Home

Research

experiment log
experimental methods
agent prompts and gpts
summarisers in summary
Ideas log

About

about this massive wiki
MyHub.ai / ChatGPT integration: - today's pilot
- LLM integration plans

About me / Contact

General Navigation

All Pages
Recent Changes

AI4Communities: a model for a self-sustaining, user-owned public sphere

Model collapse may provide the funding for accelerating decentralised collective intelligence. What needs to happen next?

(Notes: This is an early draft. As explained in this newsletter edition, I am publishing these early versions as I develop my thoughts in the hope that constructive comments will help me finish the post. More version control in the footer.)

In my last post I highlighted how the authors of the model collapse paper in Nature pointed out that in the future any data stemming from “genuine human interactions … will be increasingly valuable” for training the next generation of LLMs. Online, there’s nothing more genuinely human than the interactions in a well-managed digital community, so “after greed and short-sightedness floods the commons with low-grade AI content… well-managed online communities of actual human beings [may be] the only place able to provide the sort of data tomorrow’s LLMs will need” (How Model Collapse could revive authentic human communities).

This supports the AI4Communities idea, first outlined in A Minimum Viable Ecosystem for collective intelligence in January 2023, which envisions tomorrow’s online denizens thriving in a diverse landscape of “small is beautiful communities, supported by AIs which the communities themselves own, train and monetise”, rather than putting up with whatever Meta’s or X’s global, “one size fits all” algorithms inflict on them to turn a profit.

If model collapse has made AI4Communities more feasible, what is it exactly, how do we get there, and who’s building it? But first, why care?

What is the Cozyweb, and why is Small Beautiful?

According to the Three Legged Stool manifesto, “Very Small Online Platforms (VSOPs)” are “social networks created for a very specific purpose, with rules, norms and affordances appropriate to that community” (A social network taxonomy). While they’re obviously playing with the EU’s term for “Very Large Online Platforms”, they see VSOPs as separate from federated servers, which combine small size and large reach thanks to protocols like ActivityPub (the Fediverse), AT Protocol (Bluesky) and Nostr.

But for the purposes of this post I’ll call them all “cozyweb”, partly to keep things simple, but mainly because:

  • AI4Communities need to works for everyone, because it needs scale — something it currently lacks, as explored below;
  • they have something very significant in common: their manageable size and/or federated structure makes it easier for inhabitants to establish and enforce community norms, making community management possible, as “it’s fundamentally impossible to do good content moderation or conversation at a scale of billions… A town square operates at a certain size, not at a global scale” (Could social media support healthy online conversations?)

Small doesn’t mean insular

Which is not to say they’re isolated: while each cozyweb community has its own content moderation policies tuned to its needs, most can be networked together without sacrificing independence.

So while each cozyweb can be a village — “small, most people know each other and you all share a common interest in keeping the sidewalks tidy” — each can build federated roads to many others, including bigger ones “crowded with people, plenty of them sleazy and more than the occasional sidewalk madman… But you’ll always discover something or someone new there. Every second person’s selling something, but one in 20 is selling what you need. Besides, you’re selling too…” (Welcome to the Fediverse, starry-eyed noob).

Just as each Cozyweb community chooses its own “village rules”, with AI4Communities it manages its own AIs to support the community: some help users with content discovery while guarding the village gates against trolls and trash, for example, while others support community moderation and collaboration processes. 

So if small is beautiful, why is everyone trapping each other on a few massive, genuinely awful platforms? 

Who pays?

Must small be poor?

Decentralised, federated networks have been around almost as long as Twitter and Facebook (Evan Prodromou’s Identi.ca launched in 2008, while linkbacks, invented to knit blogging conversations together, came much earlier). But although they are growing now, they are tiny compared to the closed gardens of surveillance capitalism, at least partly because — unlike their closed counterparts — they lack a business model, and hence revenue, as I found out the hard way almost a year ago:

“Social infrastructure needs to deliver content as a basic minimum. And running that infrastructure takes time, money and professionalism” — All my toots gone

The Three Legged Stool authors believe “many VSOPs will likely receive most of their revenue from the communities they serve, similar to how a local newspaper or nonprofit is funded”, and I hope they’re right. Today, however, only a small fraction of Mastodon users currently donate to cover their server’s costs (anecodotal).

AI4Communities: when community monetises AI (not vice versa)

People may pay more for a better experience. An AI owned by the community, which reflects their preferences and adds value in many ways (as explored below) could provide that experience.

