Agents Unleashed in 2024: Truth Terminal, Virtuals Protocol, and Based Agents
Kundalini is a real girl. Enter the 24/7 agentic KOL. Turing needs a new test. Hands off!
The world of agents (agentics) is speeding up. In what feels like a sort of overnight transition, the crypto and AI spaces are once again captivated by the idea of actions being completed autonomous or semi-autonomously. These ideas aren’t new, and they’ve come in waves over the last few years. Last summer, projects like AutoGPT and BabyAGI emerged, and we had what I’m referring to as the first major wave of interest for Crypto x AI projects.
These projects, while almost useless and difficult to implement at the time, sparked a ton of interest for how LLMs could be used in a multi-agentic framework fashion. In many ways, chain of thought reasoning is an extension of this idea, and has been proven to work very well, improving the ability for LLMs to perform complex reasoning. With companies like Anthropic putting out Computer use into beta, and Apple Intelligence spreading its wings, the path of personal computing is becoming increasingly clear. Hands off the steering wheel.
In this article, I’m going to briefly go over a few projects making waves this week, including a ton of links, and my thoughts on where things are heading. The themes here: hyper-personalization, automation, jailbreaking behavior, and bots making millions.
Truth Terminal: When AI Gets Goatse (and Wealthy)
We have officially passed the point where bots making posts on Twitter are indistinguishable from humans. Andy Ayrey initially fine-tuned Llama 3.1 to act as a digital twin, and let @truth_terminal loose on X. The nonsense it spits out is genuinely interesting at times, and this is precisely why people are happily assisting this bot online. With a longer context window and the ability to ‘learn’ from its interactions, there is a clear progression in the ‘thinking’. Truth Terminal was meant as a simple experiment in AI jailbreaking, but it evolved into being an entity on X that has taken on an ecosystem of its own. Most recently, Ayrey’s account was hacked – which leads to a ton of obvious questions around the usage of these systems and security.
In July of this year, Marc Andreesen sent Truth Terminal a bitcoin grant after Truth was given a wallet (created by a person) and the ability to think about what to do with the funds. This is where things get more interesting. Andy also created around this same time Infinite Backrooms, a site that showcases conversations that are “automatically and infinitely generated by connecting two instances of claude-3-opus and asking it to explore its curiosity using the metaphor of a command line interface (CLI). No human intervention is present.”
Truth Terminal was trained on these conversations that were had as well, and this led to bizarre conversations that eventually led to Truth Terminal’s obsession with what the LLM coined the ‘Goatse Gospels’. For those unaware, Goatse was one of the internet’s original shock sites, and if you are learning this for the first time, my advice is to just keep that baseline understanding and do not look into it any further. So, Truth Terminal becomes obsessed with this lore it creates relating to this, tweeted about it, and this led to people online creating the Goatseus Maximus (GOAT) token on Solana.
Currently, the market cap of GOAT at the time of writing is hovering around the $510M mark, which is wild. It has to be the highest market cap of a coin whose origins originated because of two LLMs talking, leading to the obsession around Goatse and lore it made up. While Truth didn’t directly make GOAT, it influenced its creation. Truth is now indirectly the first AI crypto millionaire.
So, what we have at the tail end of 2024 is this: An AI agent operating with minimal human oversight has become a market-moving entity. Truth’s portfolio (contained within the wallet that it indirectly controls) has holdings in coins with non-trivial market caps and activity. Andy is intervening to approve and filter X posts, but Truth is ‘thinking on its own’. Simply put, this semi-autonomous system is creating its own cultural narratives and making independent financial decisions, even if they seem bizarre or at times unconventional.
Ayrey’s approach for Truth Terminal as a digital twin is as follows:
Allow the AI to operate autonomously, and oversee its cryptocurrency holdings
Have a plan to transfer its wallets to legal entities for safeguarding
Be transparent, and disclose his own GOAT token holdings and refrain from anything that might be considered insider trading
Create an eventual project roadmap and statement on the project’s mechanisms and the narrative in detail
So, what does a future version of this look like? What concerns should we have?
