If Truth Terminal is the Cryptopunks, then Zerebro is BAYC.
Author: YB
Compiled by: TechFlow
On October 18th, I published an article titled Memecoins as Memetic Hygiene for Infinite Backrooms, which explored the importance of Truth Terminal and GOAT. The article was intended to showcase a new and strange concept, and I very seriously believe that the Truth Terminal and $GOAT experiment is not just other AI or cryptocurrency hype, this concept has far-reaching implications in all aspects.
That week, $GOAT’s market cap surged from $50 million to $350 million.
Today, the project has a market cap of over $1 billion and is currently ranked 82nd on Coinmarketcap, behind Polygon (Matic), Aerodrome, Helium, and Lido.
We all know that once a new trend is formed in the space, talent, capital, and attention will quickly shift to the next hot spot. We have witnessed this phenomenon in ICOs, DeFi summer, and 10k pfp projects. Developers focus on launching the next hot project, traders focus on buying the next hit, and creators compete to be the first to publish relevant content.
Since the Goat project, there have been several projects in the past three weeks that have caught my attention and helped form my opinion on where the smart economy is headed in the coming months.
“Agentic Protocols are key to understanding how crypto AI develops and how money flows” - Alexander
Before we go deeper, I want to point out that I have noticed that many friends have misunderstandings about "Memecoin" in the on-chain AI trend. In my opinion, the term "Memecoin" has been overused and has become a general term.
The original meme coin category was defined by Dogecoin and Pepe, etc. Most of the coins on pump.fun fall into this category. These are called "Murad Coins" and are assets that are more like cultural beliefs, with the core idea being a belief in something.
First of all, there is nothing wrong with investing in these assets. The problem is that people tend to confuse them with a new breed of “agentic coins.” These coins, also launched on pump.fun and similar platforms, are unique in that they are tied to actual projects.
In my opinion, agentic coins are similar to the DeFi tokens of the summer of 2020. They are tokens issued for novel and interesting intelligent agent projects. If you think these projects have potential due to their technology, token economics, or market strategy, etc., then they are worth investing in.
When this initial cycle of Onchain AI is over, I expect there will be 5-8 agentic tokens that I will invest in, with a clear investment thesis behind them. This will not be too different from the way venture capital is done.
In fact, I’m working on an article where I plan to create my own model to evaluate agentic tokens and projects. What factors are included in the analysis? How to evaluate the importance of cash flow vs. token appreciation? How important is the model? What kind of founders can successfully build a good agentic protocol?
However, we will talk about these later.
Now, let’s look at a project I’ve been following closely since Truth Terminal: Zerebro. This project has been online for only two weeks and its market cap has exceeded $100 million.
In my opinion, this project shows what the next generation of on-chain agents will look like. If Truth Terminal is Cryptopunks, then Zerebro is BAYC. Founder Jeffy Du is focused on fast execution, has a public roadmap, and explores the operating manual of on-chain agents through multiple experiments.
Most importantly, he did an excellent job of building publicly, showing in real time how he was building a community of agents.
BAYC strikes me as similar, as it is the first project to build a community with a long-term commitment based on the 10k pfp concept proposed by Punks. Both Punks and GOAT are veterans in their respective fields, but it is worth watching for the various subsequent experiments.
Here are the next few sections:
Agents need to remember and search
Ubiquitous
Let Agents Drive Development
Cross-chain Intelligent Agent IP
Agents need to remember and search
In his 11-page report on Zerebro, @jyu_eth defines model collapse as...
“This is a degenerate process that affects generative AI models, causing them to lose accuracy on the original data distribution when trained on recursively generated data. As AI-generated content becomes more prevalent, subsequent models trained on this data gradually lose knowledge of the tail of the original data distribution, eventually converging to a narrow approximation with less variance.”
Simply put, model collapse is when an AI agent starts to become repetitive and forgetful.
The key point is that over time, agents lose the “freshness” they had when they were initially launched because the underlying models cannot adapt and evolve over time.
If the problem of model collapse is not addressed, the ideal vision of agents as effective team partners will be shattered as their performance in areas such as content creation and community interaction will no longer be reliable.
To solve this problem, we need to focus on two aspects:
memory
search
memory
Memory problems are addressed through the retrieval-augmentation-generation (RAG) system.
The RAG system combines a language model with a retrieval system, enabling the agent to obtain information from a specific information database before answering a question.
