(I originally wrote this to be a twitter thread, but I don’t think I have the hang of writing good twitter threads yet.)

The internet has connected the world. But despite the wealth of “possible” connections at our disposal, many still feel alone and isolated. Why is this?

In this post, I want to look at this question from the perspective of the structure of our social graphs (who is connected to whom)—or perhaps more accurately, the hyperparameters which influence how our social graphs form: Things like the recommendation systems or the presence/absence of physical/location constraints. While this contemplation is based on the prompt, “Why are we all lonely?” the topic of network hyperstructure and how it affects our lives is really a creature all on its own. That is, it probably doesn’t come close to explaining fully why we might be lonely, yet it also helps explains a lot more about modern (online) life than mere loneliness.

The simplistic take

Let me start by offering a very simplistic take linking modern social network structure to loneliness:

The transition of social life from offline to online has seen geographically localized networks merge into massive, delocalized online networks. And the patterns of connection which tend to arise in these massive networks are less amenable to friendships. Why is this?

1. As someone with finite time and cognitive resources, I cannot attend individually to each person in a massive network. Only the most salient/popular/prolific/interesting individuals will receive my attention. In a network like Twitter, I’ll follow them and possibly make some bid for connection.
2. But because everyone else is likely following a similar strategy, these salient individuals will be saturated with such bids; they certainly won’t be connecting in deep, meaningful ways with each one of their followers.

Put differently: as the global network de-localizes, the predominant emerging network structure puts much more attention on a much smaller group. And since collective attention is finite and fixed, this means less attention for the rest of the network. But since reciprocal attention is the basic starting point for the formation of friendships and relationships, much of the network fails to get friendships off the ground.

Beyond isolation

Let’s suppose that the simplistic take above describes an actual phenomenon in a roughly correct way. A next question that we can ask is what kinds of incentives this creates for members of a network. The best spoils of participating in a social network come from having many reciprocal connections, which requires getting through the attention filter. How do I succeed at this?

There are certain pretty obvious incentives, like that of being prolific. You have to post a lot before most people will notice you. And what counts as “a lot” is something that can be subject to escalation.

I’m a bit more interested in other incentives, which I’ve seen less discussed. Any network can have the effect of pushing someone toward a persona which isn’t overly represented within their network. Siblings do this within the highly localized network of a family. If an older sibling has already established themselves as the one who likes A and is good a B, a younger sibling may find that they can compete for attention more easily by being interested in C and pursuing D.

In the context of a global network, almost any profile of personality, likes, or competencies is likely to be well-represented within the nexus of the network. Does this mean that people find themselves pushed into ever more niche ways of expressing and representing themselves in order to vie for attention? It feels to me that this is likely the case.

Other equilibria

Again, let’s suppose that the dynamics that I’ve outlined above are more or less correct. A next question is whether there are other network hyperparameters–things like recommendation algorithms, etc.–which lead to generally more favorable outcomes.

The tricky question is how to measure the goodness of outcomes. Perhaps the network hyperparameters of Twitter are uniformly better for certain people than the network hyperparameters of a network like Facebook or one’s physical neighborhood; but worse for others. Changing the hyperparameters could make things better for many at the cost of making it much worse for a few.

With this caveat out of the way, I do find interesting the question of what could happen if you optimized profile recommendation algorithms for the objective of “two-way follow rate” instead of the objective of “one-way follow rate.” If someone designed a network like this, would most people choose it over a network like Twitter?

But my attention is not saturated

One of the things that I’ve spent time doing in the last month or two is looking at the actual structure of my local Twitter network. I’m not plotting graphs or anything. Just looking at who follows whom.

I’ve done this in the course of searching for accounts that were a little less prolific, a little less niche, a little less on the “cutting edge” of the Twitter vibe frontier, but still had interesting and thoughtful things to say. These accounts, comprising the long tail of Twitter, are where you might expect to find friends, I reasoned.

My results have been a little bit mixed. It’s certainly not the case that everyone follows everyone, but it’s also been harder than I expected to separate the “long tail” from the “nexus nodes.” Some possible reasons for this:

1. Finite attention is just not a limitting factor in the way that I’ve assumed.
2. People who can’t manage to get through the attention filter eventually throw in the towl and go home.

One reason that the first possibility could be the case is that, once you are within a community with content or interest boundaries that are narrow enough, you once again have a local subnetwork whose size is not so much of an issue. I wonder whether my experiment with Twitter will serve as a bit of a canary for which of these two is true, since I probably will never be incredibly prolific or niche. Stay tuned.