I would like to see your solution for these equations for normal human behaviour. For some reason, I think you cannot even formulate the objective, let alone solve it. This model might be suboptimal, but that doesn't make it worse than random guessing.
How exactly did this get modded to 5? This is taking the computer metaphor way, way, waaaaaaaay too far. Human memory in no way resembles computer memory, as has been shown over and over again. You've clearly got no idea. First open a freshmen grade book on cognitive psychology, then think again.
I think you'll find that a few general mechanisms, such as chunking and association, have been described a long time ago and explain much better what you observed. Much research has been done since. And this particular bit of research is about one of the molecular mechanisms that subserves our memory.
OSX isn't half as counter-intuitive as Linux can be (e.g. network administration is really easy under OSX; nothing in it reminds you of the actual/etc/whatever file structure, unlike some Linux tools I could mention), according to me, but tastes differ, ok, I get that.
What I don't get is the rest of your complaints. - Programs such as VLC and NicePlayer play full-screen, and they're free. And, coincidentally, QT 7.2 does allow fullscreen. - Under Linux you can also get demo apps. That's because they're commercial programs. You can download nearly any Linux OSS app and run it under OSX with X11. And even the app you mention you prefer under Ubuntu is proprietary! - Spaces isn't revolutionary at all, you're right there.
If you want to give OSX another chance, just have your account settings open the Terminal.app OR X11 from at login, and you can be very happy. For the rest, check out DarwinPorts: a lot of programs have been ported to OSX and installing them is pretty easy (easier than rpm IMHO, although that's RedHat, I suppose).
Well, GNUstep does not really implement all of Cocoa. The datamodel and bindings (KVO, KVC) spring to mind (the Wiki says they would like to implement it, but haven't even thought about when to start). A lot of modern Cocoa apps (bindings are available since 10.3, data-models since 10.4) use that, so Cocoa is really a lot more than Openstep v2. I don't think compatibility is around the corner...
Holy Lord. You call Photoshop a niche application? A bloody rich niche it is then. Whole hordes have weekly photoshop competitions, and they run it on their own computers. And Linux also lacks a decent Office suite (yes, I've got OpenOffice and AbiWord running on all platforms and Word, as sucky as it may be, still beats them).
The Open Source model just doesn't suit mass markets: mass markets generate mass money, which is a big stimulus for companies. Thus Desktop Linux will always stay in its own niche where userfriendliness is not that important and you can have flame wars over KDE vs. Gnome (vs. twm vs. fvwm vs....) instead of combining efforts on a unified GUI. Linux could succeed on servers because they're operated by geeks anyway.
Until total computer literacy breaks out, Linux is stuck with geeks...
1. "most things humans can do to improve compression algorithms are already being done in the best algorithms". Yes, that's true and that's just because the structure of words is so predictable. E.g, if you see a word starting with "struct", the next two characters are "ur" (assuming no typing errors etc). So, in that ideal case, these two characters can be represented by 0 bits. The same holds for many vowels. However, people don't realize this by themselves, and they don't always know which vowels you can leave out and which ones you can't. So I would say our intelligence in compressing is lower than that of even such a simple scheme as Huffman.
2. "Also, if your text can't be unambiguously recovered, then you can't claim to have losslessly compressed it." That by itself is a whole world of discussion, since the goal *according to me* is compression where you don't lose knowledge. However, losing irrelevant text is not a problem. Otherwise, why is there any understanding needed? So, you have to squeeze out every bit until you can just unambiguously recover the knowledge, but that has a very, very soft lower limit. So that (according to me) makes the contest rather weak.
3. "The claim is "we are one step closer to having a machine that passes the Turing Test,...". My counter-claim was that we've since long passed that point. If you ask a normal person on the other side of a TTY to give a good compression (plus decompression algorithm) of a certain sentence, I think they would utterly fail.
4. A statistical compressor can choose the correct form of "gd" if necessary (note that when it has to compress lossless, it cannot and should not choose, so wouldn't need that bit of intelligence). If you're interested in that kind of thing: there are spell checkers that take a (probabilistic) distribution of the context into account and can come up with quite good corrections. I don't think any of them has so far made it into a commercial product, but a typical example from a PhD work on this topic is the correction of "onjections" in "most vehement onjections lodged against" or in "painful intramuscular onjections received daily" to "objections" in the first case and "injections" in the second.
