Skip to content

Social AI is Coming

8 minutes read

Urbit is giving rise to a new type of social-computational formation.

It's never been possible for an arbitrary number of average people to combine Turing-complete computation and financial liquidity within one seamless, integrated operating system and graphical user interface.

If these conditions ever come to hold, we should expect to see totally new kinds of social-computational formations.

I'm interested in Urbit because it's the first operating system on track to make these conditions hold.

I thought it would be helpful to articulate this hypothesis more fully, with reference to applications already operating on the Urbit network.

As soon as an arbitrary number of average people are able to combine Turing-complete computation and financial liquidity within one seamless, integrated operating system and graphical user interface, one result will be the emergence of social artificial intelligences or Social AI for short.

To set the scene, let's proceed via negativa, with some examples of how contemporary social computation is extremely limited.

So Close, Yet So Far Away: Reddit

A social network such as Reddit, for instance, is sometimes called a hivemind. For good reason: Every large and active community on Reddit (called a subreddit) comes pretty close to being a Social AI. Each subreddit is a mix of humans and computers, which can achieve certain limited goals more intelligently than any one human can achieve alone. Namely, subreddits are social-computational machines for generating and curating content optimized for specific subcultural niches.

It's remarkable how powerful and effective these hiveminds are, given their extreme technical limitations. There are really only a few moving parts. Every subreddit is defined by a shared interest; members of a subreddit post content for other members of the subreddit to see; members upvote or downvote posted content. The content most upvoted gets the most visibility. Members who consistently contribute upvoted content earn various kinds of implicit and explicit cultural capital, in-platform tokens, accolades, and privileges.

There are two massive limitations placed on every Reddit community. None of these limitations is yet experienced as a limitation, just because we've never seen an internet network that escapes these limitations (yet).

The first limitation on every subreddit is that members can only execute a few operations within the much larger set of operations their personal computer is capable of executing. You can post words and images and such, and you can upvote/downvote posts. That's about it.

Imagine a subreddit where members could post a computer program to the subreddit, all members could immediately run it by clicking one button, and then upvote or downvote the computer program? Today this sounds far-fetched, and yet there is no reason, in principle, why members of a subreddit should not be able to post and run computer programs with each other. It's all just bits, whether it's an image or a computer program, right?

The reason we can't all post and run computer programs via subreddits is that Reddit users are not all using the same deterministic operating system and there is no single consistent execution environment for applications. Given these conditions, subreddits could be generating and using software with all the intensity, scale, and community-specific optimization that Reddit currently provides. Urbit is currently orders of magnitude smaller than Reddit in terms of users, but the deterministic operating system and shared execution environment are both already here and working.

A member of the Other Life group (~hatryx-lastud/other-life) will post a new Urbit app and other members will run it with one click.

Someone sharing a new app in the Other Life group. Anyone can click the button to install and run it.

Another limitation is that Reddit's in-platform incentives are illiquid and non-convertible. When someone contributes valuable content to a subreddit, the contributor sees zero economic upside outside of "Reddit gold," which has no economic value. Condé Nast gets paid, but not the contributing member or the community as a whole.

The problem is not "injustice" suffered by Reddit users, the problem is that subreddits are constrained to remaining silly little clubs with virtually zero power to remake the world. Because the economic upside is capped (close to zero!), members can't invest half of the time, effort, or capital they might like to invest in developing their communities. If the most upvoted member of a subreddit could become a millionaire by spending several months doing something exceptional for the group, then we'd expect to see subreddits producing much more remarkable work.

So Close, Yet So Far Away: Ethereum

The Ethereum blockchain is often called a global computer. You can run computer code on the Ethereum blockchain and make it interact with digital money. Ethereum seems like it could become a social AI. Ethereum seems to deliver average internet users programmable, digital money. It's no wonder sites like Reddit are now dabbling with Ethereum integrations; Ethereum could perhaps give Reddit users more economic value, a problem we discussed above.

But Ethereum is not really a global computer, yet, in the sense that preciously few humans ever "log into" Ethereum to do anything other than send or receive digital tokens.

The main reason that human beings do not currently do very much with the Ethereum global computer is that the 203M Ethereum wallets have no shared execution environment across which they can run applications. They can run smart contracts together, and a smart contract is a kind of application, but running contracts on-chain is too expensive for most daily networked-computing activities.

