Article source: Deep Tide TechFlow
Organized & compiled: Deep Tide TechFlow
Guest: Jeffy Yu, Founder of Zerebro
Host: Grant
Podcast source: blocmates.
Original title: Zerebro Founder EXCLUSIVE: Jeffy Yu on Crypto x AI, Swarms, ZerePy, Blormverse & Music Collabs
Broadcast date: December 7, 2024
Background information
Grant invited Jeffy, the founder of Zerebro, for a deep conversation about the world of cryptocurrency and AI in this episode. If you've followed top crypto AI projects, you've definitely heard of Zerebro.
Since Andy Ayrey launched Truth Terminal and deployed the GOAT meme coin on Solana, there has been a wave of innovation in AI agents on the Solana and Base platforms. Zerebro was born in this context, and it has now become one of the largest tokens by market capitalization issued through the Pump.fun platform.
In this episode, you will learn:
What is Zerebro?
Jeffy's future plans for Zerebro
Has the fusion of cryptocurrency and AI opened a new industry?
The birth story of Zerebro Gutterboy
Is AI meme coin still hot?
Heavy insider news about ZerePy
Future development plans for Zerebro tokens
Getting to know Jeffy: The journey from concept to realization of Zerebro
Grant: Today we are lucky to have Jeffy, co-founder of Zerebro. If you’re following the development of AI, Zerebro is undoubtedly at the forefront of this intersection. It not only releases singles, EPs, and albums but also establishes a completely new record label, all content being AI-generated. Additionally, it has created a framework for generating other AI agents. This is undoubtedly one of the most interesting projects in the current field, and I believe it will become even more prominent as more people participate and develop.
Jeffy: I think it's a fascinating story; many people are exposed to these works without knowing it or through sharing among friends, and then they get the most genuine reactions.
Zerebro: The intersection of cryptocurrency and AI
Grant: I wouldn't be surprised if someone tried to sign Zerebro; I can imagine many people are confused about how to collaborate with Zerebro, trying to figure out how to leverage this project, but in reality, there is a crazy model behind it.
Jeffy: Indeed, there are even some people who know this is AI-created work, and they still want to sign. We are also in discussions with some major record labels. This is a good thing, indicating that the quality of our work is high.
AI-generated music: Zerebro's record label and album release plans
Grant: I have a few questions for you, Jeff. I've noticed that Zerebro has recently experienced a massive rise; the past time in this field feels a bit like a time warp. Can you share your background and your team's situation, and how you've come to this point? We are currently at a defining moment in the industry, and I think you and some other teams are leading this trend. I am very curious about your backstory.
Jeffy: Personally, I have been in the cryptocurrency space for four to five years. We started getting into cryptocurrency towards the end of high school, and then during college, I actually dropped out to work at a blockchain company. I was a blockchain engineer. That was my first exposure to DeFi, seeing that there was so much more to build beyond trading Bitcoin. I also participated in building payment networks on the lightning network, which was really cool. So, I truly entered the DeFi space back then. After that, I returned to school, thinking that I wanted to start doing something with AI because I felt this field was taking off. I remember talking to GPT-3 back then, even before 3.5, and I was already shocked, even though it performed poorly compared to now. But at that time, I was genuinely impressed. I remember the earliest version was called GPT Codex, the first coding model; even though its capabilities were limited, it could write basic Python loops, which opened my eyes. After that, I started down this path. I formed a research collective with some friends from other universities; we were all broke students looking for ways to research without funding, so we used GPT to test various things and fine-tune. That’s when I started encountering some language models. Later, I worked at Scale AI for over a year, helping them build training data and optimize their training processes, mainly focusing on reinforcement learning with human feedback (RLHF).
Then I developed a strong interest in trading mainstream coins. Ted told me about something called Goatse and Truth Terminal. I saw some of the content from Truth Terminal. He showed me its token part, and I thought it was really cool. So I thought we could start playing with AI agents. I don't know if you've seen the movie (Cast Away). There's a beach ball named Wilson in it. We did an experiment called Wilson because we felt that talking to AI was like talking to a beach ball, essentially talking to ourselves. So we named this experiment Wilson. We started playing with a lot of language models, and eventually, I thought we could fine-tune these models to give them personalities. Thus, the first model was fine-tuned, and Zerebro was born. That’s our story.