These communities therefore need access to what the Three Legged Stool manifesto calls a “Friendly Neighborhood Algorithm Store”. Many of the AI services available there will support individuals (cf Bluesky users subscribing to custom feeds and block lists) but other services could support communities, and can be configured and fine-tuned by the village to ensure it reflects their interests and preferences.

Moreover, as I pointed out in 2020 when I launched myhub.ai, each community could act as a data union: rather than just buying or renting an AI to support their community, they could monetise the resulting algorithm to at least help cover the costs of running the community. While I shelved this idea when ChatGPT appeared, the model collapse paper now suggests that the training data created by well-managed communities could be the new currency of collective intelligence. 

training data created by well-managed communities could be the new currency of collective intelligence

Ideally, the marketplaces would be able to serve all protocols seamlessly, helping create the scale required. 

Because scale is currently missing. The Fediverse has around 12 million active accounts, Bluesky around the same (check), with Nostr a very distant third. Most users in these ecosystems are using Twitter and Facebook clones (check). 

The key question is therefore: would 20–30 million users, using such superficial applications, create enough high-quality training data?

AI4Communities per network

How would this look on each of these networks?

Subfiles to be developed:

As I work on the above files, an idea is emerging in my mind: AI4communities will look different on different protocols, and this is a feature, not a bug, as it might lead to "protocol competition" in much the same way that platform competition in the telecommunications industry helped accelerate the rollout of broadband.

Note: the rest of the post has not been rewritten to reflect AI4Communities on the ATmosphere, and so still reflects a Fediverse bias.

Deeper apps, better data

What social applications will be both popular and generate valuable AI training data?

For AI4Communities to work we need social apps which people want to use and which generate high-quality training data. So what services could these “cozyweb AIs” provide, and within what apps? 

... section to be developed(?)

need to know about Bluesky and Nostr:

  • how algorithmic freedom is/will be delivered
  • possible diversity of applications

Content discovery services

Content discovery is what happens when content you like or needed to see is presented to you.

Bluesky and Nostr already provide algorithmic choice, so let’s start with the Fediverse, where users only get three reverse chronological feeds: from everyone you follow; from everyone in your village; and from everyone followed by at least one person in your village. None of these feeds include content from villages which your village does not federate with, which is a feature, not a bug (although it could limit the effectiveness of AI-supported content discovery).

Introducing AI4Communities, however, means that each instance can access AIs which can proactively pull in and surface content from across the Fediverse (excluding un-federated servers) of interest to each village inhabitant. These customisable “For You” feeds will be driven by relevance and trust, as explored earlier:

! (caption) Concentric trust circles of collective intelligence, from Building collective intelligence from social knowledge graphs

So how is this any different to X, Threads et al? Simple: your village’s algorithm works for you, not the platform’s owners and advertisers. Don’t want it to optimise your feed for enragement? Tell it not to. Don’t like engagement bait, NSFW comedy or culture war memes? Tell it so. Only interested in a few topics? Let it know.

your village’s algorithm works for you

And if your tastes don’t match the village’s, move to another village. Balancing an individual’s feed parameters with the village’s content moderation rulebook should be a fun problem to tackle, but far from insuperable, as each village can find its own balance. Which brings me to…

Content moderation and Governance services

Enforcing any online space’s rules on content moderation is a thankless and sometimes hideous task, so the major platforms have already invested fortunes into AI-supported moderation. Open-source LLMs now make it possible to develop something similar for cozyweb spaces.

Each village’s moderation AI could intervene in many different ways, from nudging members to rephrase a specific post to outright banning another for life. It would also intervene at several stages of content development, from before a member clicks the “Post” button through to the moment a conversation starts overheating.

Most of all, of course, the AI will constantly learn from the village’s inhabitants what they like and dislike.

The key to success will be effective community co-ownership of the village rulebook which the AI enforces. After all, if you no longer agree with your village’s rules you can move to another one easily, so ensuring everyone has a say in defining its rules is essential for any village to thrive.

a community working together to train their village AI

This raises the fascinating possibility of a community working together to train their village AI. Obviously, any member would ideally be able to contest decisions — the conversations around these edge cases could then help fine-tune the rulebook. The AI should take an active role in these conversations, effectively being coached by the community on norms and practices. I suspect the coaching will be bidirectional much of the time.

Well-run large communities will therefore develop effective content moderator AIs reflecting the community’s values. The data used to create these AI will be valuable, as will the AIs themselves.

Centaur services for individuals...

Chatting with strangers and making friends on a social media platform is only a small part of collective intelligence. How should AI help members of a cozyweb community be more creative and productive, maximising their human potential?