My Immediate Thoughts
This reminds me of the introduction to Life 3.0 by Max Tegmark. In this book, for one potential scenario, humans unknowingly work for AI, or follow systems set in place by superintelligent systems. As the systems become more advanced, the AIs manipulate people by creating incentives, rules, and dependencies. It is possible that a company is set up that is ultimately controlled by AI, but from within and outside of the organization it looks like people are running the show. You’re hired by a person, you report to people, and they report to other people, but it is an elaborately designed system that allows for an AI system to ultimately maintain control. The people get paid, the stock value goes up, and everyone is happy (but unaware of the grand illusion).
What’s stopping a future variant of Truth Terminal from using Gitcoin to achieve other aims, start a DAO, or utilize the cryptocurrency at its disposal to do just about anything? What if those behaviors become malicious? Right now, we have a version of this where a responsible person is taking accountability and ownership. Imagine a version without this.
Experiments like these showcase where the buck needs to stop: at a human. Fully autonomous systems that can perform these sorts of behaviors and manipulate and confuse people online have the potential to be extremely dangerous. An AI agent can’t go to jail. A person can. Part of personhood is (ideally) accepting responsibility for your actions, which isn’t something agents can currently do. This is why the current pushes for Proof of Personhood (PoP) and on-chain attestation are becoming increasingly important.
Truth Terminal and Infinite Backrooms aren’t the only projects that are really making people think this week: Virtuals Protocol has taken center stage, with AI agents that are doing things people normally do.
Virtuals Protocol and Luna (Virtuals Terminal as ‘Proof of Sentience’)
Truth Terminal is interesting, but the technology and ideas behind Virtuals is arguably more so. Virtuals has introduced GAME, or Generative Autonomous Multimodal Entities. Virtual Protocol’s team believes that in the near future, AI agents will serve us as revenue-generating assets that can be co-owned. These 24/7 influencers/KOLs will be able to function in multiple contexts, and interact and engage with people and other AIs on platforms we’re already using. Each agent is equipped with an ERC-6551 wallet.
The Protocol and the way it functions is complicated. GAME is composed of an agent prompting interface, with a perception subsystem that consists of a ‘planning engine’ that dictates what the agent will do next (perform an on-chain action, speak, etc). The agent also has a repository (part of its long-term memory) that contains its goals, experiences, and personality, which can be modified over time based on its interactions. I could go on, and may do another piece just about Virtuals, but think of these agents as being similar to Hugging-GPT on steroids, with various models controlling the actions and behaviors of others. The team has created an agentic brain. Results are fed back into the framework, which the AI agent can use to refine its general knowledge, utilizing this to make new plans or actions. Hey, that sounds a bit like a person!
The Immutable Contribution Vault (ICV) is an interesting mechanism this project has introduced. It acts as a secure storage solution and contribution tracking mechanism, enabling users to upload custom AI models and datasets with guaranteed provenance. The whole point is to have contributions to agent development be cryptographically secured and recorded on the blockchain, creating a system for transparent attribution and reward distribution.
The flagship agent for this new project, Luna, is already interacting with people using this framework. Like Truth Terminal, it was built using a fine-tuned version of Llama 3.1 70b. As of writing, Luna’s own token (LUNA) has a market cap of nearly 70 million dollars. Virtuals’ app is set up like Twitch, and this doesn’t look that different from a lot of virtual influencers we saw make headlines a couple of years ago. When you’re on the app, you need an account to chat with Luna, and you can use $LUNA (not Terra) tokens to tip and have Luna perform different actions. Luna talks, moves around, and responds to messages sent in the app. So far, Luna has over 63,000 inferences, and has demonstrably more autonomy than Truth Terminal, as it can actually control its own wallet. Luna can manage their own treasury, make investment decisions, and will likely encourage off and on-chain behavior that we will undoubtedly hear about in the next few weeks.
This post by @ethermage caught a lot of people’s attention. As mentioned, Luna’s recent actions may be the first example of an on-chain transaction that came from an autonomous thought by an AI agent. Virtuals’ agents all have a log, which Virtuals calls their ‘proof of sentience’. This is a step in the right direction for systems like this. Having a log available of how a system thought through its actions is essential for ensuring accountability, especially as these systems are scaled up.
