Personally, I get a lot of use out of C++ and actually enjoy using it. Sure, it can get complex at times when doing really hard stuff, but it'll force you to stop and understand the implications of what it is you are doing thus forcing you to do it correctly.

C++ won't let you get away with half the crap that many other languages will. Knowing C++ will make you a better programmer in all languages. Try it and see.

A little background: I started coding in x86 ASM (because some idiot told me that was the only language computers understood when I was 14), switched to Pascal, then C, then C++, then I got my 1st job as VB6 oracle-to-form data pusher, then I did 5 years of very exciting C++ library development, i.e. other programmers were my customers. [highly recommend this - much easier to market&sell than any other market, learn from DHH, Trolltech or MySQL AB]

In 2002 I moved to higher level languages: Python, some Perl and lots of Ruby and Javascript. And honestly I'm not that more productive coding Ruby than I was ever in C++. Yeah, it's nice to read an entire file, break it down into an array of lines, grep them with some regexp and apply a function on each result - all that in a single line of code, but... well, read on:

C++ allows for some beautiful type-safe and performant high-level design constructs that Java, Python or even Ruby can't touch: just combine multiple inheritance with medium-complexity templates and see what you can do. Only Haskell gives me the same type-safe, nearly as fast and flexible (yet very high level) material to build software of. Being a realist, I don't see Haskell ever "making it" into the industry, but the updated C++ definitely will.

Duck typed languages are nice, but require enormous layers of unit tests to protect yourself from stupid things like misspelled variable names or even renamed/missing code files. Building with Rails without unit testing is a suicide. Doing Haskell without unit testing is actually isn't that bad. And this requirement to test the shit out of your code [2-to-1 ratios aren't uncommon] kills the spirit of exploratory programming for me. Moreover, 200x speed penalty means that simple O(n) algorithms won't cut it for an average data set, forcing you to build more complex solutions just because your language can't scan a big-ass array under 10ms. Moreover, most high-level languages live inside of a VM, which makes them useless for general purpose software development: they represent "fuck you" approach, i.e. they'll easily call your C library, but won't let you easily call them. JAR is as good to me as my swap file: I can't afford to carry around JVM on my back just because I want to call your Java code from Haskell or C++.

Which is why most of us don't write unit tests when we're in the exploratory phase of something.

"Moreover, 200x speed penalty means that simple O(n) algorithms won't cut it for an average data set, forcing you to build more complex solutions just because your language can't scan a big-ass array under 10ms."

I think that you and I have different ideas of what an "average" data set is, then, or the performance constraints expected of it. Although given the number of mathematicians and professional scientists who crunch their data in Python these days (see: NumPy, SciPy), I'd also suspect that maybe you should investigate this a bit more deeply.

For myself, once I got to a point where I could "think well" using syntax at both a high level of abstraction and write solid C code at the system call level I found that the space of "appropriate" problems for C++ had shrunk to almost nothing. By itself, if it's the only language you're ever going to learn, C++ is not a bad choice. But in the modern world of polyglot development on CPUs too fast to measure, it's almost never the best one.

I could also keep the systems stuff in C functions and use an FFI to call them, but I've never found a language that enabled me to write GUIs more easily than with plain old C++ and Qt. Java (or Java from Clojure) is slow and ugly as sin, and most other languages either use bindings to C/C++ libraries or have incomplete, mostly unsupported kits. C# and family is perhaps an exception, but not very portable.

No one doubts that C++ has been used to do hard things. They doubt that it was really better than using C++ with some higher-level language.

This will facilitate the "Accelerated C++" approach, namely teaching the libraries first, and the base language later.

The other side of this is that it will likely also widen the gap between those who use C++ libraries, and those who write them.

None of the corner cases that make C++ so hard to parse have actually been removed, so tools are still hard to write and unreliable. Nothing has been done to allow a user to express where pointer aliasing can occur and, more importantly, where it cannot, so many forms of important numeric optimizations are still extremely difficult. There are still at least three languages possibly intermingled at any given time--macros, template computation (a la Boost), and regular C++, each with its own style of assertion now. Rather than relaxing a restriction on type ordering they invented a whole new function syntax. Managing the variable closure around upward funargs (returning a lambda) is the programmer's responsibility, with a subtle stack smash if you get it wrong, or if anything you call with a lambda stores it somewhere behind your back. Two types of characters aren't enough; now char, char16_t, char32_t, and wchar_t all coexist. You can now create a const char string two ways, one of which is UTF-8 encoded and one of which is encoded any damned way the compiler writer pleases. The raw string syntax, another invention, further complicates the lexer. Objects now silently need a move constructor in certain circumstances to avoid silent, behind the scenes copies. You can't even reliably say "x = y;" any more; operator = can be deleted from whatever type x is. Rather than put the onus on the (admittedly horribly overworked, due to the previous crap) compiler writers to constant fold, now we have constexpr, just like const except different!

This is not a programming language. This is a train wreck. Supposedly this language is "close to the metal" but even the simplest operations can perform unbounded amounts of work behind your back (for example, due to implicit conversions, or operator overloading). This language is consumed with a focus on the irrelevant.

The Reg article, while very snarky, is not an unreasonable position to be taking here.

Though my bet is still on spam and anti-spam, they have better incentives.

With the operator suffix extensions, though, I think C++ has broken new ground again. It is now the only allegedly serious language that requires a symbol table to lex.

Yick. Arc may be a "hundred-year language", but if the C++ committee delays to 200A, that would make C++ a six-thousand-year language (8202 in decimal).

[Confession: I once caused the Bell Labs 'cfront' tape to be ordered, and loaded into unsuspecting Vaxen innocently running BSD Unix, at an otherwise reputable school. I allege, in my defense, that I mostly used it only as "a better C" (back then, C did not have function prototypes); and that, at the time, no other grad students were noticeably (further) deranged by the availability of C++.]

I'd say that for some it makes sense. Updated C++ will be a fairly high-level language with absolute access to bare metal and near 100% compatibility with C, that's very unique. And the world is in a desperate need for a deterministic (i.e. optionally garbage-collected) bare-metal replacement for C for systems programming. D could be it, but Walter just won't let the community to have some control of it.

Moreover, I suspect some of these improvements around the type system will allow C++ compilers to generate faster code than C compiler could ever achieve because programmer's intentions will be better expressed.

I only wish they expanded the standard library more. As far as I'm concerned, C++ doesn't really have a standard lib: STL is often awkward to use and very limited, for instance std::string is next to useless and iostream is such a pain... Then we have Boost, it comes with insane build times, incomprehensible compile error messages and multi-platform issues, so most C++ programs mostly fall back to C lib for basics like I/O, or to platform-dependent frameworks like MFC.

I also hope that external templates will work as promised. In the current form, templatized C++ works well only in a form of one giant file where everything is included in the correct order. Perhaps this is why every single C++ project in my relatively long C++ career has been using the language in "C with classes" form despite my best attempts to convince people that templates aren't always evil.

Speaking of Python and C++: possibly some of us would wish to write Python and get C++. Mark Dufour has been persistently plugging away at a Python-to-C++ compiler (and welcomes help): http://shed-skin.blogspot.com/

I don't doubt that most of the (non-idle, non-phone/TV/etc.) CPU cycles in the world are probably being burned by executing code written in C/C++. But the stuff in my shopping bag probably also did better than 90% of its movement by truck/train/ship.

P.S. Mark Dufour's ShedSkin may have you covered for 100x of that 200x speed hit you mentioned in another thread: http://www.hxa.name/minilight/#comparison is instructive (one small case, restricted, etc. Void where prohibited by law).

http://www.jwz.org/doc/worse-is-better.html