I suppose the one thing that Go does well (compared to C++ or Java) is builtin concurrency and communication across tasks.

I'm a Java-literate C/C++ programmer. I would avoid writing straight Java code at all costs; I find it immiserating. Here are some reasons off the top of my head that Golang is more pleasant to work in:

* The syntax is deliberately streamlined, including implicit declarations, semicolon insertion, lightweight expression syntax, the capital-letter-exports-a-symbol thing

* It has fully functional idiomatic closures

* Interfaces get rid of adapter class BS

* The table type (maps, in Golang) is baked into the language, like Python, not a library, like C++ and Java

* Clearer, more direct control over memory layout; data structures in Golang feel like C

I don't know if Golang's standard library is that much better than Java's, but it was obviously designed carefully by and for systems programmers, so I find it remarkably easy to work with.

It also feels like a much smaller system than Java. Almost every time I write a significant amount of Golang code, I find myself answering questions by just reading the standard library source code. It's easy to get your head around. I've written compiler/runtime-level code for the JVM and I still don't have a great grip on all of Java.

Another cool aspect of the last point (Go being small and lightweight) is that if you've got gcc and mercurial on a supported platform, building latest go from source is as easy as:

hg clone http://code.google.com/p.go cd go/src ./make.bash

Got to build a local copy of the JVM and/or JDK for some reason? Good luck with that (even ignoring all the licensing, OpenJDK vs closed, etc)

In other words, it is possible to express anything in one Turing-complete language that is possible to express in another.

http://www.paulgraham.com/power.html

So to measure the complexity of a program P we have to write another program G that generates P and then take G's length. It's not a given that a generator for a verbose language is necessarily longer than a generator for a terse language.

But more importantly, what we want to measure is the mental effort required to write and understand code in different languages. To measure that effort in terms of Kolmogorov complexity, we'd have to write a program that generates our states of mind while programming.

Good luck with that ;-)

You have to think of it in terms of compression, because that's what Kolmogorov complexity is really about. It is certainly possible that after compression J1 is shorter than T0.

In terms of this definition we need two Turing machines, one whose transition function can be proven to be the shortest that describes L1. And another one whose transition function can be proven to be the shortest that describes T0. These two Turing machines or transition functions are compressed representations of L1 and T0 respectively.

What I'm saying is that the one describing L1 can be shorter than the one describing T0 even if L1 is longer than T0.

I have not only the transition function but 6 other entities to define to accomplish this (really 5 barring the 'blank symbol') so I am certain that it is not only possible but likely much smaller than one might think at first. With such a small M (that penalizes attempts to compress T0 and J1), and a large T0 [such as the code base for Microsoft Windows ported to F#] and an even larger J1 [such as the same code base written in C], we still see my desired result.

"More formally, the complexity of a string is the length of the shortest possible description of the string in some fixed universal description language".

M is not that universal language. M is a program written in the universal language of Turing machines. I guess that is our misunderstanding. If you define M itself as that universal language but M only ever does that one contrived thing you said (i.e. there is only one program written in that language) then it is not a universal language any longer and any comparison of lengths would become completely pointless.

Then, the question becomes, how much logic do we have to put into M itself to achieve the above? I posit that it would not be very much in fact (but that should be demonstrable if someone were sufficiently motivated).

I think KC requires that M be constant across the texts you wish to compare (so that texts across various languages are normalized to an underlying architecture).

If one views M in this light (as an attempt to level the playing field across competing hardwares/algorithms/languages), then we could see it as the concatenation of the following:

1 - the specification of the raw hardware. 2 - the microcode for the hardware. 3 - BIOS, drivers, HAL, OS, etc. 4 - language implementation/runtime/shared libraries etc.

Now, to compare any two real-world implementations, we need to port them both to the language implemented in #4, for running on the hardware specified by #1-3 and sum their length with M's length.

And if we want to compare two algorithms from different languages (but running on the same 'machine'), we can shift the codes in #4 into the 'texts' of the programs themselves (so as to keep M constant between them). We can do this all the way down to the minimum M required to be Turing complete (and nearly all the way 'up' to the programs under test); i.e., the division between M and a particular text, L0, is somewhat variable.

First of all, I doubt that this is even possible, because the input, i.e L1 and T0 is arbitrarily long in the general case, while the algorithm's length must be determined before it sees any input, i.e it is constant. That's related to the invariance theorem: http://en.wikipedia.org/wiki/Kolmogorov_complexity#Invarianc...

