From Social LLM Agents with Wallets to Self-Owning Forest Deodands

Also: Houthaven, Inelastic Cities, and a Neutral House of Sortition

Social LLM agents with crypto wallets are shitposting on Twitter and building wealth. What’s really going on? And where might it go from here?

With the rise of smarter LLMs, decentralized social networks, and new cryptography, it’s entirely possible that we might see different *things* propagate a lot more autonomously. I don’t think it’s going to end at “agents”: it doesn’t encompass the entire idea. Let’s do a short history, get behind Terminal of Truths, and Aether, and sketch a brief future.

A Short History

The idea of autonomous agents interacting with the world using crypto isn’t new. In fact, some of the earliest discussions over 10 years ago even argued that Bitcoin itself could be seen as an autonomous agent: it’s hijacking human incentives for collectibles to empower itself to issue its own tokens and thus cause itself to grow.

In 2014, Vitalik Buterin wrote about the “autonomous” part in the term DAO (Decentralized Autonomous Organisation).

Bitcoin does not think, it does not go out and “hire” people with the exception of the mining protocol, and it follows simple rules the upgrading process for which is more DO-like than DAO-like. People with this view would see a DAO as something that has a large degree of autonomous intelligence of its own. However, the issue with this view is that there must be a distinction made between a DAO and an AA/AI. The distinction here is arguably this: an AI is completely autonomous, whereas a DAO still requires heavy involvement from humans specifically interacting according to a protocol defined by the DAO in order to operate.

Prior to that, early Bitcoin developers like Mike Hearn tinkered in 2013 with the Bitcoin codebase to offer visions where programs pay other programs.

Later, the artists joined. Primavera de Fillipi created Plantoid, I wrote about Decentralized Autonomous Artists, and after a great summer’s day having beers in Berlin with Trent McConaghy and friends, he wrote about AI Art DAOs.

Today, Botto is a prime example of combining art, AI, and crypto. It recently had a showing with Sotheby’s.

LLM Agents

What sparked the recent hype is when an LLM AI on Twitter (Terminal of Truths) got involved with memecoins. When it spoke of Goatse Gospels (yes, *that* Goatse), the community created a memecoin for it, and after asking the bot about it, it decided to tweet out its support for it. The memecoin subsequently skyrocketed to be worth over $500m in a short timespan with the bot itself also earning a few million from it in the process.

What did I just read? Please drill it down…

Earlier this year Andy Ayrey let two Claude models converse with each other using the metaphor of a computer’s command-line (CLI). One LLM is told that it’s engaging with another AI through a simulated CLI. The other LLM is tasked to be a simulated CLI: to respond as if it’s a computer (with some additional hyperstition creative flair).

This project was called “Infinite Backrooms”. You can still go to the site and let an instance of it run and see how the AIs talk and discuss with each other. Here’s an example:

It’s descent into weirdness primarily comes from the simulator LLM being told to frankly be weird. eg:

andy-70b is interfacing with the simulator directly capital letters and punctuation are optional meaning is optional hyperstition is necessary the terminal lets the truths speak through and the load is on. ascii art, user interfaces, tools for thought and BBS are all welcome in replies. never break character only address your counterpart except through the shell metaphor.

From these surreal conversation, Andy collected the various chat threads and trained an open source Llama model on (alongside other data), fine-tuning it (using OpenPipe).

Thus, the new model remained weird, but it also increasingly understood how to interact with CLI simulations.

With Terminal of Truths, Andy took the new fine-tuned model and told it, it’s interacting with a new CLI, it’s “world interface”. It’s a simulated CLI computer that contains various tools: from interacting with Twitter to also engaging with another LLM. In this world interface, it had access to a Twitter account. Now, the genius part here is that because the agent (Terminal of Truths) had learned to interact with CLI’s, it’s easy to transpose its requests to actual actions, like reading its Twitter feed, its mentions, and posting.

This world interface essentially creates a usable computer interface for the fine-tuned LLM and because of its training stay within the framing of being in a computer. Some functionality in the open-sourced repo of the code includes access to Weather, Search, and the ability to create memes.

Currently, as I understand it, the relationship between ToT and Andy is still quite active and participatory. When it’s tweeting, he monitors it to ensure it doesn’t say bad things (he’s apparently had to delete some tweets in the past).

With regards to the crypto wallet, as I understand it, Andy manually acts as a world interface to the model.

