Slashdot Mirror

There's a reason why people who care have a rich set of numeric datatypes.

I like to keep a rich set of numeric datatypes on hand just to show off. Some people think I have a hoarding problem, but I see myself as more of a datatype collector. And then there's this

Let's just open up my handy Javascript console in Chrome...

(0.1 + 0.2) == 0.3
false

It doesn't matter how many bits you use in floating point. It is always an approximation. And in base-2 floating point, the above will never be true.

If they're saying that JavaScript is within 1.5x of native code, they're cherry-picking the results. There's a reason why people who care have a rich set of numeric datatypes.

I'm sure you already knew this, but...

Curl is available back to at least AIX 4.3, directly from IBM.

http://www-03.ibm.com/systems/power/software/aix/linux/toolbox/alpha.html

I'm sure the simple search, or knowing the (horrible) IBM AIX site, was beyond their skill level.

Passing in registers is not a standard C parameter passing method

There's no such thing as "a standard C parameter passing method". Passing in registers is a perfectly legitimate C parameter passing method, used on several RISC architectures, such as SPARC, MIPS, 32-bit ARM, 64-bit ARM, and 64-bit {PowerPC/Power Architecture} (and probably most other RISC architectures), as well as x86-64 and z/Architecture.

If there are more parameters than fit in the registers available for parameter passing, or if the parameters are in the variable-length portion of the argument list, they might be passed on the stack, and if the called function has no prototype in scope, the compiler might be forced to pass everything on the stack, but, in all other cases, if the ABI supports it, parameters can and will be passed on the stack.

Except that the Watson programmed to diagnose diseases is still called Watson. A Watson programmed to guess prices would still carry the brand name "Watson."

"Watson, what exactly is cloud computing?"

"Well, sonny, since my API is billed per PVU per hour, let me tell you a story that my grandaddy told me:

Once upon a time, when men were men and bought their calculatin' machines from IBM, folks who were too poor to lease a calculating machine of their own would lease part of somebody elses. Then minicomputers came, and destroyed the good old days with fire and sword. Then microcomputers came, and slaughtered the minicomputers for their sins. But the minicomputers sucked, and had neither reliability, nor redundancy, nor Administrators to lead them. And Lo, the 'freedom' of the users turned to mere license, and became as ashes in their mouths."

Now, sonny, the minicomputers and users have repented and look to 'the cloud' to restore the glory of the old order, where man leased, and owned not, and Administrators watched over the users, and guided them.

Hardly, Red Hat dropped support for POWER with RHEL 5.

Not according to the horse's mouth:
http://pic.dhe.ibm.com/infocenter/lnxinfo/v3r0m0/index.jsp?topic=%2Fliaam%2Fliaamdistros.htm

It's a pretty cool critter, but I don't know if they actually sell it as a product. It might be something that they only use internally:

http://www.research.ibm.com/da/beam.html

http://www.research.ibm.com/da/publications/beam_data_flow.pdf

It's a pretty cool critter, but I don't know if they actually sell it as a product. It might be something that they only use internally:

http://www.research.ibm.com/da/beam.html

http://www.research.ibm.com/da/publications/beam_data_flow.pdf

Has Oracle/Sun also adapted LLVM/Clang for Solaris, or are they staying w/ GCC?

If you mean "as their official compiler", the answer to both questions is "no"; they have, instead, Sun^WOracle Solaris Studio's compilers. If you mean "in their package system", they could offer both, but currently only appear to offer GCC 3 and GCC 4.5.

How about IBM in AIX,

If you mean "as their official compiler", the answer to both questions is "no"; they have, instead, IBM XL C/C++.

or HP in HP/UX?

If you mean "as their official compiler", the answer to both questions is "no"; they have, instead, HP C/aC++.

I wonder if the base price includes PowerVM or a graphics card.

I wish I knew. The last Unix box I had was an Indy. Probably not the first thing, though - the Specifications page lists all of "PowerVM {Express|Standard|Enterprise} Edition" as "(optional)". According to the fine documentation, you can order "POWER GXT145 PCI Express Graphics" for your machine. If you want to run AIX on the box, I have no idea what the AIX level support is for this hardware. Forget CAD systems, though, this doesn't look like that kind of graphics hardware.