First point: f ts s s t pt th vwls bck n, gd sttstcl cmprssr wll s vr lttl spc fr thm.
Second point: even a human cannot unambiguously recover the previous sentence (hint: singular or plural).
Third point: proxies are not very interesting. Language itself is interesting, but language is not intelligence either: it's just the communication protocol between humans (which I happen to have been researching for the past 20 years).
Intelligence has many other appearances: spatial, mathematical, verbal, social, etc. I don't think these are all really separate forms of intelligence (as many people seem to claim), but neither is intelligence as one dimensional as IQ suggests, and therefore quite difficult to measure...
OK, I should have been clearer: the Hutter prize people don't assume compression is AI because of the incorrect reasoning I gave here. That was an attempt at humorous insult.
However, the link between compression and intelligence is total nonsense. If I'm going to ask a couple of randomly selected, intelligent people to compress Wikipedia, they won't reach anywhere near 1 bit per character. Perhaps if I get together a whole team of expert researchers and compression experts, they will reach this "magical" threshold, but a normal individual won't, and that's still the standard for intelligence.
And Shannon's number comes from the redundancy in our language and its written representation and has nothing to do with intelligence. Its value also differs per language, at least as far as statistical measures are concerned. English is relatively well structured, but free word order languages have a considerably higher entropy when you just look at character sequences, and that's what Shannon did. Instead, we should do as I outlined above: ask normally intelligent people to compress a large body of text and only then we can compute what the threshold for AI is. And, as you can read above, I think we long surpassed this threshold. Even gzip will do better than most humans.
So, the assumption behind the Hutter prize is bollocks. Satisfied?
It is not equivalent, so I'm not surprised you didn't get it. As far as I know, the reasoning goes as follows: Shannon estimated that each character contains 1.something bits of information. Shannon was an intelligent human being. So if a program reaches this limit, it is as smart as Shannon.
And yes, that's absolute bollocks. Shannon's number was just an estimate and only applied to serial transmission of characters, because that's what he was interested in. Since then, a lot of work has been done in statistical natural language processing, and I would be surprised if the number couldn't be lowered.
Anyway, since the program doesn't learn or think to reach this limit, nor gives a explanation how this level of compression is intrinsically linked to the language/knowledge it compresses, it cannot be called AI; e.g., it doesn't know how to skip irrelevant bits of information in the text. That would be intelligence...
Or you'd need the aliens to be directing this, handing out tiny little tidbits of information to the scientists, and either swearing them to secrecy or using some sort of mind control on them. OH MY GOD! You have found out our secret!!! And have made it public. You know what will happen to you know, don't you?
If you know what it's like to be on hold and have some empathy, this story (http://macslash.org/) will make your toes curl. The guy got an iPhone at the Apple Store and tried to get it activated by AT&T by telephone. Bad idea...
Actually, I hold a post-doc in language processing with neural networks, so I know a bit about the topic, and we're not even going to mimic a mouse brain anywhere in the coming decades. A lot of the brain's functionality depends on specific physical characteristics that we don't understand at this moment. E.g., a group of my colleagues claims that synchrony plays an important part in communication between groups of neurons. They have some evidence (in Science this year), but it is not conclusive, so we still don't know. And we know even less about its functional computational role. And there are other gray areas of our understanding of the brain, so any artificial model is going to be far from a good copy for a long time to come.
I guess by "man-made" they mean artificial and that it will REVOLUTIONIZE(tm) computing since these artificial brains are going to be built in to every PC. Where did I hear that before? I think at the time they grossly overstated the capacity of computers such as the original IBM PC. So perhaps Moore's law applies to hardware, it surely doesn't apply to exaggeration.
Anyway, who needs an electronic brain? Now I can at least yell "idiot" to MS Word when it joins sections or splits pages without it getting offended. Can you imagine Clippy looking angry and saying in this cute cartoon like blob "Now I'm not going to erase your document, you asked for it".
Oh, come off it! For every company that drops on the ranking, there is another one that goes up, so while that specific company might earn less, another one will earn more. Complete economic fairness. And while I'm not a great fan of the ever expanding Google, it is good to remember that they are there to serve *us*, the content searching end users, not product offering companies.
