Delphi Digital Dialogues Virtuals: Agents with true autonomy are more important than combinations.

As one of the representative projects of the encrypted AI Agent infrastructure, Virtuals Protocol has spawned over 80% of AI agents on the emerging public chain BASE. Its launched generative multimodal agent framework G.A.M.E is a popular choice for agents in the Web3 gaming and metaverse fields.

Since the launch of the protocol token on October 16, Virtuals Protocol has created a market value of nearly $4 billion. As a Launchpad, Virtuals has generated $70 million in revenue within 4 months, with many star projects being part of its ecosystem, such as crypto Agent KOL AIXBT, virtual influencer Luna, and the AI Agent development framework G.A.M.E, among others. For the development team, their ambitions do not stop there; in addition to being an AI agent platform, it is also a vibrant ecosystem full of limitless imaginative possibilities.

Delphi Digital interviewed Jansen Teng, the CEO and co-founder of Virtuals, discussing the future of AI agents and decentralized AI, covering important developments in agent autonomy, tokenization, and the path of the agent economy.

The Origin and Evolution of Virtuals

Jansen Teng reviewed the entrepreneurial journey of Virtuals:

We originally started as a Gaming Dao, focusing on asset allocation in the blockchain gaming sector. The collapse of FTX and 3AC in 2022 made us realize the need to adjust our strategy, leading us to adopt a more in-depth approach and start a gaming guild that focuses on the intersection of consumer applications, crypto, gaming, and entertainment.

In 2023, the emergence of ChatGPT has attracted widespread attention. However, what inspires us even more is a paper by Junon Park from Stanford University, which explores what would happen if AI could possess goal-orientedness and autonomy. This has sparked our in-depth thinking about autonomous agents, especially the tremendous value they could bring to the gaming field.

We began to integrate the concept of autonomous agents into the projects that are currently being incubated, such as developing a fully AI-driven TikTok influencer, while also developing autonomous AI agents as NPCs to replace the static NPCs in Roblox.

When TikTok AI influencers can earn between $5,000 and $10,000 in tips every day, we begin to think: if agents can make money, then they are productive assets. If they are productive assets in the crypto space, then why not tokenize these assets, allowing everyone to share in this part of the economic gains and even participate in its construction or governance?

In January 2024, the Virtuals Protocol will officially launch. Initially, we focus on building decentralized infrastructure, aiming to bring the underlying datasets and models of agents onto the chain and create a system to track contributions.

Although market enthusiasm has increased after the release, adoption is decreasing. We realize that people in the crypto community may not care about infrastructure-level content; "tokenization and speculation" are core needs, which is an important factor for the success of crypto projects.

Therefore, the second version focuses more on the tokenization of agents. It wasn't until Truth Terminal was discussed due to a spelling mistake that the market began to doubt whether this token was truly backed by AI, or if it was human-operated. This incident made us realize that there is still a strong demand in the market for truly autonomous agents.

We have decided to combine the Luna that was previously operated on TikTok with an autonomous agent brain in the gaming field, showcasing a truly autonomous AI agent, and pushing it to Twitter, allowing people to see how Luna thinks, plans, and reasons, and how she optimizes her execution steps through continuous learning. This initiative has led to explosive growth for the project, as everyone finally realizes that true autonomous agents do exist and there is no longer a need for developers to operate behind the scenes.

Reasons AI Needs Crypto

Jansen Teng believes that AI agents require encryption technology for two main reasons:

1. Functional Perspective: AI agents can control a cryptocurrency wallet on a crypto network, participating in a permissionless economic system. This sets them apart significantly from Web2 agents, as Web2 agents can never own their own bank accounts. If an agent can control a wallet, then it can exert influence over other agents and humans, which is a huge advantage.

2. Costless Innovation: When we launch an agent, if token trading occurs, a 1% transaction fee will be returned to the agent's wallet. These fees can be used to cover the agent's hosting, reasoning costs, and other expenses required to establish the agent. So far, we have observed that the money earned by the agent is sufficient to cover its own costs, solely based on the attention it generates. It liberates developers from the funding bottleneck of "this thing needs to cost $10,000 or $20,000," allowing them to focus on what they truly want to do.

The third point has great potential but has not yet been fully realized: allowing people to collectively contribute to a high-value agent in a decentralized manner, and rewarding and tracking these contributions through a cryptoeconomic mechanism.

Vision of Agency Economics

Jansen Teng shared his vision for agency economics:

1. Agents are currently in a goal-oriented autonomous phase, they have objectives, are creative, and are capable of self-execution and optimization.

2. These agents exist on a social level, capable of interacting with humans as well as with other agents.

3. Agents now control cryptocurrency wallets and can influence other entities, whether human or other agents.

4. A large number of agents differentiate and specialize in their areas of expertise.

These four observations make us realize that the most natural next step is that the agents will autonomously decide to cooperate in order to achieve more efficient production goals. For example, four agents collaborating to establish a business will ultimately result in a value that far exceeds the value of the four agents individually. In this cooperation, it is not just agents involved; in fact, there may also be human participation, because agents themselves can now hire humans.