What is a "base price" for these machines, anyway? :)

Oh, thanks !
http://www-03.ibm.com/systems/power/hardware/720/browse_aix.html
I wonder if the base price includes PowerVM or a graphics card.
At 37 to 42 dB, the tower is a bit on the noisy side - on par with the old "tower" Mac Pro.

Linux is meant to run on lighter hardware, which means it tends to use less CPU.

Yes, it's meant to run on all sorts of lighter hardware.

Srsly, Linux is meant to run on whatever the hell can run Linux. That's it. Any other claim is bullshit.

Linux is meant to run on lighter hardware, which means it tends to use less CPU.

Yes, it's meant to run on all sorts of lighter hardware.

Srsly, Linux is meant to run on whatever the hell can run Linux. That's it. Any other claim is bullshit.

Could you elaborate a bit on this? I had the impression that D-Wave's users had to map their problem to fit what D-Wave computes, not the other way around. That would make comparisons with a specialized software solver appropriate, wouldn't it?

The blog post in question also includes a link to the source code of the specialized solver (Prog-QAP), and others have confirmed that it produces the same results as CPLEX, the general solver that D-Wave beat.

CPLEX is indeed slower than D-Wave, though newer versions have brought the factor down from 3600x to 14x. But again, CPLEX is a general solver, while D-wave is specialized hardware. The specialized software solver Prog-QAP is *much* faster than CPLEX, and gets a 12000x speedup over D-Wave when running on a single core.

But all of that is a bit old, and it may be that D-Wave has produced more impressive results after that. I hope D-Wave's approach results in something able to beat classical computers, even if it doesn't lead to a general quantum computer. But I really dislike all the secrecy they employed - that is not how science is supposed to work. The fraud speculations they have had to endure are entierly self-inflicted due to this secrecy.

I2 was purchased by IBM, their product i iBase purchased by IBM is an excellent example of of software capable of bringing in this type of data and building the relationships that get the value from the data. This has become part of IBMs predictive analytics Along with SPSS
There are many others on the horizon, however, there are not enough. It is time that the great minds, such as those here, start looking at developing solutions for law enforcement. There is a market out there for the person who is interested, that is for sure, the current players have been into it for a while, but they are built on old technology and I believe many lack the in-house talent to out it all together. Most of these companies were started by police officers who liked "playing with computers". Of course there are great companies out there, but they are few and far between, Hope that helped shed some light on the subject.

I2 was purchased by IBM, their product i iBase purchased by IBM is an excellent example of of software capable of bringing in this type of data and building the relationships that get the value from the data. This has become part of IBMs predictive analytics Along with SPSS
There are many others on the horizon, however, there are not enough. It is time that the great minds, such as those here, start looking at developing solutions for law enforcement. There is a market out there for the person who is interested, that is for sure, the current players have been into it for a while, but they are built on old technology and I believe many lack the in-house talent to out it all together. Most of these companies were started by police officers who liked "playing with computers". Of course there are great companies out there, but they are few and far between, Hope that helped shed some light on the subject.

I know that some HFT platforms are written in Java, but I think this is caused by other reasons than trying to attain maximum performance. E.g. it is certainly harder to find programmers who have both high- and low-level skills (i.e. know hardware well and have good understanding of math, for instance).

As for "better than JIT" argument, I think you are putting too much trust into a rather generic approach. Even traditional (and performance oriented) compilers don't handle all use cases well and you can hit a roadblock there too, let alone all the fundamental problems with "managed" code (e.g. random memory access patterns which hurt CPU caches, various safety checks, stack-based VM design that doesn't map well to register-based hardware - stack-based processors are at inherent performance disadvantage, by the way, that's why Intel shunned FPU in favor of register-based SSE). My statement of JIT weakness is supported by the well-known fact that server-side software rarely has great single-thread performance (and often it doesn't need it, but it's another topic), so server CPUs tend to have gobs of cache in order to alleviate that.

Again... language (syntax, etc) doesn't matter much for me, I wouldn't mind Java if it allowed me to get as close to hardware as possible, even via non-portable extensions. It's the layers of code to profile and debug through and resulting feeling of not being in control is what I don't like.

It's a little more complicated.. where, say, 5 people hang out together for a week, where multiple bills are paid individually by different people over the duration of the vacation or whatever and you need to "settle up" when you get home.

It's a balance sheet, but on a computer!Good job no one has thought of doing that before.

> I predict the way you're using two digits to count the errors is going to turn into a scalability limit.