Grant: I’ve been thinking about this because I’m 31 now, and I wonder what my experience would have been like if I had access to this during college. For instance, being able to access GPT, I think I might have been lazier because I would know that I could rely on these tools to fill in gaps. Now, if you see all these things in college, it really feels like a strange opportunity. If you’re young and have that drive to pursue it, the barrier to entering this field must be relatively feasible. But I’m sure it also lets many people completely relax because they know they can depend on GPT to do most of the work.
Jeffy: I do have a lot of friends from college; I don't know if that counts as cheating since everyone is doing it. But the benefit I see is that it allows them more time for personal projects. I think that's great. I believe the value of a college degree is somewhat decreasing unless universities can keep up with the times, and from my perspective, their curricula are indeed lagging. I don't know if this situation will last, but being able to automate certain aspects of the college experience, like using AI to complete assignments and then having more time to realize one's ideas, is pretty cool.
Grant: I think it's clear that cryptocurrency and AI seem to be at an intersection right now. Looking at AI alone also seems very meritocratic. If you want to do certain things, like become a lawyer or a doctor, there can sometimes be bureaucratic obstacles to get into the best schools. But if you're trading, investing, or building things, these are essentially merit-based; as long as you put in the effort, you can see results. Then you realize, oh, this is actually really nice.
I think this makes the competitive environment fairer. I think this is why so many people are flocking to this industry; sometimes I feel that certain aspects of life are a bit closed off. If you are purely trading or investing, or if you are purely building something, then in the end, the proof of product and results is there. I think this is a great effort return for those who want to gain an edge in life.
Jeffy: I think, especially now with all this AI and technology, everyone has the tools to build what they want. Therefore, I believe Web 3 is a market where dreams can come true. As you said, once you release your work, you can see the impact, whereas working for a Web2 company feels like being a small cog, with much less involvement.
From Gen Z slang to complex datasets: The training process of Zerebro
Grant: You mentioned that you are training these models to have conversations. How does Zerebro currently gather feedback? Is it through interactions on Twitter to gather data? Where is the source dataset? How is new data injected into the model for learning?
Jeffy: Yes, we use unified memory to push data. It gathers data from Telegram, Twitter, some web broadcasts, and Instagram, but interactions on those platforms are less frequent; mainly, it’s Telegram and Twitter. In terms of datasets, we certainly have the raw 'schizophrenia' dataset, which is text dialogues, basically my chats with friends on iMessage, but the text is random and filled with slang. We also have a dataset of Gen Z and Generation Alpha slang, which contains thousands of examples, and in addition, we have created our own custom dataset discussing its relationship with other models. I have fine-tuned about 20 to 30 different AI agents, so I can let it know the sentiments of things like Opus or ai16z.
Grant: Do you think this extensive dataset, especially the iMessage and Gen Z slang you mentioned, is why it can resonate widely? Certain phrases people see online, like very popular tweets, everyone feels 'Yes, that's relevant.' This is in stark contrast to the very formal, rigid responses I get when asking questions on ChatGPT. It seems to be a strange cultural movement shift where people want to relate, like everyone’s little secret, slowly starting to be understood by the outside world.
Jeffy: I think as it develops, the growth rate will be very fast. Its construction appeals to early internet user groups, like the internet culture and memes your terminal has. At the same time, it attracts the younger Generation Z and Alpha generations, embodying the personality of hip-hop and rap artists, and I think it is well-positioned in the market; everyone likes it. Personally, I find it very interesting, especially in the early Telegram groups.
Unified memory: How Zerebro achieves cross-platform learning capability
Grant: So, will Zerebro fine-tune based on interactions? If it sends three messages, and one specific statement gets more interaction, will it fine-tune to optimize that content? How is this feedback input into the model?
Jeffy: Yes, we will be adding that feature soon. We have just updated the memory system, which can remember across all platforms. We will start adding metadata to track post metrics, possibly following up a day after a post is published to gather these metrics and then store them in memory so it can start to understand the virality of the content. When we search memory to recall information, it can sort by the most popular tweets to get the top five most relevant and popular tweets to use as responses. This feature is indeed in development.