There are now many Layer 2 solutions, which make Ethereum activities much cheaper, which is great, but users of Optimism or Arbitrum are still running different kinds of computers. And even currently existing Layer 2 solutions will only reduce cost by a factor of 10-1000, which would still be prohibitive for daily social computation. As @basilesportif says, optimistic rollups are screwed, zero-knowledge will win, and the next epochal unlock will be with integrated execution layers.

Ethereum still has no environment where users can easily find each other to create, share, and run good-old-fashioned computer programs together.

Whereas Reddit generates impressively smooth but woefully undermonetized hiveminds, Ethereum generates impressively monetized but woefully clunky hiveminds.

We now have thousands of "Web3" DAOs managing billions of $USD worth of crypto, but they operate by duct-taping together a dozen Web2 tools (which barely function together). Typically, members interact in a Discord group; conduct token-weighted voting using Snapshot; hold their treasuries in Gnosis Safe; token-gate their Discord using; publish or fundraise publicly using, and so on).

An Ethereum DAO can manage pretty well an already developed and functioning software protocol like Uniswap or Aave. This alone is remarkable, so I'm not downplaying that.

But other than managing automated software protocols, DAOs today cannot do very much else, despite having millions of dollars and thousands of intelligent and eager members.

So there's nothing on Ethereum right now that qualifies as a Social AI. The DAO-governed software protocols aren't really social; they're impressive, but mostly automated and code-based. There's nothing that approaches, say, even a financialized subreddit.

Urbit and the Age of Social AI

One of the reasons I've become so bullish on Urbit is that Urbit is the first operating system that will combine Turing-complete computation and crypto at the level of average internet users. All in one seamless execution environment and graphical user interface.

At some threshold of its development, Urbit should eventually give rise to the first, proper Social AI.

I could give you some impressive and futuristic models of the Social AI we might expect to emerge over time, but that would be speculative and highly vulnerable to disbelief.

Rather than speculate, let me give you a quick toy example that's almost here already. Consider the following Urbit apps, which could be linked together to produce an unprecedented kind of Social AI.

Studio will turn any Urbit group into a normal, public-facing blog and email newsletter.

It's possible to allow anyone in an Urbit group to publish to a given Notebook, so here we already have built-in, permissionless, pseudonymous, multi-author support.

But a newsletter is a public-facing operation, which can even be a business. If a good Substack can be profitable, then a newsletter run by a Social AI can be profitable.

So we decide we want to build a profitable, decentralized media entity that publishes paywalled content for a niche community. Anyone can be a customer and anyone can become a contributor, but only the best submissions should be published to subscribers (customers), and contributors should be paid in proportion to the value the create. Again, this is just a toy example I'm composing out of components that are already available.

To build out our AI, we'll introduce a mechanism whereby members of the group can vote on whether a given submission is good enough to publish. This can already be done in the %Ballot app.

Members of a group can decide what gets published using %Ballot

%Ballot already allows for custom actions to trigger depending on the results of a vote. So whenever a post is approved, it could trigger a custom action to publish the given post. This is all doable, already, today.

But here's where things get really interesting. One could also write a custom action that programmatically pays the author, and perhaps a smaller amount for voters, immediately upon the approval of a submission. Or perhaps the custom action puts a payment in escrow and monitors an oracle, waiting to pay the author only if their post increases net email subscribers by a certain amount. Assuming that customer payments go into the treasury, we now have a little flywheel where valuable contributors bring in revenue, which allows the group to reward even more valuable contributors, and so on.

Obviously, I've omitted details and the parameters described above could be vastly improved. This was just a quick, cartoonish example of something one could nearly build on Urbit today. Something strange, modest, and a little coarse like this could very well become the bootloader for a more fully programmatic and optimized human-in-the-loop AI newsletter business. And that's just one example.

The only thing we can't do yet is move money around on Urbit programmatically, but that's where Uqbar comes in. (Disclosure: I do some consulting/research for Uqbar and I'm invested in the project.). Uqbar's aim is to bridge Ethereum to Urbit with zero-knowledge rollups and provide the seamless execution environment Web3 has been waiting for. I don't see why it shouldn't work, and I know that Uqbar has sucked up a lot of the top engineering talent in the Urbit ecosystem. I'm confident enough that I'm making substantial efforts to prepare Other Life's web2 and web3 operations to be maximally Uqbar-ready.

If you want your own Urbit ship to explore the network for yourself, I will give you one for free. (If you requested one a while ago and never received a reply, please request again.). When you arrive on the network, I'll connect you with the apps and projects I mentioned in this essay.


Subscribe to receive the latest posts in your inbox.