Rob RicciThanks very much for these thoughts! I really appreciate your help and patience filling in the gaps in my knowledge here. FWIW I'm playing around with these things hands-on to learn more, I set up my own PDS last night and hope to try my own relay, either with the reference implementation or blacksky's rsky this week.
On PLC: The number of did:web handles for the network appears to be - according to another dataset I found that scrapes handles - 159. I'm not confident in saying that's the complete set, but the number of PLC handles would appear, conservatively to be north of 99.9% (the dataset I am using, which is coming from the bluesky relay, contains 402k handles, so that's assuming that I'm seeing a representative sample of the full set of handles)
My conclusion here is that the B-Index for atproto is, depending on your viewpoint, very boring or very revealing. Unless the handle data I'm looking at is very incorrect, the B-Index for the *current* atmosphere is effectively $B_X$ = 1 for all $X$ at least up to 99.9; whether or not you believe they *would*, through administrative decisions or company failure Bluesky PBC could cut off any or all users from 99.9% of the network. I'm sure some people find that totally unsurprising, while others are probably quite surprised.
On your third point, this is actually kind of what I'm trying to capture with B-Index. Yeah, it does seem like (maybe other than the PLC) that someone who is using blacksky's appview, PDS, and relay, is going to be just fine within the blacksky network - but they will be cut of from 99.999% or so of the atmosphere. The B-Index is not trying to model whether they care - some people will be perfectly fine with that scenario, but rather model the centralization of power that gets one to that point; ie. one company can decide to cut me off from 99.999% of the world, vs. if m.s blocks me (worst case on the fediverse) that cuts me off from 25% of the world.
On the Mississippi situation, my answer is kind of similar: how many companies (or instance admins, etc.) does it take to make a decision that affects X% of users. On the VPN thing, I think we can't model it, since we can't really model the how accessible VPNs are. On the alternative appviews, we can, though, if we can get user counts for them; here the B-Index may be higher, because more entities have to make the decision to block.
So some other thoughts here:
* One of the interesting things (I hope) about B-Index is that you can use it for little 'what if' experiments, eg. we could move on from the PLC, assume that it will not be used to revoke DIDs and that it will not go down, move on to the next 'largest' resource and go from there: I assume that's likely to be the bluesky relay or maybe the appview. But this is where I need to understand better what someone on, say, blacksky, who has been banned at the appview and/or relay level from bluesky can or cannot see. Maybe once I have my own relay to play with I'll have a better understanding.
* I think it would be interesting to create some version of this metric that looks not at blocking but credible exit, separately for social graph and user data (which as you point out is going to cover an exceedingly small number of Fediverse users today). Though it sounds like atproto might run into the same problem with PLC - but we can use the same technique of 'here's what it looks like today, here's what it would look like if the PLC were in the hands of a consortium, or truly decentralized, etc.
* Labeling and feeds are also going to be interesting here, I'm not giving up on trying to look at those just yet :) This will probably be the case where, assuming I can get appropriate data, the *fediverse* is the one that's complicated to model, since each instance has its own blocked
* As you point out there are DNS and other external dependencies, and I'm interested in those as well, I did some rough numbers for the fediverse in terms of networks that hosts are on a while ago, I'm sure this could be formalized: https://discuss.systems/@ricci/114396317436420669
* I think it would be interesting to extend the kind of thing noted in my previous point to legal jurisdictions as well.
All GNU social JP content and data are available under the