No, the parent is completely right. I too worked on code that had quite a few ifdefs to maintain compatibility across the different versions of Windows. Actually, since it ran as a DLL in Office, it also had ifdefs for different versions of Office.
Games are different beasts: quite a few of them only read the keyboard and the mouse events and show an image on screen. They interact very little with the OS. And they usually break down in the parts that are not play-related, such as starting up, self-updating, etc.
Did any read it? You might better ask: did anyone understand it? Because that doesn't seem to be necessary to get +5 insightful for you comments.
About the interference from previous experiments within the same animal: it depends on what you want to investigate. Of course there is neural adaptation, but in most regions it is quite slow. So after one experiment you might let the animal run around under normal conditions and it'll be as ready as ever, if only for another experiment.
Plus, when you're able to suppress unrelated activity, you might get a much cleaner signal and thus need less animals.
Just for comfort: I hope your karma will be fine. I find the GNU prefix a load of gnu sh*t as well. It's just ego talking. And you're right about the GPL, at least before someone invented a clause that said you had to accept the clauses of future GPL licenses.
You know what the worst thing is that can happen? That someone mods you up a few points and then down and then up again, etc. Then you'll be banned from posting for some time. It's quite ridiculous, but with so many zealots around, it just might happen. I've been there for saying something innocent but apparently really upsetting to some...
If he said that, he's very wrong. The cortex consists of dozens areas with different cyto-architectonic (that means cellular structural) properties, see http://spot.colorado.edu/~dubin/talks/brodmann/bro dmann.html for a nice map. Brodmann counted 46 of them and modern views distinguish sub-areas in most of them. E.g., BA44 (Brodmann's Area 44) is considered to be involved in language processing (amongst other things), but is usually divided into 44a, b and c (there are different ways of naming these, too; e.g. the pars opercularis or Broca's Area for BA44).
So, the cortex contains many different areas with different physical properties and these are commonly tied to specific functions: e.g. language processing involves a few areas, and motoric processing involves a few different areas, and these never overlap. Consequently, any model that wants to approach the cortex at neuronal level should account for this.
And the cortex of a whale is much larger than that of humans...
He'll just repeat that. These actions were voluntarily. All the study shows is that bits in our brain have a different activation pattern when you try to hold on to your decision to add or your decision to subtract.
Furthermore, nobody can deny that subconcious parts of our brain will have a bias to some choice before that choice has to be made. But that does not imply that we don't have a free will. All it says is that we identify our conciousness one-to-one with our mental processes instead of accepting that it is more a reflective system by which (amongst other things) we judge our intentions and actions.
This has been reported before on SlashDot. I work in the field, I've read the article, and I'll say again: the claims are not that strong. There are bits in the brain that are more active when someone has taken the decision to add numbers that are going to appear in a few seconds and there are bits that are more active when when that same person has taken the decision to subtract. There are also bits that show a different activation pattern for adding and subtracting while performing the actual operation. They did get a correct prediction measure by training on a subset of the data and measuring actual performance on the part left out. That it wasn't done in real-time is not that important. And the success rate was 71%, whereas simple guessing would give you 50%.
So, can we read intentions? In practice: no. fMRI is way too slow and a lot of material is needed before a "prediction" can be made. And MRI scanners are not really portable. In theory: neither, since the success rate is rather low and the choice is binary (which real-life intentions are certainly not), artificial (make a "random" decision and maintain it for a couple of seconds) and forced (normal decisions arise spontaneously from events around you). Other experiments with "brain reading" show much more interesting results, but could not predict either. They might simply have measured people holding on to their intention, response to conditioning or even trying to be as random as possible.
Well, it's from Nature 406 (6793): 277-279 Jul 20 2000. The article is cited 315 times and seems dispted. Here is the abstract for those poor souls without access to Nature, Web of Science, Scopus, etc.:
Einstein's theory of special relativity and the principle of causality(1-4) imply that the speed of any moving object cannot exceed that of light in a vacuum (c). Nevertheless, there exist various proposals(5-18) for observing faster-than-c propagation of light pulses, using anomalous dispersion near an absorption line(4,6-8), nonlinear(9) and linear gain lines(10-18), or tunnelling barriers(19). However, in all previous experimental demonstrations, the light pulses experienced either very large absorption(7) or severe reshaping(9,19), resulting in controversies over the interpretation. Here we use gain-assisted linear anomalous dispersion to demonstrate superluminal light propagation in atomic caesium gas. The group velocity of a laser pulse in this region exceeds c and can even become negative(16,17), while the shape of the pulse is preserved. We measure a group-velocity index of n(g) = -310(+/-5); in practice, this means that a light pulse propagating through the atomic vapour cell appears at the exit side so much earlier than if it had propagated the same distance in a vacuum that the peak of the pulse appears to leave the cell before entering it. The observed superluminal light pulse propagation is not at odds with causality, being a direct consequence of classical interference between its different frequency components in an anomalous dispersion region.
