Shirky on Internet Groups July 2, 2003 5:16 AM Subscribe
Shirky on Internet Groups: All groups of any integrity have a constitution. The constitution is always partly formal and partly informal. At the very least, the formal part is what's substantiated in code -- "the software works this way."
Article about groups (including Metafilter) on the internet. Quite Interesting.
Article about groups (including Metafilter) on the internet. Quite Interesting.
I just want to know...
Why is the font so big on the linked speech?
posted by jpburns at 7:10 AM on July 2, 2003
Why is the font so big on the linked speech?
posted by jpburns at 7:10 AM on July 2, 2003
4.) And, finally, you have to find a way to spare the group from scale. Scale alone kills conversations, because conversations require dense two-way conversations. In conversational contexts, Metcalfe's law is a drag. The fact that the amount of two-way connections you have to support goes up with the square of the users means that the density of conversation falls off very fast as the system scales even a little bit. You have to have some way to let users hang onto the less is more pattern, in order to keep associated with one another.
This is an inverse value to scale question. Think about your Rolodex. A thousand contacts, maybe 150 people you can call friends, 30 people you can call close friends, two or three people you'd donate a kidney to. The value is inverse to the size of the group. And you have to find some way to protect the group within the context of those effects.
Sometimes you can do soft forking. Live Journal does the best soft forking of any software I've ever seen, where the concepts of "you" and "your group" are pretty much intertwingled. The average size of a Live Journal group is about a dozen people. And the median size is around five.
But each user is a little bit connected to other such clusters, through their friends, and so while the clusters are real, they're not completely bounded -- there's a soft overlap which means that though most users participate in small groups, most of the half-million LiveJournal users are connected to one another through some short chain.
IRC channels and mailing lists are self-moderating with scale, because as the signal to noise ratio gets worse, people start to drop off, until it gets better, so people join, and so it gets worse. You get these sort of oscillating patterns. But it's self-correcting.
And then my favorite pattern is from MetaFilter, which is: When we start seeing effects of scale, we shut off the new user page. "Someone mentions us in the press and how great we are? Bye!" That's a way of raising the bar, that's creating a threshold of participation. And anyone who bookmarks that page and says "You know, I really want to be in there; maybe I'll go back later," that's the kind of user MeFi wants to have.
You have to find some way to protect your own users from scale. This doesn't mean the scale of the whole system can't grow. But you can't try to make the system large by taking individual conversations and blowing them up like a balloon; human interaction, many to many interaction, doesn't blow up like a balloon. It either dissipates, or turns into broadcast, or collapses. So plan for dealing with scale in advance, because it's going to happen anyway.
Until I got to there--man, this scrolls s-l-o-w-ly on dialup--I thought, No, not Kay-Cee Nicole again... But this discussion on scale is to the current point.
After all, The core group has rights that trump individual rights in some situations. Ah, but where do we draw the core group line? Are the pre-9/11 the anointed? Or shall it be the pre-9622 Wacky Fun Site schism members? Inquiring minds will want to know... ♥ your old timer status. Such as it is.
posted by y2karl at 9:13 AM on July 2, 2003
This is an inverse value to scale question. Think about your Rolodex. A thousand contacts, maybe 150 people you can call friends, 30 people you can call close friends, two or three people you'd donate a kidney to. The value is inverse to the size of the group. And you have to find some way to protect the group within the context of those effects.
Sometimes you can do soft forking. Live Journal does the best soft forking of any software I've ever seen, where the concepts of "you" and "your group" are pretty much intertwingled. The average size of a Live Journal group is about a dozen people. And the median size is around five.
But each user is a little bit connected to other such clusters, through their friends, and so while the clusters are real, they're not completely bounded -- there's a soft overlap which means that though most users participate in small groups, most of the half-million LiveJournal users are connected to one another through some short chain.
IRC channels and mailing lists are self-moderating with scale, because as the signal to noise ratio gets worse, people start to drop off, until it gets better, so people join, and so it gets worse. You get these sort of oscillating patterns. But it's self-correcting.
And then my favorite pattern is from MetaFilter, which is: When we start seeing effects of scale, we shut off the new user page. "Someone mentions us in the press and how great we are? Bye!" That's a way of raising the bar, that's creating a threshold of participation. And anyone who bookmarks that page and says "You know, I really want to be in there; maybe I'll go back later," that's the kind of user MeFi wants to have.
