expect the type of such a function to be: Integer -> Integer -> Integer

Except, that's wrong. When we raise an Integer to a negative Integer (2^-2), the result will be some fractional value. If we want to retain full numeric precision, it'll have to be something like a Rational. So now your function's type signature is: Integer -> Integer -> Rational.

But that's also wrong. So what are we to do? Maybe positive and negative integers should be separate types? PositiveInt and NegativeInt, let's say. But then what about subtraction? Should it's type signature be: PositiveInt -> PositiveInt -> PositiveInt? What about 2 - 3? Whoops!

So, there's no way, with types alone, to simply and accurately encode all basic math operations. All you can do is throw a Domain Error when someone violates certain boundary conditions, and tell them to check their arguments before deciding which function to call.

...meanwhile, dynamic languages will happily return whatever type is most appropriate for the arguments given.

In a dynamic language, your mistake passes silently; the result you thought was an integer is actually a fraction, but your program carries on going. Maybe it turns it into a string and uses it as a file path, so something you thought was a simple filename has a slash in it. Maybe that breaks your security model. Pretty unlikely in this specific case, but that's how bugs happen.

What really interests me are languages like Julia or Dart, with their gradual typing, or Clojure and core.typed. You retain the flexibility and speed of development of a dynamic language, with the possibility of later adding type safety guarantees to your code without having to rewrite it from scratch.

How is this different? When you compile, your compiler will say "Hey, the requirements have changed, your code is wrong here, here, here, and here," and you modify those things, and you're done.

1- You wrote tests which start failing. You must now spend time changing them. You don't manage to avoid working in order fix your safety net (be it static type checking or your test suite).

2- You didn't write the right tests, and don't notice the change in behavior. The program later fails in production.

The big theoretical drawback is that you get an unbounded increase in program length. The reason is that programs in a typed language are proofs, and proof length cannot be bounded above by any computable function of proposition length.

The big practical drawback is that we just don't know how to explain proofs to a computer very well.

It can express that 2^2 is positive and an integer, that 2^-2 is positive and a rational, and that 2-3 is an integer but not necessarily positive.

I don't know. Haskell and OccamL, for one, are much, much faster than Python and Ruby.

Of course, there was no Ruby back then. The comparison was against C++, Ada, Awk (!?), something called "relational Lisp" and a variety of in-house languages. Haskell comes just after Lisp in development time, with a mind-boggling 8 hours! (No C++ times were reported... heh heh). The Haskell solution also got top marks in general, not just in development time. As a downside, some of the reviewers didn't understand it because they weren't well-versed in functional programming, and felt some of the Haskell code was "tricky".

Because ad-hoc tests, tests, tests, ... with the verbosity they add (tons of uneeded tests that could be type checked instead), and the arbitrariness (you have to remember to add a test -- instead type checks are always run) and checking only some code paths (type checks propagate everywhere) are in any way better?

Not to mention that you can have type checks AND (less, complimentary) tests at the same time.

At least us software guys have an advantage: the hardware guys will always have to live with failure rates.

But the standard happily declares software to be bug-free, as long as you do the required testing and stuff. ;-)

But aren't those tools (that can build trust) something that could be built-in & available to all users of the language, therefore avoiding "reinventing the wheel"? Why would you want to do by hand something that can be automated?