Slashdot Mirror
I would like to see your solution for these equations for normal human behaviour. For some reason, I think you cannot even formulate the objective, let alone solve it. This model might be suboptimal, but that doesn't make it worse than random guessing.
How exactly did this get modded to 5? This is taking the computer metaphor way, way, waaaaaaaay too far. Human memory in no way resembles computer memory, as has been shown over and over again. You've clearly got no idea. First open a freshmen grade book on cognitive psychology, then think again.
I think you'll find that a few general mechanisms, such as chunking and association, have been described a long time ago and explain much better what you observed. Much research has been done since. And this particular bit of research is about one of the molecular mechanisms that subserves our memory.
OSX isn't half as counter-intuitive as Linux can be (e.g. network administration is really easy under OSX; nothing in it reminds you of the actual /etc/whatever file structure, unlike some Linux tools I could mention), according to me, but tastes differ, ok, I get that.
What I don't get is the rest of your complaints.
- Programs such as VLC and NicePlayer play full-screen, and they're free. And, coincidentally, QT 7.2 does allow fullscreen.
- Under Linux you can also get demo apps. That's because they're commercial programs. You can download nearly any Linux OSS app and run it under OSX with X11. And even the app you mention you prefer under Ubuntu is proprietary!
- Spaces isn't revolutionary at all, you're right there.
If you want to give OSX another chance, just have your account settings open the Terminal.app OR X11 from at login, and you can be very happy. For the rest, check out DarwinPorts: a lot of programs have been ported to OSX and installing them is pretty easy (easier than rpm IMHO, although that's RedHat, I suppose).
Well, GNUstep does not really implement all of Cocoa. The datamodel and bindings (KVO, KVC) spring to mind (the Wiki says they would like to implement it, but haven't even thought about when to start). A lot of modern Cocoa apps (bindings are available since 10.3, data-models since 10.4) use that, so Cocoa is really a lot more than Openstep v2. I don't think compatibility is around the corner...
Amazing how that could be modded insightful...
Holy Lord. You call Photoshop a niche application? A bloody rich niche it is then. Whole hordes have weekly photoshop competitions, and they run it on their own computers. And Linux also lacks a decent Office suite (yes, I've got OpenOffice and AbiWord running on all platforms and Word, as sucky as it may be, still beats them).
...) instead of combining efforts on a unified GUI. Linux could succeed on servers because they're operated by geeks anyway.
The Open Source model just doesn't suit mass markets: mass markets generate mass money, which is a big stimulus for companies. Thus Desktop Linux will always stay in its own niche where userfriendliness is not that important and you can have flame wars over KDE vs. Gnome (vs. twm vs. fvwm vs.
Until total computer literacy breaks out, Linux is stuck with geeks...
Umm, let's see.
...". My counter-claim was that we've since long passed that point. If you ask a normal person on the other side of a TTY to give a good compression (plus decompression algorithm) of a certain sentence, I think they would utterly fail.
1. "most things humans can do to improve compression algorithms are already being done in the best algorithms". Yes, that's true and that's just because the structure of words is so predictable. E.g, if you see a word starting with "struct", the next two characters are "ur" (assuming no typing errors etc). So, in that ideal case, these two characters can be represented by 0 bits. The same holds for many vowels. However, people don't realize this by themselves, and they don't always know which vowels you can leave out and which ones you can't. So I would say our intelligence in compressing is lower than that of even such a simple scheme as Huffman.
2. "Also, if your text can't be unambiguously recovered, then you can't claim to have losslessly compressed it." That by itself is a whole world of discussion, since the goal *according to me* is compression where you don't lose knowledge. However, losing irrelevant text is not a problem. Otherwise, why is there any understanding needed? So, you have to squeeze out every bit until you can just unambiguously recover the knowledge, but that has a very, very soft lower limit. So that (according to me) makes the contest rather weak.