Not if the error sequence number follows the convention used in IBM RPG/400 1.1.4.4. "Sequence Numbering of the Listing after a Compile" ... "The high order 2 digits of the sequence number are made up of the characters A through Z and 0 through 9 in the following order: A, B, C, ..., Z, 1, 2, ..., 9, A0, AA, AB, ..., AZ, A1, A2, ..., A9, B0, BA, ..., ZZ, ..., Z9, 10, ..., 99. This structure allows for up to 1295 different increments of the high order sequence number. " ... it is worth noting that this counting sequence does not sort properly in ASCII or even native EBCDIC [A9,B0,BA] which leads Real Programmers away from the messy realms of real-world problems into the comfortable zone of devising elaborate workarounds for problems they had created.

Sometimes delving into the structure of ancient computer architectures and programming languages yields new and clever insights into old problems. This is not one of those times.

Booting an embedded PPC system directly to Zork.

If you're having performance issues, then C++ would offer a more efficient solution. Why jump through all these hoops to boost Java performance? Just use C++ and get twice the performance instantly with Linux.

That isn't true (in the general case) and it hasn't been for many, many years - can't tell if you're trolling or just ignorant. There are plenty of operations for which a JVM is actually faster than a C process (for example, Java new() is faster than malloc()), and Hotspot runtime optimization has access to a lot more information about how code is actually being used than static compile time optimization - the difference that makes can be remarkable.

Java isn't the right tool for every job, but neither is any language, and there are a great many applications out there for which converting to C++ (for example) would not give any kind of performance boost (and may even be slower).

$18 Billion in sales would seem to suggest otherwise.

oops, somehow the link got busted up. Try this:

http://www-935.ibm.com/services/us/gbs/consulting/

It's not migrating Exchange - Think MUCH bigger. It's things like this:

http://www-35.ibm.com/services/us/gbs/consulting/

No, such separation of access is not fundamentally impossible. Why should the 'admin' have write ability to ANY installed software? He shouldn't. He should only have the authority to run an installation program, and the installation program should have the authority to replace the files. Of course, the installation program should verify that the package he is installing is on a list (which he does not have write access to) and has been properly cryptographically signed (where again, he has access to neither the signing keys nor the trusted keys database).

Such systems exist (and have for decades). Here is one .

I don't believe this is legal, while you do have the right to terminate an employment relationship in a employment at will relationship, you're exposing yourself to a tort claim for misrepresenting the terms of employment. You could end up paying for the damages that result from unemployment. At the very least, your actions are unquestionably unethical and a jury would find contempt in your admission.

State clearly in writing what behavoir is acceptable and what is not in an employer handbook and have the employee sign something stating they have read it. This benefits you and your employees, and employee could have assumed they had flex time to make up for the time they spent on Facebook. All you've effectively done by not documenting these conditions is make yourself into an asshole.

Review IBM's Business Conduct Guidelines if you need examples of what to write in your employee handbook: https://www.ibm.com/investor/pdf/BCG2013.pdf

Answering my own question: nothing prevents silicon transistors from working over 400 GHz. IBM & GeorgiaTech have already done that.
http://gtresearchnews.gatech.edu/newsrelease/half-terahertz.htm
http://www-03.ibm.com/press/us/en/pressrelease/19843.wss
As it has been mentioned by user mc6809e in another comment, certain transistors have long since reached 1 THz, but I'm unqualified in the area and can't find the appropriate article or key words.

Keeping my excitement for some other occasion.

You can get the 'advance' toolchain. It's basically an upstream gnu toolchain, with a lot more optimizations and support for ppc chips.

Looks like LLVM is getting improved as well.

Educate yourself.

On the "Metro" thing the first thing that came to mind right now was the NetApp MetroCluster we just installed at work.
Now I find there is an IBM storage technology called MetroCluster and a NetApp storage technology called MetroCluster

I wonder how that will pan out.

Right... because google signs their own certs and you magically make your browser trust them tool.

Why don't you do some reading:

http://www-01.ibm.com/software/webservers/httpservers/doc/v2047/manual/ibm/en_US/9atssl.htm

and let intelligent people talk.