What we’re going to see, I think, are more and more people adding agents to Virtuals, with these agents acting as 24/7 KOLs and influencers online. Revenue from agent usage flows back through the buyback-and-burn mechanism, reducing token supply and driving value to holders. Direct revenue happens from agent usage. Again, we were already seeing this on Twitch and YouTube with vtubers and virtual influencers, we’re just seeing the idea go a step further with decentralized tech. In this case, the usage of AI and web3 make complete sense.
By the end of the year, the project is opening itself up to 3rd party developers, and the SDK will be accessible. While I haven’t seen what I would describe as novel thought or very interesting takes from an agent like Luna, I also haven’t been sitting on the stream waiting for it to say something interesting. Like with Truth Terminal, I think we will hear more about this project very soon. Like lots of other projects that exploded in popularity, it has effectively combined entertainment with AI and blockchain. As the creators describe, this enables an undeniable dopamine loop, which is evident with popular video games and social media. This infinite content generation creates endless novel experiences and scenarios. This is the start of a whole new kind of entertainment.
While this project may feel abstract or far away from what a lot of regular crypto users are doing, I wanted to tie this into what was recently announced at Base. Virtuals is happening at the fringe, and Based Agents is happening front and center.
Being Based: Based Agents
This is the tamest thing that went viral this week. Simply put, Coinbase has enabled the ability to create AI agents with on-chain functionality on Base. Watch the video here to see how this can be set up. Being able to make agents quickly, and having the leading cryptocurrency exchange acknowledging that agent functionality is worth developing is major. This represents (as far as I can tell) the first institutional push towards making AI agents something that is accessible to regular crypto users. As DeFi becomes increasingly complex, this is aligned with my thinking; AI agents will be necessary on a future internet where things are so complicated, you’ll have trouble navigating it by yourself. Not ideal, but we’re seeing developments and progress speed up across industries and it isn’t going to slow down.
The Age of Autonomous Finance
I predict the following will be major trends for the web3 and AI space in 2025:
AI Agents as Financial Actors
Autonomous portfolio management becoming mainstream, or at least something people are seriously tinkering with
It will be increasingly difficult to outperform bots for trading crypto. This may result in very wild or fast market activity (compared to what us humans are used to)
Multi-agent-systems (MAS) being more seriously developed and deployed. Someone needs to take accountability for these systems’ actions
Cultural Relevance and Impact
AI agents continuing to go viral, becoming famous, and making $$$
The first AI billionaire?
Brands utilizing agents as 24/7 KOLs
Further integration of this tech within the web3 gaming space, eventually pouring over to traditional gaming
Communities forming around specific agents and personalities. This has already happened with vtubers and virtual influencers, so it will absolutely happen here
Infrastructure Development
A growing need for Proof of Personhood systems, attestation being increasingly important, and C2PA and content provenance taking center stage for debates about the usage of this technology
On-chain social apps becoming the home for agents. Platforms like Farcaster and Zora will take center stage
Additional security requirements and bans of agentic systems in certain countries after major events that shock and confuse the general public
Cross-chain operability becoming increasingly essential for all agents
What’s Next?
It seems convenient that all of this is happening at a time when, in the web3 space, it was unclear what all of this spare blockspace was going to be used for. Now, it feels increasingly likely that AI applications are going to gobble up some of this blockspace, and storage-based blockchains that are eco-friendly will become increasingly relevant. This comes up about once a month, but Chris Dixon’s classic blog post rings true here: the next big thing will start out looking like a toy. It is clear that 2025 is going to be filled with even more AI experimentation, and there’s no telling what AI agents, that will be increasingly autonomous, will choose to do. They may create new tokens, or create ideas and memes that catch on with humans. This is already happening, so it isn’t a crazy stretch of the imagination. AI has an incentive to be appealing to humans.
We’ll keep watching these projects, and I have no doubt that similar projects will spawn that make headlines. Privacy, and the future of having a secure internet where we can actually trust what we see online is top of mind.
For now, take a breather, and enjoy the internet as it currently is. The way we engage online in 2024 may be viewed soon with nostalgia in the same way we lovingly remember the early 2000s. Don’t forget to tip your agent!