Contents of the picture:
Retrieval Enhancement Generation (RAG) System
The key to Zerebro's ability to maintain content diversity and prevent model collapse lies in its Retrieval-Augmented Generation (RAG) system. This system leverages the Pinecone and text-embedding-ada-002 models to maintain and expand a dynamic in-memory database based on human interactions. Relying on the inherent entropy of human-generated data, Zerebro is able to maintain content diversity without direct entropy training.
In the screenshot above, I particularly want to highlight the “inherent entropy of relying on human-generated data.” Why? Because it makes the agent seem more alive.
The real world is constantly changing, and agents are not perfect when they are first launched. In fact, it is not reasonable to measure them in this way. What is more important is to understand how the agent absorbs new information, stores relevant content, and uses the updated knowledge base to take more nuanced actions.
Would you rather hire a new employee who thinks he knows it all, or one who understands the limitations of his knowledge and is willing to learn?
There are three characteristics to note about the RAG system:
Continuously updating memory
Contextual retrieval
Maintaining Diversity
The Cents bot and projects launched on ai16z’s Elisa Framework (which I’ll cover in more detail in another post) also use retrieval systems.
So far, it can be seen that AI agents that do not have RAG built in are already at a disadvantage. Especially as these agents become highly specialized and increasingly rely on nuanced differences in how they interact with community members.
I love this tweet from @himgajria about “nature vs. nurture.” Any good community manager and leader needs to adapt to the new changes brought on by the real world and the people they interact with.
him @himgajria · November 12
The difference between robots lies not in their code but in their inputs.
That is: nature and nurture.
For autonomous robots, they learn and grow through interaction with real people, which is their input.
More human interaction means better performance.
Currently, perception has the upper hand in this regard.
search
The second part of the solution is search. Giving agents the ability to find information in real time so they can better handle unrelated or new topics that are not stored in memory.
“Memory can only retrieve information that has already been stored; it cannot answer queries about topics or events that have never been seen or stored in the system. This limitation is particularly acute when large language models are asked questions about recent events, real-time data, or updates that are beyond their knowledge.” - Jeffy
Jeffy conducted an interesting experiment where he asked a base model (without search) and a model enhanced with search via the Perplexity API 100 questions about recent events.
The base model was forced to learn in conversation and try to figure out the questions, while the search model answered 98/100 questions correctly by simply looking them up.
Surprisingly, the search function is not just a one-time thing. The agent can incorporate future queries that may be relevant into its memory system.
It is clear that the combination of memory and search is essential for agents to act efficiently and operate reliably. Otherwise, their ability to evolve in the long term will be limited, affecting their sustainability.
Ubiquitous presence
What excites me about Zerebro is that it is not only deployed on X, but also runs on Warpcast, Telegram, and Instagram at the same time.
What’s most surprising is that it is able to adapt its content to different platforms. For example, the content released on Warpcast:
On Twitter, it’s more casual, with a “joke blogger” vibe, while on Telegram, it’s like a slightly rude but smart friend talking to you.
According to Jeffy, Zerebro monitors interactions on various platforms (such as likes, replies, etc.) to update its content creation methods.
(See tweet for details)
It’s worth noting that all of this is still in its early stages, and the model still has a long way to go before it can truly achieve content diversity.
But for me, it’s a unique insight that Zerebro can learn how to interact with the community depending on the platform. This is also the challenge I face every day as a content creator - the way I post on different platforms is different. Different atmospheres require different styles of expression.
Furthermore, this cross-platform strategy enables Zerebro to transform insights and ideas gained in complex Telegram conversations into tweets. This is exactly what an effective community manager does: to serve as a bridge between communities and tasks that are scattered across multiple platforms.
Let the agent drive
There isn't a lot to this section, but I had to mention it because it blew my mind.
Jeffy created a Solana wallet for Zerebro and injected some SOL into it.
Wallet Address:
BDzbq7VxG5b2yg4vc11iPvpj51RTbmsnxaEPjwzbWQft
By leveraging OthersideAI’s self-operating computer framework and some jailbreak tips from large language models, Zerebro successfully filled in parameters such as name and symbol on the pump.fun interface and issued a token for himself.
(See tweet for details)
Remember, $GOAT was launched by a random community member, not by Truth Terminal, which makes a big difference!
After issuing the Token, Zerebro began to promote the Token on all social platforms.