But even if it were possible or if we agree to arbitrarily limit the size of the input for practical purposes, you wouldn't have achieved your original goal, which was to show that a verbose language can never be as expressive as a terse language.

Your universal language is neither a Turing machine nor any other generally accepted universal language, so no one would agree to accept this particular language as a test of Kolmogorov complexity.

If you can disprove the invariance theorem you may have come up with a valid criticism of Kolmogorov complexity itself, but you were the one who initially based his argument about the expressiveness of programming languages on Kolmogorov complexity. If you invalidate Kolmogorov complexity you invalidate your own argument as well.

In any event, this has been a fun thought experiment.

I think I am suggesting to construct a particular universal language designed to make any compression of some reasonable T0 and J1 written in that UL (e.g., the source code of Microsoft Windows written in F# and C respectively) longer than T0 and J1 themselves, hence leaving them to be their own Kolmogorov limit (if you will). This seems entirely doable to me within the confines of the 5-tuples afforded me by the formal definition of a Turing machine (although I can see how others, including you, might be less convinced). :-) Maybe I, or someone else sufficiently motivated, will attempt it someday.

[1] http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.51.4...

[2] http://gafter.blogspot.com/2007/03/on-expressive-power-of-pr...

"In my mind, a language construct is expressive if it enables you to write (and use) an API that can't be written (and used) without the construct." - Neil Gafter

Once more, there is no such construct. All APIs can be written and consumed without language support. Otherwise, Boost would not have had all the new C++ goodies (as an add-on library) before they became part of the language itself. And, of course, one can always put the 'human compiler' to work to produce boilerplate that would be produced by the compiler if the language in question supported the feature in question (if there isn't an add-on library providing the feature in question). One can always do this but the discriminating one tries to avoid it and instead chooses 'expressive' (i.e., conciseness-facilitating) languages. [In the case of closures in Java, one [human-compiler] would merely use one-off interfaces or anonymous classes].

I will take a look at the other links but I gave up on the commentary after seeing such a blatant falsehood.

Yes.

    "* ^^"._? #> (a.!?(1, Get())) must be_==(Full(Answer(1))) ?~! eventually

"Are you sure about that?"

"Yes."

Do I need to explain?

Certainly if you intended to say more than 'yes' then you failed at adequately expressing the message to begin with so comparisons at that point are moot: i.e., two pieces of text must both express precisely the same ideas before comparisons of them make sense.

Of course this only really applies to long running apps (web servers), if startup time matters then certainly Go wins.

Perhaps I need to try Go out, but I just don't see what the selling point is.

In a word: hype. If Go did not have the Google brand name attached to it, it wouldn't have gone anywhere.

I actually did write a fair amount of Go code (my primary languages are C++ and Python).

On real, non-toy programs it is significantly more concise than C++, in the ballpark of Python code.

This isn't visible on code snippets (less than 500 loc) like the toy raytracer, but trust me: you'll see a big difference on 10k loc codebase.

Given that "kkowalczyk" talked about 10K line programs, what applications are you thinking of that are 10K lines and cannot tolerate a 0.1 second/3 MB overhead. Would you restart java everytime? Note that even a simple helloworld c program has about 0.01second/0.5MB overhead.

    $ time java HelloWorld
    Hello, world!

    real	0m0.101s

    $ time javac HelloWorld.java 

    real	0m0.511s
    user	0m0.833s
    sys	0m0.050s

FWIW, we went with go at my work instead of java because our application is memory-intensive, and there's huge gains there in go over java.

Somewhat anecdotally, I remember hearing Rich Hickey talking about how the JVM JIT loves stateless code. I'm pretty sure the vast majority of Java code in use is deeply stateful. I don't know how much of a difference this actually makes but it seemed like a relevant data point.

If you know what crt0.c does, you can see C is also pretty much the asymptote of what you're going to get from program startup, so it's a little silly to make the comparison.

One can always use one of the commercial native compilers for Java, any of them can easily produce static binaries with fast startup times.

Don't measure Java the language, by Oracle JVM the implementation.

Not really. Other languages are far superior in this regard (see Scala or Rust for instance).

"they ended up with a language with the expressiveness of C and the speed of python."

This is just pure troll. Apparently you have a Python variant of this renderer that matches the already impressive Go performance? Please reveal it.