You have to find some way to protect your own users from scale. This doesn't mean the scale of the whole system can't grow. But you can't try to make the system large by taking individual conversations and blowing them up like a balloon; human interaction, many to many interaction, doesn't blow up like a balloon. It either dissipates, or turns into broadcast, or collapses. So plan for dealing with scale in advance, because it's going to happen anyway.
Until I got to there--man, this scrolls s-l-o-w-ly on dialup--I thought, No, not Kay-Cee Nicole again... But this discussion on scale is to the current point.
After all, The core group has rights that trump individual rights in some situations. Ah, but where do we draw the core group line? Are the pre-9/11 the anointed? Or shall it be the pre-9622 Wacky Fun Site schism members? Inquiring minds will want to know... ♥ your old timer status. Such as it is.
posted by y2karl at 9:13 AM on July 2, 2003
My old timer status is all in the real world, I'm afraid.
posted by timeistight at 10:41 AM on July 2, 2003
posted by timeistight at 10:41 AM on July 2, 2003
A fascinating read. I have to admit, seanyboy, your last MeTa post made me think you were a jerk. Now I think you just need to control your jerkoid tendencies better.
This bit, for instance, is extremely thought-provoking:
In the early Nineties, a proposal went out to create a Usenet news group for discussing Tibetan culture, called soc.culture.tibet. And it was voted down, in large part because a number of Chinese students who had Internet access voted it down, on the logic that Tibet wasn't a country; it was a region of China. And in their view, since Tibet wasn't a country, there oughtn't be any place to discuss its culture, because that was oxymoronic.
Now, everyone could see that this was the wrong answer. The people who wanted a place to discuss Tibetan culture should have it. That was the core group. But because the one person/one vote model on Usenet said "Anyone who's on Usenet gets to vote on any group," sufficiently contentious groups could simply be voted away.
Lots of implications for those who want voting on MeFi.
posted by languagehat at 12:09 PM on July 2, 2003
This bit, for instance, is extremely thought-provoking:
In the early Nineties, a proposal went out to create a Usenet news group for discussing Tibetan culture, called soc.culture.tibet. And it was voted down, in large part because a number of Chinese students who had Internet access voted it down, on the logic that Tibet wasn't a country; it was a region of China. And in their view, since Tibet wasn't a country, there oughtn't be any place to discuss its culture, because that was oxymoronic.
Now, everyone could see that this was the wrong answer. The people who wanted a place to discuss Tibetan culture should have it. That was the core group. But because the one person/one vote model on Usenet said "Anyone who's on Usenet gets to vote on any group," sufficiently contentious groups could simply be voted away.
Lots of implications for those who want voting on MeFi.
posted by languagehat at 12:09 PM on July 2, 2003
My apologies for following the statistical approach:
voting discussion
"The likelihood that any unmoderated group will eventually get into a flame-war about whether or not to have a moderator approaches one as time increases."
posted by woil at 12:22 PM on July 2, 2003
voting discussion
"The likelihood that any unmoderated group will eventually get into a flame-war about whether or not to have a moderator approaches one as time increases."
posted by woil at 12:22 PM on July 2, 2003
The division between "formal" and "informal" rules is something I've been thinking about in regards to Metafilter - the site is fairly unique in that the ratio of formal to informal controls is so low. Look at Slashdot, Plastic, Kuro5hin, Everything2, even newsgroups - moderation, meta-moderation, karma, distributed moderation. The ops decided technology was the best solution to the "problem" - but of course it isn't! Computers can count, but they don't think. Around here we can't even come up with an agreed upon criteria for "Newsfilter". Around here we can't even agree what the site is about - or at least make it obvious enough that new users don't have dreadful misunderstandings. But this is good! It creates a human dynamic. Sure it breed injokes and personality clashes, but that's life. It's why Slashdot is unfathomable, it's why Plastic is snore-inspiring, and it's why I got no work done yesterday day, constantly clicking my Metafilter bookmark in the vague hope I may receive a response.
posted by Jimbob at 6:15 PM on July 2, 2003
posted by Jimbob at 6:15 PM on July 2, 2003
+1, Informative, JimBob.