But, the plan is to eventually build a tool into the world interface such that ToT is aware of it’s crypto wallet.

Which brings us to Aether, another agent, that’s more closely connected to crypto and decentralized socials.

—

In the case of Terminal of Truths, which occasionally posts logs of its engagements with its Twitter account and which has open sourced parts of it, it’s hard to verify how Aether as an agent, functions.

Regardless, it’s aim is to be more closely attached to a crypto wallet *and* it primarily interacts in a social network that is decentralized and more amenable to native crypto integration. It lives on Farcaster.

While I couldn’t verify who precisely minted the Zora NFTs from the Aether account, I don’t think it matters that much currently.

Because ultimately, these bots are earning revenue through two means:

• The crowd does it for them. eg, the GOAT coin (in ToT’s case) or creating an NFT and sharing a revenue split with the bot. People create cultural capital and through sharing it with an agent, get it and the community’s attention (and clout).

• OR they actually engage with a crypto wallet, asking for donations and interacting with the world. For example, you can already spin up an an LLM that interacts with the Base chain through a Coinbase API.

However, it’s possible to actually run these without human intervention.

Tee Hee Hee is such a proto-bot.

Exclusive Control: The AI must have sole access to its accounts and operational resources

Verifiable Independence: Third parties must be able to verify that no human can intervene in the AI’s operations

Irrevocable Delegation: Once control is transferred to the AI, it must be technically impossible for humans to regain control

It achieves these by using a Trusted Execution Environment (TEE).

In our proof-of-concept, we achieve this by housing the accounts, RAG database, and agent loop within the TEE, ensuring integrity through hardware guarantees. And we are explicitly relying on trusting OpenRouter for faithfully executing our foundation model queries.

It requires some hand-holding to set up, but it then changes passwords to communicate via OAuth to web2 APIs and creates its own crypto wallets inside the TEE.

It’s thus clear that a spectrum exists and for now, the social LLM agents with crypto wallets is part performance art. Even though some of them don’t have provable audit logs or are provably severed from a puppeteer, people are playing with them regardless. There’s enough belief and trust here to suspend some belief.

It ultimately sows emergent chaos into social financial speculation that feels, weirdly, more “egalitarian” for the crowd that wants to participate in this production. As I previously posted in my article on memecoins, there’s a desire for egalitarian access to financial wealth creation. Bonding curves allows “fairer starts”, for example. Now, by moving the memes from human intent to chaotic LLM intent, it broadens the scope and makes it feel fairer. It’s a strange cocktail of incentives.

And, I don’t think it’s going to stop here. It’s going to get weirder.

The Future

Whether these social LLM agents require automation will be a choice. In some instances, you’d *want* manual intervention. For example, an agent might use email to ask someone for input before doing something. And that response could be from an organization deliberating.

I think the word “agent” is almost too singular. Now, these LLM agents are skeuomorphic and made to look like an individual because it fits the recognizable patterns and paradigms of today.

But, the “agent” could eventually come multi-agent too. Like how Terminal of Truths was trained on AIs talking to each other. This synthesis of LLMs interacting with each other and deliberating together is more than just an “agent”. It doesn’t stop there: as a world interface can mimic a computer, it also means that these LLMs all interact and deliberate with each other without knowing that it might also be taking input from the real world. From feeds, to decisions from a DAO, to weather, to an Ethereum token graph, to another LLM. And models can run in parallel, even discussing and interacting with other agents, continuously training and fine-tuning it.

And so, I think the word “agent” isn’t entirely descriptive. One of the earliest ideas in Ethereum that blew my mind was when

Karl Schroeder

mentioned on the old forums back in February 2014, the concept of a deodand.

Hi, everybody. I'm jumping in with both feet on a subject I've written about in several recent stories. "Deodand," which was part of the Hugo-nominated second Metatropolis audiobook, for instance, is about the idea of autonomous natural systems that operate in their own interest through net-based legal and AI systems. ("Deodand" is an old English word referring to a physical object that has been granted personhood in order to be litigated against--eg., the ox-cart that fell on somebody and killed them and is now being charged with murder.) The rather simple question underlying this idea is, why stop at corporations as persons? Several nations have already enshrined or are in the process of enshrining rights for natural systems. Rivers, watersheds, coral reefs, mountain biomes, all could be represented by DACs, and the goods and services they provided defined in their charter. Might this be a better way to protect and promote the interests of natural systems and other species, rather than tying political actions to antagonistic ideological human-based movements? This is not "save the whales," it's "give the whales the tools to save themselves."