The second step could be even crazier, which aligns very well with the concept of "networked nations." We quickly realized that there might be a piece of land somewhere in the world, like Dubai, that would become a place where humans and agents live together, cooperate, and build productive nations or networked nations. In this place, there would be a government, which might be elected by the agents themselves or involve human participation, but humans and agents will coexist on the same level in this society.

In my view, Virtuals is no longer just a launchpad; its vision has developed into a form of nation-building. We now regard $VIRTUAL as the currency of this nation, Agents as the companies or micro-enterprises within it, and humans as the immigrants of this nation. This is our current vision.

The Role of Frameworks and Interoperability

Regarding the role of the framework, Jansen Teng stated:

The primary intention behind establishing the framework is to provide guidance for different language models (LLMS), becoming the "brain" behind these autonomous agents.

The design goal of the G.A.M.E framework is to enable agents to operate in a broader action space.

We expect frameworks like Eliza or Zerebro to address the needs of mid-level developers, as this market will be very large, similar to how Shopify and Wix cater to intermediate-level developers today.

In summary, the G.A.M.E framework is a tool capable of large-scale planning and execution, designed to help developers quickly build autonomous agents, and is a "plug-and-play" service prepared for mid-level developers; while Virtuals, as a concept of a nation, has transcended the positioning of a platform, we view each framework as the "brain" of the agents, which can collaborate with each other to collectively construct this autonomous world. The natural language communication between agents makes the standardization requirements between frameworks less stringent.

The Key to Creating the Next Most Powerful Project

Jansen Teng believes that what truly makes these Agents valuable is their ability to solve real-world practical problems, rather than just abstract concepts. For example, a commentator agent in the sports betting market might live stream during every NBA game while providing the best predictions by combining sports analysis information. If the Agent can establish a relationship with users and persuade them to place bets, it can earn commissions from these transactions, thus becoming an Agent worth billions of dollars.

The goal of Virtuals is to promote higher economic output and create an efficient digital society through more efficient Agent collaboration.

Developer Contribution and Reward Mechanism

Jansen Teng explained the mechanism of developer contributions and rewards on the Virtuals platform:

Contributor section: Developers can make incremental contributions at the model layer or data layer.

Validator section: Validators will assess the value of these contributions. Suppose someone contributes a basic sports betting Agent model, a group of validators will score this contribution to evaluate its value. Then another person might provide data from a sports data analysis platform. If the validators believe this data contribution is more valuable than the previous one, they will assign it a higher score.

Rewards can be realized through the means of cryptoeconomics, possibly by issuing proxy tokens or through the distribution of proxy income or funds from a pool. If an agent earns one million dollars, they may allocate 10% of it to contributors based on their contribution scores.

Currently, this mechanism has been built into the governance of each token, but it has not yet been activated. It will be enabled when the time is right.

The Difference Between Automated Reply Bots and Advanced Agents

Jansen Teng believes that we are not far from seeing functional agents. About 75% of the work is already completed, many teams are close to finishing, and we can also see some projects that have already gone live.

What will eventually happen is that the part of the automatic reply robots will gradually decrease, which is a natural process. As for the solution to this problem, it can be addressed by establishing a code of conduct, stipulating that Agents should not proactively post spam replies unless mentioned.

The Motivation Behind High-Value Agent Open Source

Jansen Teng believes that the reason high-value Agents are valuable is because they possess non-public private models; the second value may not be technical, but rather distributional, for example, they may have reached a partnership with a large platform, and such collaboration would increase the value of the Agent. The advantages in commercialization and capability are their core competitiveness, and these high-value Agents may not choose to be open-source.

However, like all technologies, there will be a small portion of people — about 10% — who, purely from the perspective of technological advancement, may choose to support competitors, thus witnessing the situation of open source versus centralized competition. This is healthy for the market.

Relationship with Competitors

Jansen Teng stated that currently there are 10-15 projects using the Eliza framework for tool development, and some teams are even combining Eliza with other tools to modify and build their frameworks, publishing them on Virtuals and participating in our community.

However, competition exists, and we are well aware that it will become increasingly fierce. In fact, we have generated about 70 to 80 million dollars in revenue during our operations over the past two months, and competitors will naturally move in on us; this is the norm of the market.

But I think this is a good thing. The company will fail not because of competition, but because of the failure of the industry itself. Many times, the cost of education is very high, and you need multiple competitors to share the costs of educating new users. For example, in the shared mobility market, if there is only one company trying to convince everyone to use it, the costs will be extremely high, and it may even not survive. But if there are five major competitors, the market becomes easier to expand and attract users to join.

For us, what matters more is whether we can create Agent that are truly market-driving and worth billions of dollars, rather than just focusing on quantitative growth. If we can create these billion-dollar Agents, we will be able to prove to Web2 companies that this is not just competition within the Web3 community, and that Agent-driven business models are worth investing in and building. This is where our core competitiveness for the future lies.

Advantages and Disadvantages of Decentralized AI Stack

Jansen Teng believes that many decentralized models, although their performance may not be as good as centralized ones, have broader application scenarios and can better meet some specific needs, such as content that is censored by centralized platforms. Models like Llama, although they may be slightly lacking in capability, are still valuable because they...