Actually I think Shirky's got it wrong, on at least two counts. First, he seems to presume that the power-law distribution of group size/activity in Yahoo groups is anomalous, and can be rejigged with some informed social-technical engineering, such as constitutions. I think however that the distribution in Yahoo groups is about what should be expected. Talking off the top of my head, you would be unlikely to get a distributed electronic communication system in which group size/activity follows a normal distribution; rather, you'd keep getting a 'Zipfian' distribution, with a few very large groups, and many many small groups, which is in fact what you find. MeFi shows Zipfian trends in a number of areas, for instance: in having a few very prolific posters and many irregular/lapsed posters; in having large numbers of referrals from a few sites and many referrals from many small sites.
Second, Shirky's use of Metcalfe's argument against scaling doesn't work either, cuz it assumes that, if speakers are vertices and conversations are edges, that the resultant graph is homogeneous. Whereas, assuming a Zipfian distribution, we would assume that a few people are having conversations with a lot of people, whereas most are just having one or two with one or two other people. Scaling a Zipfian distributed rather than a homogeneous graph means scaling a graph that is relatively empty (I think).
Upshot is, I don't think 'successful' groups are engineering themselves in readily identifiable ways that produce 'success.' I also don't think that 'unsuccessful' groups can make themselves a 'success' by copying 'successful' groups, or by attempting to formalise (for instance in constitutions) whatever it is they think that 'successful' groups are doing, that they (as 'unsuccessful' groups) are not.
*finishes talking out of hat*
posted by carter at 7:37 PM on July 2, 2003
Actually I think Shirky's got it wrong, on at least two counts. First, he seems to presume that the power-law distribution of group size/activity in Yahoo groups is anomalous, and can be rejigged with some informed social-technical engineering, such as constitutions. I think however that the distribution in Yahoo groups is about what should be expected. Talking off the top of my head, you would be unlikely to get a distributed electronic communication system in which group size/activity follows a normal distribution; rather, you'd keep getting a 'Zipfian' distribution, with a few very large groups, and many many small groups, which is in fact what you find. MeFi shows Zipfian trends in a number of areas, for instance: in having a few very prolific posters and many irregular/lapsed posters; in having large numbers of referrals from a few sites and many referrals from many small sites.
Second, Shirky's use of Metcalfe's argument against scaling doesn't work either, cuz it assumes that, if speakers are vertices and conversations are edges, that the resultant graph is homogeneous. Whereas, assuming a Zipfian distribution, we would assume that a few people are having conversations with a lot of people, whereas most are just having one or two with one or two other people. Scaling a Zipfian distributed rather than a homogeneous graph means scaling a graph that is relatively empty (I think).
Upshot is, I don't think 'successful' groups are engineering themselves in readily identifiable ways that produce 'success.' I also don't think that 'unsuccessful' groups can make themselves a 'success' by copying 'successful' groups, or by attempting to formalise (for instance in constitutions) whatever it is they think that 'successful' groups are doing, that they (as 'unsuccessful' groups) are not.
*finishes talking out of hat*
posted by carter at 7:37 PM on July 2, 2003
carter, some things are not clear in your comment:
Shirky: 'If you go into Yahoo groups and you map out the subscriptions, it is, unsurprisingly, a power law.'
you claim: "he seems to presume that the power-law distribution of group size/activity in Yahoo groups is anomalous"
"Scaling a Zipfian distributed rather than a homogeneous graph means scaling a graph that is relatively empty (I think)."
could you please give more details. I know power-law is scale free.
In our context, adding more members and maintaining the same gamma coefficient for the power law, we will end up with a strange fellow who will comment in every post, with a contribution index of ~50. Strangely, s/he would post interesting comments since too much noise in not tolerated here.
copying: then why do we see groups imitating the 'winning strategy' of successful groups? i.e. the karma model
posted by MzB at 8:29 PM on July 2, 2003
Shirky: 'If you go into Yahoo groups and you map out the subscriptions, it is, unsurprisingly, a power law.'
you claim: "he seems to presume that the power-law distribution of group size/activity in Yahoo groups is anomalous"
"Scaling a Zipfian distributed rather than a homogeneous graph means scaling a graph that is relatively empty (I think)."
could you please give more details. I know power-law is scale free.