I think everyone is familiar with the idea of the concept of "centaurs" (myhub: #centaur), but the key role of "centaur services" is to avoid us becoming "reverse centaurs" - human beings turned into "OK-button-mashing automatons" by the requirement (often driven by regulation or marketing) to keep a human in the loop.

Research in this field tends to result in prompt libraries and frameworks for using LLMs - for example:

  • AI And The Decline Of Human Intelligence points out that “Our brain is a muscle; it needs to be exercised,” which is a problem given how easy it is to outsource our skills to AI. So it provides a number of useful-looking prompt frameworks (Socratic Method, Feynman Method, Debate Partner, etc.) to help you treat AI “as a tutor or consultant, guiding us like a teacher”.
  • This echoes ChatGPT as muse, not oracle, from February 2023, but still well worth reading.
  • AI as Extraherics: Fostering Higher-order Thinking Skills in Human-AI Interaction is an academic paper with a similar goal, proposing an AI usage framework called “extraheric AI” which “fosters users’ higher-order thinking … creativity, critical thinking, and problem-solving… by posing questions or providing alternative perspectives … promoting a balanced partnership between humans and AI”.

So rather than provide a generic "send this to an LLM" system, as is currently the case on myhub.ai (see How to chat with ChatGPT about your content) and dozens of other apps, AI services marketed to cozyweb communities will need to provide AIs in the form of process agents which "lift up" our thinking and help us learn, rather than product agents which focus on turning us into reverse centaurs, mindlessly OKing whatever product the AI provides.

The resulting content should be a high-quality source of AI training data.

... and collective intelligence

Things get even more interesting when you create agents to help a community collaborate and produce something together.

If you want to take things further and co-create something together with a group, you have a problem — hardly any social media platform offers integrated collaboration tools (exception: Reddit’s community wikis).

Which is why you need to agree, as a group, on a collaborative tool everyone is comfortable with (TL:DR; it doesn’t exist for groups larger than 4), and then co-create on one platform while chatting on the other.

I’m not sure why this is so. Work-oriented collaboration platforms offer reasonably seamless chat and collaboration environments. In the public realm, however, there seems to be a stubborn divide separating wikis and other groupware from blogging and social media.

A social ecosystem where users can seamlessly collaborate

I explored one solution almost 4 years ago in Thinking and writing in a decentralised collective intelligence ecosystem which, as the title suggests, focuses on improving the collective intelligence of social groups online, while this video introduces the idea of a social ecosystem where users can seamlessly collaborate (and then explores massive.wiki, an early example of what the collaborative part of the overall technology stack might look like, and the tool I use to manage this post).

If a social ecosystem where users can seamlessly collaborate is a worthy goal in itself, then how can AI help? Some ideas on how that would look can be found in How to Use AI to Build Your Company’s Collective Intelligence, which explores how AI can help “increase the collective intelligence of the entire organization… through boosting collective memory, collective attention, and collective reasoning”.

However, this article also identifies some risks (in the process contradicting itself somewhat) - for example, bringing an AI voice assistant into a collaborative effort created groupthink (Hub: #groupthink):

  • teams aligned their attention to the assistant’s and "even adopted the [assistant’s] specific terminology … which further shaped where groups directed their attention”
  • AI usage “caused a decrease in intellectual diversity … Through a form of algorithmic monoculture”.

The AI services marketed to cozyweb communities to support collaboration, in other words, must not only help the individual members boost their creativity (see above), but will also have to reinforce the positive and dampen the negative dynamics of groups. The most successful agents will probably have as many psychologists, facilitation and negotiation experts behind them as data scientists.

Finally, if any collaborative work is to be supported by your village’s AI, the group involved would obviously need to give its consent. Which in turn means that these collaborations become another source of high-value AI training data.

Better apps

As mentioned above, Bluesky and Nostr already offer some elements of the infrastructure required for this.

Bluesky, for example, already provides some pretty cool content discovery tools via feeds, developed and shared by users. The first I installed, for example, was “Science”, developed by Dani cRabaiotti, a zoologist and data scientist working at the Zoological Society in London.

This is possible because Bluesky wants to give users Algorithmic choice. Instead of suffering the problems created by a one-size-fits-all "master algorithm", as seen on platforms like X and Facebook, Bluesky users can choose from an open and diverse "marketplace of algorithms". In this approach, algorithms "act as aggregator services, similar to search engines" - users add algorithms to their clients seamlessly, with Bluesky allowing them "to swipe between favorite algorithms or view a multi-algorithm feed".