3. "The claim is "we are one step closer to having a machine that passes the Turing Test,
4. A statistical compressor can choose the correct form of "gd" if necessary (note that when it has to compress lossless, it cannot and should not choose, so wouldn't need that bit of intelligence). If you're interested in that kind of thing: there are spell checkers that take a (probabilistic) distribution of the context into account and can come up with quite good corrections. I don't think any of them has so far made it into a commercial product, but a typical example from a PhD work on this topic is the correction of "onjections" in "most vehement onjections lodged against" or in "painful intramuscular onjections received daily" to "objections" in the first case and "injections" in the second.
First point: f ts s s t pt th vwls bck n, gd sttstcl cmprssr wll s vr lttl spc fr thm.
Second point: even a human cannot unambiguously recover the previous sentence (hint: singular or plural).
Third point: proxies are not very interesting. Language itself is interesting, but language is not intelligence either: it's just the communication protocol between humans (which I happen to have been researching for the past 20 years).
Intelligence has many other appearances: spatial, mathematical, verbal, social, etc. I don't think these are all really separate forms of intelligence (as many people seem to claim), but neither is intelligence as one dimensional as IQ suggests, and therefore quite difficult to measure...
OK, I should have been clearer: the Hutter prize people don't assume compression is AI because of the incorrect reasoning I gave here. That was an attempt at humorous insult.
However, the link between compression and intelligence is total nonsense. If I'm going to ask a couple of randomly selected, intelligent people to compress Wikipedia, they won't reach anywhere near 1 bit per character. Perhaps if I get together a whole team of expert researchers and compression experts, they will reach this "magical" threshold, but a normal individual won't, and that's still the standard for intelligence.
And Shannon's number comes from the redundancy in our language and its written representation and has nothing to do with intelligence. Its value also differs per language, at least as far as statistical measures are concerned. English is relatively well structured, but free word order languages have a considerably higher entropy when you just look at character sequences, and that's what Shannon did. Instead, we should do as I outlined above: ask normally intelligent people to compress a large body of text and only then we can compute what the threshold for AI is. And, as you can read above, I think we long surpassed this threshold. Even gzip will do better than most humans.
So, the assumption behind the Hutter prize is bollocks. Satisfied?
It is not equivalent, so I'm not surprised you didn't get it. As far as I know, the reasoning goes as follows: Shannon estimated that each character contains 1.something bits of information. Shannon was an intelligent human being. So if a program reaches this limit, it is as smart as Shannon.
And yes, that's absolute bollocks. Shannon's number was just an estimate and only applied to serial transmission of characters, because that's what he was interested in. Since then, a lot of work has been done in statistical natural language processing, and I would be surprised if the number couldn't be lowered.
Anyway, since the program doesn't learn or think to reach this limit, nor gives a explanation how this level of compression is intrinsically linked to the language/knowledge it compresses, it cannot be called AI; e.g., it doesn't know how to skip irrelevant bits of information in the text. That would be intelligence...
Does that mean the FSF will start sueing the Beatles as well?
If you know what it's like to be on hold and have some empathy, this story (http://macslash.org/) will make your toes curl. The guy got an iPhone at the Apple Store and tried to get it activated by AT&T by telephone. Bad idea...
Actually, I hold a post-doc in language processing with neural networks, so I know a bit about the topic, and we're not even going to mimic a mouse brain anywhere in the coming decades. A lot of the brain's functionality depends on specific physical characteristics that we don't understand at this moment. E.g., a group of my colleagues claims that synchrony plays an important part in communication between groups of neurons. They have some evidence (in Science this year), but it is not conclusive, so we still don't know. And we know even less about its functional computational role. And there are other gray areas of our understanding of the brain, so any artificial model is going to be far from a good copy for a long time to come.
I guess by "man-made" they mean artificial and that it will REVOLUTIONIZE(tm) computing since these artificial brains are going to be built in to every PC. Where did I hear that before? I think at the time they grossly overstated the capacity of computers such as the original IBM PC. So perhaps Moore's law applies to hardware, it surely doesn't apply to exaggeration.
Anyway, who needs an electronic brain? Now I can at least yell "idiot" to MS Word when it joins sections or splits pages without it getting offended. Can you imagine Clippy looking angry and saying in this cute cartoon like blob "Now I'm not going to erase your document, you asked for it".