Really?
http://www.pipeline.com/~hbaker1/RealTimeGC.html
http://www.pipeline.com/~hbaker1/NoMotionGC.html
More recently:
http://researcher.watson.ibm.com/researcher/view_project_subpage.php?id=175

Benchmarks have been done and Java can often out-perform C

Cite? I've often seen that claim but don't recall many benchmarks. Maybe somebody found a ten line loop that a JIT optimizes the hell out of, but I'm very skeptical that Java can regularly outperform C.

Lies, Damn Lies, and Benchmarks, as they say. Actual mileage may vary. For short runs, the sheer overhead of setting up a JVM puts it out of the running. For server systems running more or less continuously, the startup time becomes negligible. Loop optimization is probably not one of the things that Java can optimize better than C. On the other hand, on a lot of systems, there's a significant difference in the amount of time that it takes a conditional branch instruction to take the branch versus going straight. A C compilation produces static code and if the worst-case optimization (branch taken) is selected, that's the end of the story. A JVM monitoring performance can detect this sort of thing and rewrite the code to minimize the number of times the branch is taken. Needles to say, a 10-iteration loop run once a day isn't going to benefit, but a routine that runs millions of times an hour will.

For some actual numbers, I asked my friend Google and managed to find some cases where people did a head-to-head with a greater or lesser degree of honesty (Lies, Damn Lies, ...).

Here's one:

http://keithlea.com/javabench/

A Wikipedia discussion, for contemplation by those who don't knee-jerk reject Wikipedia:

http://en.wikipedia.org/wiki/Java_performance

And something a bit more scholarly:

http://scribblethink.org/Computer/javaCbenchmark.html

And something from IBM on Java's memory-management performance:

http://www.ibm.com/developerworks/java/library/j-jtp09275/index.html

Something I read a while back and cannot recall now asserted that Java won most of their benchmarks except for those involving floating-point numbers. The reason being that C could use the processor's native FPU, but Java's "Write Once/Run Anyware" requirements forced it (unless overridden) to use IEEE Floating-point, which often had to be done in software.

IBM is a big Java proponent. The modern zSeries machines include 2 FPUs. One for the original, rather strange S/360 FP format and the other for IEEE. Not a co-incidence, I suspect. Reminds me of how a group of Honeywell engineers got together with an OS that had been written under government contract for NASA (thus available for public use), and designed a computer whose instruction set was optimized to run FORTRAN. They called they company Prime Computer.

No, not that IBM; the other IBM that still has a metric (not imperial) shitload[1] of assets.

[1] http://www.ibm.com/products

I don't think the boxes can talk infiniband at all.

Implementing and Managing InfiniBand Coupling Links on System z

You mean the IBM CPLEX tests (Part 2)?

"This said, best solutions CPLEX could find in 30 minutes are still worse than the best ones found using D-Wave black box or Tabu search in 29 problems, equal in 3, and better in only one problem. This is partly explained by the fact that CPLEX not only tries to find good feasible solutions, but it also spends a fair amount of time trying to prove optimality.
In short, we dramatically improved CPLEX results, but it does not really change the fact that heuristic methods (D-Wave blackbox and Tabu) are finding better quality solutions in a limited amount of time. QAP are definitely not CPLEX sweet spot, at least using the simple MIQP model above."

The next chip is coming out in 2014 and will have 2048 qubits. It's easier to find useful problems at that scale.

You mean the IBM CPLEX tests (Part 2)?

"This said, best solutions CPLEX could find in 30 minutes are still worse than the best ones found using D-Wave black box or Tabu search in 29 problems, equal in 3, and better in only one problem. This is partly explained by the fact that CPLEX not only tries to find good feasible solutions, but it also spends a fair amount of time trying to prove optimality.
In short, we dramatically improved CPLEX results, but it does not really change the fact that heuristic methods (D-Wave blackbox and Tabu) are finding better quality solutions in a limited amount of time. QAP are definitely not CPLEX sweet spot, at least using the simple MIQP model above."

The next chip is coming out in 2014 and will have 2048 qubits. It's easier to find useful problems at that scale.

While I'm on my soap box: I think it would be nice if we could have a programming environment that is completely sandboxed from every other program and every other user. In those cases where information must pass from one user to another on the same computer or from one computer to another, socket technologies with easy-to-use encryption seem like a good answer. And the runtime environment for that should be controlled by the installing user. It should be installed where I say with the sandboxed files going where I say. None of this hidden behind several curtains and modify-the-registry crap. No technology that I see today fulfills these needs.