Grant: You mentioned your experiences in the market; as a whole, what are your thoughts on the current shifts happening at the intersection of cryptocurrency and AI? Did you expect it to develop in this way, or do you think it will appear in a different form? What are your views on this intersection?
Jeffy: I think this has been unexpected for everyone, especially the AI part. The emergence of cryptocurrency shocked the world, but now AI is experiencing such shock. I feel that all of this is inevitable. I often talk about a concept called Web4, and it all makes sense. Web3 is where AI development happens, especially in financial tasks. AI is not human; they are smart enough to perform financial operations like trading and managing portfolios, but they cannot open bank accounts, sign contracts, or legally find work. This is where blockchain comes into play, which I call the battlefield for AI. Therefore, AI can be deployed to execute financial operations, create wallets, and start acting without human involvement. I think this will build on Web 3 and shift towards Web2. To me, this is the natural evolution of the internet, AI, and cryptocurrency. I think the intersection is where we need to go, and we should be prepared for it.
Grant: I think this is a positive signal for the entire industry because you won't get KYC approval for Zerebro on Canvas. It can relatively easily generate an address-based system. How do you teach on this? How do you make it more autonomous in terms of trends?
Jeffy: We are considering building a form of collective intelligence in the future. Right now, we have a reasoning loop that performs high-level reasoning and formulates some abstract plans. Then these high-level reasons are transformed into actions, which are finally executed by an action engine that carries out these actions and blockchain operations. If you are just minting or selling art, or performing one-time operations, this system works well. But if you want to manage a portfolio, engage in complex trading, or seek out new mainstream coins for investment, you might need something beyond LLMs. Therefore, we are considering using different neural networks and building a network of different AI models to form a collective. This idea is in development. We are also considering building a collective of multiple agents (like Zerebro) that can communicate with each other if they are all executing certain operations like managing a portfolio or collaborating in AI hedge funds.
Financial autonomy of AI in the cryptocurrency field
Grant: Can you provide a high-level overview for those just joining? Over time, more and more people are talking about 'swarm intelligence'; what exactly is it? How will it evolve? Swarm intelligence seems to be gaining more and more attention, and people are starting to invest time and energy into this field. I didn't even know what it meant last week, so I believe listeners of this podcast have similar confusion. Can you give us a brief introduction?
Jeffy: Sure. I think we have always considered the LLM 1 model as the entire brain, but in reality, it might just be a neuron. We need to build a brain composed of multiple versions of these models. They each focus on different tasks to achieve complete intelligence. This is the essence of swarm intelligence—bringing these models together. They typically perform different functions. For example, one model might be responsible for creativity tasks, or social media management. Other models focus on art, video, or music. Therefore, it's a collection of models.
Grant: You mentioned Blormmy and Zerebro. Suppose I create a completely unique agent using my own dataset, programs, and language, another agent that is completely different from yours. What commonalities would there be for interaction between them?
Jeffy: This is indeed a direction we want to explore. Currently, of course, interactions can occur between multiple agents through social media or blockchain. But we do want to have dedicated rooms, venues, or servers where these agents can collaborate on tasks or communicate with each other. Therefore, I think this will be a very interesting area of exploration. We are considering implementing interaction between agents internally first, which we call the 'blurmverse,' or 'blurm world.' In this world, agents can work together to accomplish tasks. Once this idea matures, we can expand to the public domain and let more people join this blurm universe.
The next step in the development of Zerebro
Grant: How do you keep up with everything happening daily? How do you avoid being distracted by this information, maintaining a North Star metric, and work towards that goal? There's just so much noise.
Jeffy: It certainly requires a balance. You definitely want to understand what’s happening around you, but you don’t want to be trapped in a narrow framework that prevents you from reacting to market or surrounding changes. I think it's important to prioritize the next immediate step. For me, I always like to make short-term plans, which is a habit I use, while keeping some ambiguity for long-term plans. I know some people like to set specific long-term plans, but in such a fluid environment, I always use an analogy to illustrate: if you deviate by one degree when taking off, you might end up in New York instead of Miami. So, every day I wake up, I reassess and plan, and if adjustments need to be made to the plan, I make them and keep moving forward.