In our context, adding more members and maintaining the same gamma coefficient for the power law, we will end up with a strange fellow who will comment in every post, with a contribution index of ~50. Strangely, s/he would post interesting comments since too much noise in not tolerated here.
copying: then why do we see groups imitating the 'winning strategy' of successful groups? i.e. the karma model
posted by MzB at 8:29 PM on July 2, 2003
Apologies, MzB - it's been a long day! Also I am not a mathematician (makes for fuzzy claims).
*resumes headgear mediated speech*
Re: Shirky's Yahoo claim. He says that there's a few 'highly populated groups,' some moderately populated groups, and "this long, flat tail of failure." I am suggesting that in the context of Shirky's article, his characterisation of these groups as 'failures' implies that the distribution within Yahoo groups could be different. If more Yahoo groups paid attention to what it is that makes 'succesful' groups successful, then less of them would be failures, he seems to be saying.
So I'm saying that Shirky is saying that the Yahoo groups distribution is anomalous, and that the source of the anomaly is ignorance on the part of the 'failures' of the sociotechnical properties of 'succesful' online groups. Remove the anomaly, and a different distribution of groups is achieved. (I actually disagree with this on a number of levels; for instance, I'm a member of some groups which post once a week or less - but when someone posts it's usually very good for me! - so I don't see presence in the long tail as necessarily indicating failure.)
Re. scaling. Shirky makes a point that the effect of scaling in a group where everybody talks to each other is a square of the rate of increase: "The fact that the amount of two-way connections you have to support goes up with the square of the users means that the density of conversation falls off very fast as the system scales up even a little bit." In effect he says that as we won't have time to talk with everybody in a large community, and that this is a negative. However, I think this model depends on assuming that all those conversations would have existed, whereas I think they would not have. And one of the reasons I think they would not have existed is that in the prescaling stage of the model, not everybody is connected to everyone else anyway. In a Zipfian sense, a few people have lots of connections, while a lot of people have few, and this will be the situation both pre- and post- scaling. This is why a Zipfian distribution is said to be scale free.
To put it another way: Shirky suggests a large (and alienating) increase in the complexity of a community with scaling. I suggest that this is not necessarily the case - and perhaps the level of complexity we perceive might remain about constant. There might just be more stuf with that level of complexity.
Re. contribution index of ~50 - MeFi might have to scale through a number of orders of magnitude to get to that. What are we at at the moment - max of about 8? (can't find link this late at night) And even then, that end of the curve is less well-defined; it's only a minute probability of a single occurrence, I think.
Re. copying. Yes people do copy - but is karma necessarily a winning strategy? Does it lead to better community or better comments? It does work for filtering 400 post /. threads, although there are also side effects (trolls proud to lose karma for example); but perhaps works less well for smaller groups - especially where people do know each other.
posted by carter at 9:27 PM on July 2, 2003
*resumes headgear mediated speech*
Re: Shirky's Yahoo claim. He says that there's a few 'highly populated groups,' some moderately populated groups, and "this long, flat tail of failure." I am suggesting that in the context of Shirky's article, his characterisation of these groups as 'failures' implies that the distribution within Yahoo groups could be different. If more Yahoo groups paid attention to what it is that makes 'succesful' groups successful, then less of them would be failures, he seems to be saying.
So I'm saying that Shirky is saying that the Yahoo groups distribution is anomalous, and that the source of the anomaly is ignorance on the part of the 'failures' of the sociotechnical properties of 'succesful' online groups. Remove the anomaly, and a different distribution of groups is achieved. (I actually disagree with this on a number of levels; for instance, I'm a member of some groups which post once a week or less - but when someone posts it's usually very good for me! - so I don't see presence in the long tail as necessarily indicating failure.)
Re. scaling. Shirky makes a point that the effect of scaling in a group where everybody talks to each other is a square of the rate of increase: "The fact that the amount of two-way connections you have to support goes up with the square of the users means that the density of conversation falls off very fast as the system scales up even a little bit." In effect he says that as we won't have time to talk with everybody in a large community, and that this is a negative. However, I think this model depends on assuming that all those conversations would have existed, whereas I think they would not have. And one of the reasons I think they would not have existed is that in the prescaling stage of the model, not everybody is connected to everyone else anyway. In a Zipfian sense, a few people have lots of connections, while a lot of people have few, and this will be the situation both pre- and post- scaling. This is why a Zipfian distribution is said to be scale free.