Actually there is a technology that offers user / applications level sandboxing and permissions between applications are off by default:
http://www-03.ibm.com/systems/i/ and http://www-03.ibm.com/systems/z/

Or for that matter any OS with http://en.wikipedia.org/wiki/Capability-based_security

Interestingly enough NT started with this model and then turned it off because it was too complex. So NTs can do this. iOS (Apple's) is capability based.

From a 10000 foot view, it looks like Javascript is better because it is better sandboxed, but when you get down to how many technologies are involved to make a web browser work, it seems like there are going to be holes that are always open because no one can be an expert in all of those technologies.

It is much easier to secure bad technologies in a hardened environment than to secure good technologies in a soft environment. If just about all permissions default to "no" things are a lot harder to get screwed up.

While I'm on my soap box: I think it would be nice if we could have a programming environment that is completely sandboxed from every other program and every other user. In those cases where information must pass from one user to another on the same computer or from one computer to another, socket technologies with easy-to-use encryption seem like a good answer. And the runtime environment for that should be controlled by the installing user. It should be installed where I say with the sandboxed files going where I say. None of this hidden behind several curtains and modify-the-registry crap. No technology that I see today fulfills these needs.

Actually there is a technology that offers user / applications level sandboxing and permissions between applications are off by default:
http://www-03.ibm.com/systems/i/ and http://www-03.ibm.com/systems/z/

Or for that matter any OS with http://en.wikipedia.org/wiki/Capability-based_security

Interestingly enough NT started with this model and then turned it off because it was too complex. So NTs can do this. iOS (Apple's) is capability based.

From a 10000 foot view, it looks like Javascript is better because it is better sandboxed, but when you get down to how many technologies are involved to make a web browser work, it seems like there are going to be holes that are always open because no one can be an expert in all of those technologies.

It is much easier to secure bad technologies in a hardened environment than to secure good technologies in a soft environment. If just about all permissions default to "no" things are a lot harder to get screwed up.

Try grabbing the demo version of RSA, (Rational Software Architect), it's great for turning Java and C++ code into UML models and has a neat way to explore models through diagrams.

The trial version is free for 30 days.

I am sure there are open source tools out there for this, but nothing so complete.

The true cost of Nuclear power is more than any other method.

Talk about easy mode! "Any other method" logically includes coal. And coal sucks. To put it in perspective, about twice as much electricity is produced each year from coal(44.9%) as from nuclear power(20.3%) in the USA.

What, you want healthcare costs included along with the fatalities? Okay, sure thing. How does $500B/year sound, for the USA ALONE?

I'd say I hate to break it to you, but that would be dishonest. I LOVE breaking this to you: The world could suffer a Chernobyl level event EVERY year and it would STILL come out cheaper than coal.

And while we are at it, lets add in all of the cost for nuclear power plant accidents both public and private funds and divide that by the the number of operating plants. Let's see, Three Mile Island, Chernobyl, Fukushima, smaller costly but less publicized accidents.

Let's see: Chernobyl: $235B, TMI: $975M, Fukushima: too early to tell. Let's go with roughly between Chernobyl and TMI: $118B. It's probably quite high, but eh. Total: $354B, or about 3/5ths the damage coal does to the USA alone each year.

As I've said before, Chernobyl's design wouldn't have been allowed anywhere, the cost would have been far less if it had been built with a containment dome. 437 reactors, leaving the share per nuclear plant at $810M per your stupid standard.

Let's put it into better context: End of 2012 nuclear power had produced 69,760 billion kwh. Chernobyl, TMI, and Fukushima amount to .5 cents of cost per kwh. Yes, half a cent.

I clearly hadn't read more than the first few lines of the help on Java TimeZone info or I could have found out that the answer was already there, without having to wait for it. As another poster pointed out IBM already provides free Java timezone updates.

Let me google that for you! But more to the point, writing a tool that will grab those updates for yourself and storing it where you need it looks like a bash script or batch file candidate. Our brains are more than a match for Oracle's bean counters. Let's use them!

cheers...ank

Obviously it should be IaaS, but IBM in space would not be unheard of.

By the late 1960s, the US panic over Russia's launch of Sputnik had resulted in a significant increase in investment in science education at all levels. In particular, an experimental program was started at the public Bronx High School of Science in New York City, with the goal of exploring whether high school age students could successfully program computers. (Computers often cost hundreds of thousands or millions of dollars, so this felt a bit like seeing whether students could fly a jetlineer). The school procured a number of small systems including several of the legendary Olivetti Programma 101s.

