A popular old saw due to Robin Dunbar, which had a resurgence in the blogosphere a couple of years ago for the most classic of reasons for something to have a resurgence on the blogosphere - it seemed to be about the blogosphere.
The original paper regresses mean primate group size against “neocortex ratio” (ratio of neocortex volume to the volume of the rest of the brain) for various species including humans, identifies a goodish fit and posits that neocortex capacity serves a basically social function whose limits are demonstrated by this group size limit. He then goes on to note that it is of course a bit more complex with humans, and suggests a tri-modal distribution of group sizes for human based upon hunter-gatherer societies (who are the most ‘natural’ groupings.) Specifically, human groups tend to be organised in fluid, nested hierarchical sizes which cluster around group sizes of 30ish, 150ish and 1000ish.
Confidence intervals for the neocortical regression has a factor of 3 margin for error, and the group sizes themselves are highly fiddly to calculate, and hard to disentangle from, e.g. ecological carrying capacity constraints. But it’s a neat idea, eh? Certainly Dunbar has been able to dine out on it for years, which is the hallmark of neatness.
This is a bit too sloppy for actual predictive use. For a start, Dunbar suggests these numbers are good for predicting long-term sustainable group memberships, but that short term group sizes might be different. Next the magic numbers are all too sloppy to be of actual use. Finally, numbers like this are an invitation to confirmation bias - please, go forth and select group sizes that look like they correspond to those numbers. but even if one does a comprehensive survey of group sizes, what are we identifying here? A human cognitive constraint? A constraint on time investment in relationships? Or on task-coordination?
The web-driven resurgence in the idea has been about the discovery of many groups online that happen to fit Dunbar numbers, online communities being fertile ground for group spotting. I haven’t seen any examples of this that are wallowing in the aforementioned confirmation bias - there’s usually a way you can count community sizes that means that you found the Dunbar group size (Wikipedia editors above an arbitrary activity threshold, for example), but no special reason to believe that you have thereby found the canonical way of counting the group. I still think it’s suggestive, and provokes some interesting questions, but I’m not going to rely on these numbers in any predictive capacity.