To put it another way: Shirky suggests a large (and alienating) increase in the complexity of a community with scaling. I suggest that this is not necessarily the case - and perhaps the level of complexity we perceive might remain about constant. There might just be more stuf with that level of complexity.
Re. contribution index of ~50 - MeFi might have to scale through a number of orders of magnitude to get to that. What are we at at the moment - max of about 8? (can't find link this late at night) And even then, that end of the curve is less well-defined; it's only a minute probability of a single occurrence, I think.
Re. copying. Yes people do copy - but is karma necessarily a winning strategy? Does it lead to better community or better comments? It does work for filtering 400 post /. threads, although there are also side effects (trolls proud to lose karma for example); but perhaps works less well for smaller groups - especially where people do know each other.
posted by carter at 9:27 PM on July 2, 2003
OK, I see your point now.
'Remove the anomaly, and a different distribution of groups is achieved.'
Besides your example, I could argue that the probability distribution will not change, meaning it will still be a power-law, but with a different coefficient gamma. An 'improved' distribution would have a lower coefficient and it would look more egalitarian (close to uniform distribution). It will still have a fat tail.
Example (relates to scaling): at the beginning, most of the posts and comments on MeFi were Matt's. This is the 'absolute monarch' model with high gamma. As more people started to contribute, Matt did not contribute more (too keep up) and the value of gamma decreased (again, assuming the power-law model holds).
'perhaps the level of complexity we perceive might remain about constant.'
very interesting: pre-scaling, each user 'collaborates' (e.g. comment in the same post) with several others (but not always the same others). Post-scaling: new members are in, I cannot ignore them because they post interesting stuff, so in order for me to maintain the same level of links (same level of complexity) I have to sever some of my old links.
- pro: this makes sense if we assume some sort of constrain, e.g. time constrain. Do you have any empirical data to support this?
- con: if we get a lot of yahoo/fark standard type users, the quality of comments would go down, and one will not be tempted to post or comment. Thus both types of links (old and new) will be cut.
'but is karma necessarily a winning strategy?'
It does not have to be a winning strategy, just a better one. I agree with your argument, it does not work in 100% cases, but if it works for 70% I'm happy.
Regarding your first comment ('I also don't think that 'unsuccessful' groups can make themselves a 'success' by copying 'successful' groups'), remember, people started revolutions to imitate successful people from other countries and they copied their strategies (constitution). In most cases, changes had to be performed to adapt the law to the local conditions. So imitation plays a very important role in group organization.
posted by MzB at 5:26 PM on July 3, 2003
'Remove the anomaly, and a different distribution of groups is achieved.'
Besides your example, I could argue that the probability distribution will not change, meaning it will still be a power-law, but with a different coefficient gamma. An 'improved' distribution would have a lower coefficient and it would look more egalitarian (close to uniform distribution). It will still have a fat tail.
Example (relates to scaling): at the beginning, most of the posts and comments on MeFi were Matt's. This is the 'absolute monarch' model with high gamma. As more people started to contribute, Matt did not contribute more (too keep up) and the value of gamma decreased (again, assuming the power-law model holds).
'perhaps the level of complexity we perceive might remain about constant.'
very interesting: pre-scaling, each user 'collaborates' (e.g. comment in the same post) with several others (but not always the same others). Post-scaling: new members are in, I cannot ignore them because they post interesting stuff, so in order for me to maintain the same level of links (same level of complexity) I have to sever some of my old links.
- pro: this makes sense if we assume some sort of constrain, e.g. time constrain. Do you have any empirical data to support this?
- con: if we get a lot of yahoo/fark standard type users, the quality of comments would go down, and one will not be tempted to post or comment. Thus both types of links (old and new) will be cut.
'but is karma necessarily a winning strategy?'
It does not have to be a winning strategy, just a better one. I agree with your argument, it does not work in 100% cases, but if it works for 70% I'm happy.
