AI4Communities & Bluesky (newsletter)
I’ve been on a Bluesky deep dive, and came up hooked (for now).
(This is my mid November newsletter. Browse/Subscribe here.)
Since my last edition I’ve been on a Bluesky deep dive (12 resources and counting), and as a result just updated my AI4communities post (we’re now on v6) and my Hub’s Fediverse overview (v4).
Bluesky and AI4Communities
I haven’t got enough space here to properly explain how Bluesky works, and how it differs from Twitter or Fediverse apps like Mastodon: suffice to say that it combines the best of both worlds.
What matters is that Bluesky already supports AI4communities, but not according to the “cozyweb village” paradigm I originally developed.
Instead, communities on Bluesky can coalesce around Bluesky labellers and custom feeds: two tools (AI-powered or otherwise) anyone can build to create comfort zones within the wider network. While these zones will have both high signal2noise (ie, focused on the community’s topics of interest) and safety standards respecting the community’s norms, however, they’ll be porous:
- Bluesky’s openness means there will not be separate villages, nor private groups for collaboration
- anyone can subscribe to the labellers and custom feeds, and join the community’s conversation
Crucially, labellers and custom feeds can be built by anyone and published anywhere, without asking Bluesky for permission. So it would be quite easy to imagine a Friendly Neighborhood Algorithm Store charging access to a community’s AI-powered labellers and custom feeds, and sharing the revenue with them.
If AI4communities will look different on different protocols, this is a feature, not a bug, as it might lead to “protocol competition” in much the same way that platform competition in the telco industry helped accelerate broadband rollout.
Kicking Bluesky’s tyres for the Brussels Bubble
It’s still early days, however: while there were already ~40k custom feeds by early 2024 (source: MacKuba) I didn’t come across any service offering AI support as I started working on a Bluesky feed for the Brussels Bubble. I ended up testing two services to create two feeds (they’re both in the BlueSky Brussels Bubble Starter Pack):
- I created Brussels Bubble with the free tier of the Bluesky Feed Creator fremium service, which claims to host a little over 1400 feeds.
- I created brusselsbubble2 with the free Skyfeed service, which apparently hosts the vast majority of feeds created to date.
Left: Skyfeed is actually a Bluesky client (anyone remember Tweetdeck for Twitter before X bought it and restricted it to premium users?), and comes with a free feed builder, shown above. It has … let’s say a challenging look and feel, and was apparently built by a single developer (the github repository is anonymous, and it doesn’t even have a Wikipedia page).
While Bluesky Feed Creator (right), on the other hand, looks much slicker and easy to use, it is less powerful: Skyfeed users can use “regular expressions”, which are far more powerful than Bluesky Feed Creator’s simple keyword matching, and allows you to choose from a list of different ranking and sorting algorithms. Bluesky Feed Creator, however, does provide language filtering, which is incredibly useful (did you know that “eu” is a word in Spanish?).
So, which one actually created a better feed? There are two measures: completeness (did the feed capture all the posts it should?) and noise (did it capture many irrelevant posts?). I just checked — they couldn’t be more different:
- Brussels Bubble (Bluesky Feed Creator) presented me with 42 posts in 1 hour, of which 12 didn’t seem to have any of the keywords I was looking for — ie, just under 30% noise, which is pretty high
- brusselsbubble2 (Skyfeed) only presented me with just 3 posts in 1 day, all relevant.
I suspect the regular expressions I used in Skyfeed are to blame for the latter: I’m not a developer, and that’s the point. For communities to settle in and grow on Bluesky using custom feeds and labellers, an ecosystem of third party tools needs to emerge that someone like me can use.
Personally, I’m pretty hopeful: the developers are active on Bluesky, explaining features and answering questions, while the entire project is on Github, where users such as you and me can suggest new features and whip up support from other users. The network, finally, is growing, making the marketplace more and more compelling. And I still have a lot to learn. So say Hi if you’re on Bluesky.
Next, I’ll be checking out Nostr and (hopefully) Farcaster, where social media hits the blockchain. Wish me luck.
Stuff worth reading
More on Bluesky
If you’d like to learn more about Bluesky, everything I’ve read is on my Hub. Of these, I’d recommend:
A complete guide to Bluesky: one of many guides out there, but extensive, from “the tips & tricks that I often give to friends when I send them an invite code”, through to a brief but informative history lesson.
Bluesky’s Stackable Approach to Moderation, where the Bluesky team announce that they’d open-sourced “Ozone, our collaborative moderation tool… individuals and teams can work together to review and label content”, coupled with the ability for people and communities to “run your own independent moderation services, seamlessly integrated into the Bluesky app”.
A quick and dirty guide to making custom feeds on Bluesky, for those interested in using Skyfeed.
And for a contrary view: in Bluesky and enshittification, Cory Doctorow, who coined the term, is not convinced they won’t.
Creative centaurs
Just one paper in this edition about designing AI services to maximise human potential: in Ensuring Human Agency: A Design Pathway to Human-AI Interaction, the authors built on cognitive learning theory to propose a conceptual model for designing AI outputs which require humans to think by providing “open-ended, reflection-provoking feedback… [so] the user must expend effort to arrive at an answer.”
It does this by providing “evaluative feedback and critique… [rather than] explicit, outcome-focused recommendations, and focuses on improving a user’s understanding” of the area being explored, rather than focusing their education on how to use AI.
More centaur papers.
Ruben Verborgh
I’ve just discovered Ruben Verborgh, and have only read two articles so far:
Paradigm shifts for the decentralized Web is from 2017, yet sets out exactly what is required, allowing people to “store their data wherever they want, while still getting the services they need”, rather than having to “accept package deals we cannot customize”.
This means users become data controllers, apps become views of the data which we choose they can access, and interfaces become queries. As he points out, this approach creates “separate markets for data and applications”, thus stimulating innovation in both, which is essentially what we are starting to see on Bluesky.
He also remarks that “if we really want free options, we could even imagine paying with our personal data”. That’s the basic idea behind AI4Communities, and its something he returns to in Trust takes time, a really beautifully written essay on trust, using the author’s experience as a father to make it personal.
Online we live in a trustless world: as we click the “I accept” button “we know we are lying … as does the author of the legalese no one expects anyone to read… There are no winners when there is no trust … companies have nowhere near the most valuable data [from us], and people are getting almost no value out of their own data” (my emphasis).
It’s too long to summarise here, except perhaps for his conclusion, which is that the EU’s GDPR regime is “making it marginally more expensive to do the wrong thing… [and] substantially more expensive to do the right thing… [We need] a new generation of techno-legal systems… [to] set up and maintain long-term relationships and negotiate mutual benefit”.
PS. But how do I quit Twitter?
I know. You’ve spent years building a following on X, and the thought of having to rebuild it from scratch on Bluesky fills you with dread. But staying on X doesn’t seem so great, either.
So feel free to borrow my strategy.