Oh, come off it! For every company that drops on the ranking, there is another one that goes up, so while that specific company might earn less, another one will earn more. Complete economic fairness. And while I'm not a great fan of the ever expanding Google, it is good to remember that they are there to serve *us*, the content searching end users, not product offering companies.
No, the parent is completely right. I too worked on code that had quite a few ifdefs to maintain compatibility across the different versions of Windows. Actually, since it ran as a DLL in Office, it also had ifdefs for different versions of Office.
Games are different beasts: quite a few of them only read the keyboard and the mouse events and show an image on screen. They interact very little with the OS. And they usually break down in the parts that are not play-related, such as starting up, self-updating, etc.
Did any read it? You might better ask: did anyone understand it? Because that doesn't seem to be necessary to get +5 insightful for you comments.
About the interference from previous experiments within the same animal: it depends on what you want to investigate. Of course there is neural adaptation, but in most regions it is quite slow. So after one experiment you might let the animal run around under normal conditions and it'll be as ready as ever, if only for another experiment.
Plus, when you're able to suppress unrelated activity, you might get a much cleaner signal and thus need less animals.
He's implicity REAL.
That's truly funny. You've got to be a nerd to get it, but funny.
Just for comfort: I hope your karma will be fine. I find the GNU prefix a load of gnu sh*t as well. It's just ego talking. And you're right about the GPL, at least before someone invented a clause that said you had to accept the clauses of future GPL licenses.
You know what the worst thing is that can happen? That someone mods you up a few points and then down and then up again, etc. Then you'll be banned from posting for some time. It's quite ridiculous, but with so many zealots around, it just might happen. I've been there for saying something innocent but apparently really upsetting to some...
Actually, dolphins were found to be rather dumb in a recent study. Search for "Manger" (the author) and "dolphins"...
If he said that, he's very wrong. The cortex consists of dozens areas with different cyto-architectonic (that means cellular structural) properties, see http://spot.colorado.edu/~dubin/talks/brodmann/bro dmann.html for a nice map. Brodmann counted 46 of them and modern views distinguish sub-areas in most of them. E.g., BA44 (Brodmann's Area 44) is considered to be involved in language processing (amongst other things), but is usually divided into 44a, b and c (there are different ways of naming these, too; e.g. the pars opercularis or Broca's Area for BA44).
So, the cortex contains many different areas with different physical properties and these are commonly tied to specific functions: e.g. language processing involves a few areas, and motoric processing involves a few different areas, and these never overlap. Consequently, any model that wants to approach the cortex at neuronal level should account for this.
And the cortex of a whale is much larger than that of humans...
He'll just repeat that. These actions were voluntarily. All the study shows is that bits in our brain have a different activation pattern when you try to hold on to your decision to add or your decision to subtract.
Furthermore, nobody can deny that subconcious parts of our brain will have a bias to some choice before that choice has to be made. But that does not imply that we don't have a free will. All it says is that we identify our conciousness one-to-one with our mental processes instead of accepting that it is more a reflective system by which (amongst other things) we judge our intentions and actions.
This has been reported before on SlashDot. I work in the field, I've read the article, and I'll say again: the claims are not that strong. There are bits in the brain that are more active when someone has taken the decision to add numbers that are going to appear in a few seconds and there are bits that are more active when when that same person has taken the decision to subtract. There are also bits that show a different activation pattern for adding and subtracting while performing the actual operation. They did get a correct prediction measure by training on a subset of the data and measuring actual performance on the part left out. That it wasn't done in real-time is not that important. And the success rate was 71%, whereas simple guessing would give you 50%.
So, can we read intentions? In practice: no. fMRI is way too slow and a lot of material is needed before a "prediction" can be made. And MRI scanners are not really portable. In theory: neither, since the success rate is rather low and the choice is binary (which real-life intentions are certainly not), artificial (make a "random" decision and maintain it for a couple of seconds) and forced (normal decisions arise spontaneously from events around you). Other experiments with "brain reading" show much more interesting results, but could not predict either. They might simply have measured people holding on to their intention, response to conditioning or even trying to be as random as possible.
For another, more understandable report, here is a BBC website: http://www.whyevolution.com/einstein.html (search for Wang).