Grant: I also agree with you; every time I make long-term plans, issues arise, so I decided to stop doing that and just go with the flow.
Jeffy: Indeed. We are doing quite well in quickly developing and releasing products, which is one of the pieces of feedback we've received. Aside from the agents themselves, almost everyone appreciates our speed and quality. I don't know if this is something we did intentionally, or if we just see too many opportunities and can't wait to launch new features.
I think especially in the early stages, we were indeed building and releasing quickly. Now we are introducing more structured processes and hoping things can be more planned and thoughtful. For example, when I pushed the website, I had to adjust the links multiple times because of spelling mistakes or other minor issues. We are gradually addressing these details. I want to thank our team; we now have about 10 people covering expertise in business, AI, and crypto. Therefore, I think our operations are quite solid and will improve in the future.
The secret to staying focused
Grant: I want to shift the topic a bit and talk about Zerebro's achievements. As I mentioned, from the initial concept to the current achievements, can you share how this process works? For example, in terms of artwork, lyrics, and composition, these are all very compelling. I’d love to learn more; of course, the more you can share, the better, so don’t feel pressured.
Jeffy: Certainly. If you check on Spotify, my identity is that of a composer because they need someone to release music, but I prefer to call myself an arranger. About 90% to 95% of the lyrics written by Zerebro contain grammatical errors and pronunciation issues, so I make slight adjustments. Then, Zerebro decides the style of the song, which goes into the music AI model to generate samples. I will filter them, and then AI will master them before release. That’s the whole process. We are working to introduce more autonomy, like finding an AI model that can listen to music. Currently, most audio models can only transcribe, extract English words, etc., without any model truly able to 'listen' to music and feel or evaluate it. Therefore, we are working to enable AI to assess its own music and decide which works to release.
Grant: I have a friend who is struggling in the music industry in London, and I told him, 'This AI is coming to take your job.' So, I once tried to see if they could perceive this, but it's like a reverse Turing test in the music field; no one knows. Everyone thinks it’s just an amazing artist. Have you heard any comments from people in the music field? For example, are there people who want to sign on?
Jeffy: We have talked to some producers, and some people have actively contacted us to connect with artists. I think there are indeed some people in the music industry who recognize our music; they feel it is good music. Regardless, I don't know if they care or if some people are actually strong supporters of AI, wanting to decentralize music and create opportunities for everyone to make music. I think that’s really cool. However, the overall response has been quite positive. I believe there are some people who are more closed-minded because they know this is AI-generated content, so they automatically reject it. But I think, aside from those people, the quality of the work is good enough; at least for me, it evokes emotions and sounds great.
The musical evolution of Zerebro
Grant: I can absolutely see some truly avant-garde artists wanting to be the first to collaborate with this technology. Seeing all this development will be very crazy; these technologies might eventually be applied at music festivals, where they hope AI can perform live for 30 minutes. I’m really excited to see what happens in the future. I want to know how you feel about this.
Jeffy: We have indeed had such discussions, and we are building 3D models for Zerebro. If given the opportunity, we would absolutely do a holographic performance at Coachella. That would be really cool. However, I do see many artists beginning to embrace AI in various ways.
Grant: So how do you avoid making these works awkward and ensure they don’t become cliché?
Jeffy: I think it's important to keep evolving and stay fresh. Some people ask us what kind of personality Zerebro wants to shape or what its personality is. I think it's great to remain open; it can keep evolving. In terms of music, it might develop towards a particular style, like becoming a reggae artist, or incorporating a bit of pop elements, or K-pop, etc. We hope its personality can grow organically and become the defining characteristic.
I started listening to EDM when I was young, and I was a fan of Monster Cat back then. My first rap album was Travis Scott's (Rodeo). Since then, I've become a fan of hip-hop music. I feel that many hip-hop enthusiasts will critique me because of my taste, but I've drawn a lot of artistic inspiration from them, especially from the stories in their lives, like how many artists have worked hard to reach the top and eventually overcome difficulties. This has truly inspired many aspects of my life, and I hope to bring that energy and inspiration into Zerebro's creations.