The school's main machine, though, was an IBM 1620. It was a decimal (not binary) machine, and the high school's was had the minimum 20K digits (not bytes!) of memory. Originally, input and output was only via punch cards and the built in typewriter, but by 1968 or 1969 an IBM 1311 disk drive, which was itself the size of a small washing machine, added 2 Million digits of persistent storage, and a 1443 lineprinter was also added.

Students were taught, in this order: machine language, then FORTRAN II, then assembler. That's not a typo. Since the machine was decimal, it wasn't too hard to type in raw machine code onto cards, which could be loaded directly into memory by a short self-booting loader program, which was included on the front of each card deck. There was an assember, but until the hard disk showed up the assembler required that an intermediate deck be punched and loaded each time an assembly was attempted. So, for moderate size programs, it was often easier to write machine code directly. This typically involved manually computing and setting the absolute target of each branch. Here's a picture of the cover of the textbook that students used to learn 1620 programming.

This experiment was, in my opinion, wildly successful. I'm aware of at least 4 quite well known computer scientists who started their programming careers in the 1960s on that machine at Bronx Science. In general, many programmers from that era learned on 1620s. Like PDP-11s and Apple-2's later, 1620s were machines that an individual could get easy access to, and could learn to program at both a low level (machine or assember) and higher level (FORTRAN, LISP). Famously, the 1620 did not have a general purpose adder implemented in hardware: the add instructions would not produce the usual results until software loaded a table of partial sums into a fixed location in memory. The machine's nickname was thus CADET: Can't Add, Doesn't Even Try.

By the late 1960s, the US panic over Russia's launch of Sputnik had resulted in a significant increase in investment in science education at all levels. In particular, an experimental program was started at the public Bronx High School of Science in New York City, with the goal of exploring whether high school age students could successfully program computers. (Computers often cost hundreds of thousands or millions of dollars, so this felt a bit like seeing whether students could fly a jetlineer). The school procured a number of small systems including several of the legendary Olivetti Programma 101s.

The school's main machine, though, was an IBM 1620. It was a decimal (not binary) machine, and the high school's was had the minimum 20K digits (not bytes!) of memory. Originally, input and output was only via punch cards and the built in typewriter, but by 1968 or 1969 an IBM 1311 disk drive, which was itself the size of a small washing machine, added 2 Million digits of persistent storage, and a 1443 lineprinter was also added.

Students were taught, in this order: machine language, then FORTRAN II, then assembler. That's not a typo. Since the machine was decimal, it wasn't too hard to type in raw machine code onto cards, which could be loaded directly into memory by a short self-booting loader program, which was included on the front of each card deck. There was an assember, but until the hard disk showed up the assembler required that an intermediate deck be punched and loaded each time an assembly was attempted. So, for moderate size programs, it was often easier to write machine code directly. This typically involved manually computing and setting the absolute target of each branch. Here's a picture of the cover of the textbook that students used to learn 1620 programming.

This experiment was, in my opinion, wildly successful. I'm aware of at least 4 quite well known computer scientists who started their programming careers in the 1960s on that machine at Bronx Science. In general, many programmers from that era learned on 1620s. Like PDP-11s and Apple-2's later, 1620s were machines that an individual could get easy access to, and could learn to program at both a low level (machine or assember) and higher level (FORTRAN, LISP). Famously, the 1620 did not have a general purpose adder implemented in hardware: the add instructions would not produce the usual results until software loaded a table of partial sums into a fixed location in memory. The machine's nickname was thus CADET: Can't Add, Doesn't Even Try.

Quietly shelved? I don't think so.

You get sequencing too...
http://www-03.ibm.com/ibm/history/ibm100/us/en/icons/dnatransistor/

I believe those are actually a visualization of the atoms' electrons moving across the copper surface... you can see constructive and destructive destruction of the waves around the boy. If you look at this stm image ( http://researcher.watson.ibm.com/researcher/files/us-flinte/stm16.jpg ), the bottom right image shows the wave function of electrons completely trapped inside the circle of atoms.

Remember, this is scanning tunneling microscopy, so the electrons are not actually going in and out of the plane... what we're seeing is their potential to tunnel into the tip of the microscope.

http://www-03.ibm.com/press/us/en/pressrelease/40970.wss