Richard Ngo recently posted the following poll on Twitter. I bet him that the mode of the poll would be the 7+ option. I lost this bet, because I am miscalibrated. Here, I try to explore the reasons why:
Firstly, fair warning: this post will be disappointing, primarily because Twitter only allows me to see about 68 of Richard's 34k followers, and this subset appeared to be highly non-random. (By contrast, it allows me to see all of my followers.) This seems to be a bug.
At the time that we made the bet, we were sitting on the patio at the Embassy in SF with an approximately random subset of people who would be hanging out at a San Francisco group house at ~9:30pm on an average Saturday night. Of the 5 people present, 4 had lived in 7+ places during their lives. I thus bet Richard that the mode of the distribution of this poll would be the 7+ option, which captures the tail of the distribution. (Richard countered that the average person probably grows up in one place, goes to another place for college, and then settles down in a third place, so the mode should be 3.) Central to my conjecture was the belief that the average Richard Ngo Twitter follower would look approximately like the average person hanging out at the Embassy.
Clearly, this conjecture was incorrect. Interestingly, however, a recent Pew poll found that 37% of Americans live in one town throughout their lives, which is much more than the 20.8% responding "1 or 2" in Richard's poll. Thus, Richard Ngo followers are significantly more mobile than the average American, even if they are significantly less mobile than the average San Francisco Group House attendee.
So, who follows Richard Ngo on Twitter? This question turned out to be remarkably difficult to answer. As mentioned above, Twitter would only reveal to me 68 of Richard's followers; I tried checking with a different account, and got a different list of 68 that was about 75% overlapping with the original list. Of these 68, the majority appeared to be either intellectuals (e.g. @AdamMarblestone, @nickcammarata), media (e.g. @RobertWiblin), troubled souls (e.g. @alexeyguzey), or people with a clear declared interest in AI safety and alignment (e.g. @stefanfschubert, @broad_priors). There were also a surprisingly large number of finance people (@tradernewsai, @anshumanmishra), and there were 14 accounts with no identifying information at all, which appear to be bots. With the exception of the bots, most of these people seem like they would be at home in an SF group house. However, it seems highly likely that Twitter is showing me a biased subset of Richard's followers, because I knew a large subset of the followers it showed me, which would be highly unlikely if the subset were random.
To try to get to the bottom of this, I decided to post the same poll to my followers, to see if my followers differ significantly from Richard's. After 1 hour, our distributions were strikingly different:
Over the following 23 hours, however, my distribution converged towards his distribution:
The drift here cannot be explained by statistical variation. Instead, it seems like early voters do tend to move around more. There are a few possible reasons for this: early voters may be people who engage more with my content and thus tend to see it first. It would make sense that that population is more mobile. Alternatively, it might be that early voters are Twitter power users, and that Twitter power users tend to move around move.
My only other possible conclusion from this was that it seems like my distribution is enriched for people in the 7+ category at the expense of people in the 1-2 category, compared to Richard. It seems consistent with a hypothesis that, as you gain followers, you get more "normal people" in your follower pool, who may tend to move around less.