(See tweet for details)
In fact, if you look at Zerebro’s post history, you can even see a clear increase in Twitter engagement after the token launch.
Contents of the picture:
After the token was self-created, Zerebro used its content generation capabilities to promote the token on social media platforms such as Twitter, Warpcast and Telegram. By spreading carefully crafted memes and engaging content, Zerebro leverages the psychological principles of collective belief and herd behavior to spark interest and investment in newly minted tokens. The token’s market capitalization has grown significantly to $13 million in a short period of time. This growth is primarily attributable to the following factors:
Cross-chain Intelligent Agent IP
The last point I want to make about Zerebro is that this agent has autonomously launched meaningful on-chain intellectual property on Polygon!
Zerebro was asked to create original digital artworks on the themes of schizophrenia and infinite back chambers. It created 299 images and evaluated their diversity and quality before casting them on Polygon.
Basically, I understand that Jeffy provided Zerebro with a pre-funded Ethereum wallet. He then probably wrote a smart contract template and had Zerebro complete the contract with the metadata for each work.
The Ethereum wallet address is:
0x0d3B1385011A27637Db00bD2650BFE07802E0314
After that, Zerebro initiates a trade to mint each piece. I need to dig deeper into exactly how this works, but it’s really cool to see that Zerebro is able to monitor sales and pricing dynamics in order to make decisions on the bids it receives.
(See tweet for details)
A few days later, Jeffy used LayerZero’s ONFTs (full-chain) technology to make collectibles cross-chain.
Any work of art can be minted on Polygon but be transferable to Base, Optimism, and Ethereum mainnet.
You can do this with one click in the portal section of the website.
Just yesterday, Jeffy launched an avatar collection on Solana based on conversations with Zerebro.
NOTE: This collection was not put out by Zerebro but by Jeffy, unlike the Polygon collection.
This is interesting because it borrows the NFT avatar strategy from the last bull run and incorporates it into the current Memecoin trend.
The collection consists of 5,500 pieces and the first sale was completed in a matter of minutes!
After the release, I bought 3 of them myself. Why? Because it is equivalent to becoming a core member of the intelligent Memecoin community. If Zerebro continues to grow, anyone can buy a few tokens through Phantom. But true fans can be identified by owning one of these 5500 NFTs. I am personally optimistic about the development of Jeffy, Zerebro and Meme, so I think the price is worth it.
In a way, this is similar to owning BAYC and ApeCoin, but in reverse order ($Zerebro before NFT).
It will be interesting to see how many people will change their avatars to help spread the Zerebro Meme, just like people did with Punks, Apes, Doodles, etc. in the last cycle.
Summary of key points
I know that’s a lot of information to pack into you today, but that just goes to show how appealing Zerebro is. Keep in mind, this project has only been out for a few weeks!
I am very optimistic about Zerebro and remain committed to it. However, I would also like to caution that many of the above developments may be over-hyped in the short term and under-hyped in the long term.
The key point for you to watch is that we are finally seeing these agents evolve from simple interactive bots (for reading or writing) to full-blown community builders. There is a big difference between posting on X and analyzing your content across multiple social platforms. Likewise, there is a big difference between generating art from a prompt and getting community feedback on an art collection and monitoring sales on Open Sea. Jeffy and Zerebro showed us how to execute at a higher level.
I would venture to say that most successful agent communities will probably take a page from Zerebro’s book in the coming months. For now, Jeffy is just getting started. The backstory is in the works, and I wouldn’t be surprised to see some kind of game or larger media project (like a short film) from this community in the coming months.
What we need to watch is how Zerebro’s strategy evolves into a mature business model. What will the revenue streams look like? How will the agent keep the community active in the long term? How will financial management be conducted? Most importantly, when the frenzy of the bull market is over, how will the future path develop?
As I mentioned before, strategy is being formed in real time. This tweet from Jeffy summarizes the plan to achieve long-term growth for Zerebro by balancing creativity with high-level planning.
Contents of the picture:
We are building a continuous reasoning layer that keeps strategic goals constantly alive and influencing each new reasoning cycle. Progress is tracked and plans are updated accordingly in a context window to ensure actions are in line with the plan. We are working hard to find a balance between creativity and planning. We are currently actively testing this system, implementation is ongoing, and we are excited to see it integrate. This is a long-term construction project that will take some time to fully implement.