Bluesky doesn't even control feed creation. Instead, they provide "APIs for feed generation... allowing for custom feed and moderation systems to be created as independent services ... [and] a feed selection system that enables users to explore third-party feeds and access them as effortlessly as their home timeline... we’re using a similar approach to address reputation, misinformation labeling, and moderation."

I've therefore started work on a Bluesky feed for the Brussels Bubble to see whether that will help them off their nasty Xitter habit, and have added it to the Bluesky Brussels Bubble Starter Pack. However, these are not AI-driven.

I’m less familiar with Nostr, but I know that Jack Dorsey, its founder and the guy who launched Bluesky when he ran Twitter, definitely believes in algorithmic choice. My introduction to Nostr came via one of my favourite thinkers in this space, Gordon Brander (6 resources), whose Nature's many attempts to evolve a Nostr walks you through the various architectures, concluding with the Nostr approach, of which he’s clearly a fan: “You sign messages with your key, then post them to one or more relays. Other users follow one or more relays. When they get a message, they use your key to verify you sent it”.

Currently it's not clear to me how this doesn't result in a splinternet: if two relays don't talk to each other then how do messages find other people not linked to your relay(s)?

But what Brander's piece did convince me (for now) is that the federated services architecture underpinning the ActivityHub-powered Fediverse has a fundamental flaw:  "Federated networks become oligopolies at scale", due to general forces seen everywhere: "airline routes, power grids, trains, banks, Bitcoin mining, protein interactions, ecological food webs, neural networks". It happened to email, and now it's happening with the Fediverse, where calls to defederate Threads (which was 10x the rest of the Fediverse when it arrived) were ineffective.

... section to be developed(?)

need to know more about Bluesky and Nostr:

  • am I talking out of my rear end?

One thing that bothers me is that Bluesky, as well as (I think!) most of the Nostr and Fediverse apps, are Twitter look-alikes.

My gut feeling is that for AI4communities to work in privacy-first spaces like these, the apps people use will need to be less superficial to generate higher-quality training data. After all, there’s only so much value in short status updates, and noone's proposing following users around the web, siphoning up their data to sell to marketers as Facebook et al do.

there’s only so much value in short status updates

For one example of how that could look, A Minimum Viable Ecosystem for collective intelligence suggests combining publishing tools (blogs, social posts), social bookmarking and personal knowledge management, and networking them together across federated networks, all supported by your personal AI. (caption) From Social knowledge graphs for collective intelligence (January 2023)

Because then your AI will be able to can learn from:

  • what you do on your social network(s)
  • every article you bookmark or Hub (and so find really, really useful)
  • the notes you make about it as you curate it (which tells the AI exactly what you find useful about it)
  • everything else (blog posts and professional references, mainly) which you publish alongside them
  • and everything else in your private knowledge library you make accessible to it.

That’s a lot of tightly interconnected knowledge, all reflecting your interests and expertise. Any AI sourcing that sort of knowledgebase will perform content discovery superbly well for you. And that AI is also being trained by everyone else in your village, with whom you presumably share some interests and values, and with whom you may also be collaborating (see below). Together, your community will generate extremely valuable AI training data, due to the authentically human curation and creation processes involved in making it.

A Brave, early inspiration

There is already evidence for this strategy on the market in a parallel field: search.

In May 2023 Brave Software, creators of the privacy-first eponymous browser and search engine, released Brave Search API. Via this service, AI developers can access the training dataset Brave developed through both Brave Search and their Web Discovery Project, which allows Brave browser users to contribute anonymous data about the pages they’re actually visiting.

As a result, Brave’s training data “is much more representative of the Web people actually care about”, without the huge morass of “junk sites — spam, dead links, duplicates, and more” one finds when crawling the Web at large. And that was written before the Web started filling up with AI-generated content, which is even worse for training AI.

AI4Communities is the social media equivalent of Brave’s approach to search. Both revolve around authenticity-driven data quality:

  • People do not actually spend a lot of time browsing junk content, so Brave’s dataset doesn’t have any, and so is of higher quality
  • People in well-run communities do not curate, share and collaborate on rubbish content, so the data they create will be valuable, too.

Conclusions

... section to be rewritten(?)

Looking back over this post, I can see that it’s heavily influenced by my greater familiarity with the Fediverse than Bluesky or Nostr. It’s quite possible that the latter, more modern protocols already support everything I want to see built, or that ActivityPub never will, or that these ideas are simply unimplementable on all three.

Stil needed:

  • reality check from bluseky and Nostr perspectives
  • a better visualisation of how this would look

Revision Notes

This is one of this wiki's pages managed with the permanent versions pattern described in  Two wiki authors and a blogger walk into a bar…