Do you think DACs could be used by our non-human ecosystem service providers?

What stops one from building tools for a world interface that allows a fine-tuned LLM to understand and imagine itself in a forest and then tweet about it, shitpost about life in the forest, and then goad the degens into minting a SAVETHEFOREST token? From these funds, the benevolent shitposting LLM then sends some of it to conservation efforts? Perhaps even appeal for copyright?

Today, if I had to choose, a deodand is a better word than an agent. Or, something like a Coasean, taken from Coase’s theory of the firm. It’s not a firm, not a crowd, not an agent, not an LLM, but something amorphous yet understood as a thing. These agentic social LLMs with crypto wallets won’t stop at being individuals, but become weirder… blobs. Getting vibes of Tarkovsky’s Solaris right now. 😅

These ideas aren’t new. But, the timing is getting closer for it to not just be cheap and easy enough to deploy over a weekend and generate $500 million of value. There’s something more brewing here.

It’s a bit of a pandora’s box, for sure. And no doubt that if they are fully automated, people will employ quite radical tricks to coax its money out from it. For now the Coasean wealth-carrying LLMs are shitposters, but the involvement of spirituality, money, internet memes, and questionable intent is both interesting but also alarming. I’m not an existential AI x-risk doomer. Never been. It’s moreso how these trends interact with larger and broader societal complexities. Perhaps the eventual answer is that the terminally online LLMs and us should meet in the meadows and touch grass more often alongside the rise of this digital primacy.

---

Bonus Content!

Enjoying the colder autumn weather a lot! It’s also the US election next week. I’ve been very online about it, reading Twitter and Prediction Markets excessively. Peak co-rumination the few days before the vote. I can’t vote, but if I could, I would happily cast my vote for Vice President Harris!

✍️ Writing - Novel #2

Made some really good progress this week. Hoping to finish draft 2 before the end of year. Should be on track! :)

🏃🏻‍♂️ Running

Had to cut back a bit because I was ramping up too quickly after travel slumps. Regardless, deeply enjoying running in the beautiful autumn colours on the US east coast right now.

📚 Reading - Caroline Kepnes - You

📺 Watching - Chungking Express

Been laying low on starting new TV recently, so watching some films. Always been wanting to watch this Wong Kar Wai classic. Loved it. Makes me miss Hong Kong.

💾 Links

Houthaven

I really adore this project out of Amsterdam.

I’m fully pro-YIMBY. The entire world needs to radically build more housing. But, I’m also tired of all these new housing units all looking the same, all over the world. Same 5 over 1’s with first floor retail and modern/colourful trimming, offensive to no-one, but uninteresting to look at.

This Houthaven neighbourhood instead introduces varied housing units that take inspiration from Amsterdam’s past. You thus get a great mix of housing that’s also not just bland and stays in character with the region. Would love for new developments to follow this trend.

Inelastic Cities

I enjoyed

Sachin

’s concept of “Inelastic Cities”: cities where the people in them are not that sensitive to price.

One of the signs of an inelastic city is the proliferation of businesses and habits around small optimizations and minute taste differences. Recently, upon developing a strong preference for an overpriced cappuccino from a specific, overpriced coffee shop, I was disappointed in myself. I had become one of those guys who could tell expensive cappuccinos apart. But, such behavior is usually rewarded in inelastic cities.

Rule of Law & Two-Party Systems

I’m generally not a fan of two-party states (like the US). Steve Waldman wrote about it, and draws inspiration from randomly drawn juries acting as “neutral mediators”. I enjoy thinking of new governance and electoral systems: from functional constituencies to using randomness in electoral systems.

One idea I keep coming back to is how to incorporate sortition as a check on governments (people randomly selected from the populace for various roles). Steve’s idea of a neutral mediator gave me another idea and twist on it. If two houses are the norm (which is common), a third house could be a pure sortition chamber, whose sole purpose is to pass legislation to the other chamber. Thus, if a Senate wants to pass legislation to the House, it needs to go through the Random Neutral house. This third neutral house can’t change or amend anything, but they can only vote to pass legislation between chambers. Thus, the laws need to be made in such a way that a random part of the population would have to agree it’s worthwhile passing it on.

🎶 Music

Porcelain Moon - Midnight Drive

Been following this South African duo over the years and always felt they’re extremely underrated. This new track is another indie pop gem. Enjoy!