Regarding your first comment ('I also don't think that 'unsuccessful' groups can make themselves a 'success' by copying 'successful' groups'), remember, people started revolutions to imitate successful people from other countries and they copied their strategies (constitution). In most cases, changes had to be performed to adapt the law to the local conditions. So imitation plays a very important role in group organization.
posted by MzB at 5:26 PM on July 3, 2003
Somewhat belatedly: Yes I generally agree w/ what you say, and you also raised some points I hadn't thought about (e.g. gamma decreasing as size of MeFi increases).
Re. lower gamma approaching a uniform distribution. An 'improved' distribution would have a lower coefficient and it would look more egalitarian (close to uniform distribution). I'm not sure that fat-tailed distributions would ever get to the stage where they would look like uniform distributions (that's what makes them such distinctive signatures); although I am sure that either a uniform or a normal distribution is what Shirky thinks represents the presence of 'successful' web communities.
Re. pre/post scaling: Yup generally agree with your model. pro: this makes sense if we assume some sort of constrain, e.g. time constrain. Do you have any empirical data to support this?. I have a hunch that the constraint might be a cognitive/information processing one (of which time is an example). In other words, in a larger community I'd get to recognise a few names and keep checking out their posts - this saves me cognitive load, as I'm going with who/what I know, although I'd also check out unfamiliar stuff. Perhaps if everyone else did the same, clusters would emerge over time.
Regarding empirical data - perhaps cog psy types have done work on the amount of serious relationships people can keep track of in any community. I've heard 6 or 7 bandied about before. In terms of emergent effects, I know some archaeologists in the Southwestern United States have worked on models of rank of settlement size using measures of face-to-faceness vs. isolation/alienation as variables - that is, how big can villages get before people fell they're not in a village any more and so stop expanding - but I don't have the cites with me at the moment.
Re. Farkers, even if they came - I'd probably ignore them. Although S/N could go down in value, the absolute value of signal might increase.
Re. imitation ... I'll have to think about that one.
BTW, one of the reasons I'm interested is cuz I'm looking at the design of collaborative workspaces - and also looking for metrics that might signify a 'healthy' or 'unhealthy' community.
posted by carter at 2:34 PM on July 7, 2003
Re. lower gamma approaching a uniform distribution. An 'improved' distribution would have a lower coefficient and it would look more egalitarian (close to uniform distribution). I'm not sure that fat-tailed distributions would ever get to the stage where they would look like uniform distributions (that's what makes them such distinctive signatures); although I am sure that either a uniform or a normal distribution is what Shirky thinks represents the presence of 'successful' web communities.
Re. pre/post scaling: Yup generally agree with your model. pro: this makes sense if we assume some sort of constrain, e.g. time constrain. Do you have any empirical data to support this?. I have a hunch that the constraint might be a cognitive/information processing one (of which time is an example). In other words, in a larger community I'd get to recognise a few names and keep checking out their posts - this saves me cognitive load, as I'm going with who/what I know, although I'd also check out unfamiliar stuff. Perhaps if everyone else did the same, clusters would emerge over time.
Regarding empirical data - perhaps cog psy types have done work on the amount of serious relationships people can keep track of in any community. I've heard 6 or 7 bandied about before. In terms of emergent effects, I know some archaeologists in the Southwestern United States have worked on models of rank of settlement size using measures of face-to-faceness vs. isolation/alienation as variables - that is, how big can villages get before people fell they're not in a village any more and so stop expanding - but I don't have the cites with me at the moment.
Re. Farkers, even if they came - I'd probably ignore them. Although S/N could go down in value, the absolute value of signal might increase.
Re. imitation ... I'll have to think about that one.
BTW, one of the reasons I'm interested is cuz I'm looking at the design of collaborative workspaces - and also looking for metrics that might signify a 'healthy' or 'unhealthy' community.
posted by carter at 2:34 PM on July 7, 2003
Thanks for the answer. Regarding your research, I would suggest to look not only at a static metric (snapshot in time), but also at a dynamic one, e.g. rate of change. Some cluster analysis / neuronal networks might come in handy here.
posted by MzB at 4:00 PM on July 8, 2003
posted by MzB at 4:00 PM on July 8, 2003
You are not logged in, either login or create an account to post comments
posted by seanyboy at 5:19 AM on July 2, 2003