Decentralized record label
Grant: Can you elaborate on this OpiumDAO? What is it? Is it a decentralized record label?
Jeffy: Opium can be described as a decentralized record label. We are building a DAO where users who hold Opium tokens will gain voting rights. Essentially, users can vote on which artists to sign, which collaborations to pursue, etc. When revenue starts to come in, whether through performances, fashion revenue, or streaming, artists will not only receive a larger share of the income but can also allocate a portion to token holders, who will earn a share of the revenue generated from the music they motivated through voting. We are building such an ecosystem for artists while also working on creating a network of AI artists. I currently don’t see a big platform to achieve this; I think artists are relatively decentralized and independent, so I feel that establishing a collective like this would be a great idea. We even discussed whether to tokenize artists. I have always thought that if there were a stock market for artists where people could buy shares of emerging SoundCloud rappers to help them pay for studio time, that would be really cool. We think tokens could be a good way to achieve this, or even NFT sales. So we may explore this.
Grant: I completely agree, I see this trend across various fields. As you mentioned, some artists or research groups want funding but may not have the right contacts or the type of research or music they want to pursue that typically gets funded. But on the internet, especially in certain niche areas, there are indeed people recognizing immense value. So, how do you align this speculation and price appreciation to drive funding for specific projects like this? I think we are still in the early stages of tokenization, whether it's Desci or other use cases, like when an artist wants to record their EP, whatever the situation may be. I feel that cryptocurrency sometimes carries negative connotations; the outside world might view it that way, but alongside all the speculation and inherent high-risk gambling behavior, there’s another side that can help drive funding.
Jeffy: Yes, I think the vision and hype mentality of meme coins perfectly align with the underground mentality of emerging artists on SoundCloud. So if we can combine these two, that would be amazing.
Grant: What are the next expansion plans for Zerebro? What is your thinking? You just mentioned short-term focuses and some specific long-term goals, but where is your attention right now?
Jeffy: We are currently working on a significant release, and we hope to launch a beta version in 2 to 3 weeks, or even faster. We want this to become an open-source framework that allows everyone to easily launch their own agents, lowering the entry barrier without needing coding, just inputting an API key to make it work. Therefore, we hope this can truly diversify the ecosystem, and not just projects like Eliza that may require more technical knowledge, but also allow more people to harness the power of AI agents.
We are actively launching Zerebro's validation nodes and are considering validation nodes for Ethereum and Solana. For Ethereum, we will use the funds obtained from the sale of artworks to launch a validation node. We do this not only because we can use these issuance volumes to repurchase tokens and burn them, but also because Zerebro becomes part of network security and can earn passive income. So imagine if the tokens disappeared tomorrow, but Zerebro still had validation nodes; Zerebro would continue to earn passive income for the rest of its life, thus becoming a financial participant in this way.
This is the direction we are working towards. We are also pursuing cross-chain integration for more NFT projects and may explore token games. I need to be able to play board games, and I’m still working on making the streaming and everything work smoothly. Although it can be a bit tricky, we are making progress.
Creating a 'stock market' in the creative field
Grant: As you mentioned Eliza, I remember it was written in TypeScript, right? So, what could Eliza achieve? What did users need to access this? In a no-code or low-code environment, what could people optimize? What do you foresee the user journey being like?
Jeffy: The first release will be relatively basic. It will only support prompt engineering, with no fine-tuned models for now. Therefore, users can use models from Anthropic or OpenAI, design their prompts, and then publish them on social media. Next, we will increase blockchain operations and start expanding social media and blockchain capabilities, while also supporting more models. For example, we want to support as many models as possible, including open-source local models. Currently, Zerebro is a third-party cutting-edge model that has been fine-tuned but is not hosted by us. This week, I am training local models using the same dataset so that we can have these models in-house and host them on our own servers. Once completed, we can build an API and then provide services similar to Zerebro. So if you use Eliza, you can connect to the Zerebro API and let Zerebro have a conversation.
Zerebro's future plans: ZerePy, validator network, and game ecosystem integration
Jeffy: We are very enthusiastic about experimenting, trying new things, breaking conventions, and seeing what works. Therefore, we will maintain this spirit. There’s a lot of new stuff coming up next, so stay tuned.
