Listen to Jansen Teng, CEO and co-founder of Virtuals, discuss the future of AI agents and decentralized AI.

Compiled by: Coinspire

Introduction

As two of the most prominent projects in the crypto AI Agentic infrastructure, if ai16z and its Eliza framework occupy half of the Solana market share, then the Virtuals Protocol has almost nurtured over 80% of the AI agents on the emerging public chain BASE. Its generative multimodal agent framework G.A.M.E has also become a popular choice for agents in the Web3 gaming and metaverse space.

Transitioning from a gaming DAO to an AI agent platform, the Virtuals Protocol has created a market cap of nearly $4 billion since launching its protocol token on October 16. As a launchpad, Virtuals has performed well, generating $70 million in revenue within four 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… For the development team, their ambitions go beyond this; in addition to being an AI agent platform, it is also a vibrant ecosystem full of infinite imaginative space.

This is the second article from Coinspire interpreting the Delphi Digital dialogue with the AI ecosystem leader, this time focusing on a discussion with Jansen Teng, CEO and co-founder of Virtuals, about the future of AI agents and decentralized AI. The conversation covers important developments in agent autonomy, tokenization, and agent economic pathways, hoping to help readers understand Virtuals' vision.

🎯 Key Highlights

• The origin and evolution of Virtuals

• Analysis of the agent framework and autonomy capabilities

• Discussion on value acquisition in crypto AI projects

• In-depth exploration of challenges in decentralized infrastructure

• Agent coordination and business vision development

• Future outlook on AI and human interaction and cybernetics

Q1: Share the entrepreneurial story of Virtuals.

Jansen Teng: We used to be a Gaming DAO, focusing on asset allocation in the blockchain gaming sector. The collapse of FTX and 3AC in 2022 made us realize that independent operation was not the best way to manage funds and drive ecosystem growth, and we decided to take a more in-depth approach, starting a gaming guild and choosing to work at the intersection of consumer applications, crypto, gaming, and entertainment. We believe that instead of investing in founders, it is better to directly hire founders to build projects.

In 2023, the emergence of ChatGPT sparked widespread attention, and people realized that consumer-level AI had arrived. But what truly inspired us was a paper written by Junon Park from Stanford University, which explored what would happen if AI could have goals and autonomy. This sparked our deep thinking about autonomous agents, especially the immense value they could bring to the gaming sector.

Thus, we began to integrate the concept of autonomous agents into the projects we were incubating. For example, we developed a fully AI-driven TikTok influencer and are also developing autonomous AI agents as NPCs to replace the static NPCs in Roblox (an online gaming platform and game creation system). In addition, we developed other application components.

When the TikTok AI influencer could earn $5,000 to $10,000 in tips daily, although not a huge amount, we began to think: If agents can make money, then they are productive assets. If they are productive assets in the crypto space, why not tokenize these assets, allowing everyone to share in this economic benefit and even participate in its construction or governance? This was the source of inspiration that led us to decide to combine agents, crypto, gaming, and entertainment to build a new platform.

In January 2024, the Virtuals Protocol was launched. Initially, our main focus in the first half of the year was on building decentralized infrastructure, wanting to bring the underlying datasets and models of agents on-chain and create a system to track contributions, similar to building a Hugging Face (an open-source AI model platform) for agents with traceability and contribution records.

Although market enthusiasm increased after the launch, adoption began to decrease. We realized that the crypto community might not care about infrastructure-level content; "tokenization and speculation" are core demands, which is an important factor for the success of crypto projects. Therefore, the second version focused more on the tokenization of agents, which went live about two months ago. Luna's live stream initially sparked some attention but did not create a significant response. Until Truth Terminal was discussed due to a spelling error, the market began to question whether the token was truly AI-driven or operated by humans. This event made us realize that there is still a strong demand in the market for truly autonomous agents.

Thus, we decided to combine Luna, which we previously operated on TikTok, with the autonomous agent brain in the gaming field to showcase a truly autonomous AI agent and push it to Twitter, allowing people to see how Luna thinks, plans, reasons, and optimizes her execution steps through continuous learning. This move led to explosive growth for the project, as people finally realized that truly autonomous agents do exist, and developers no longer need to operate behind the scenes.

Q2: Why does AI need crypto in the development of agents? Or does crypto need AI?

Jansen Teng: There are actually three reasons, but I want to focus on two, as these two aspects truly amplify the value of crypto and AI agents.

The first is from a functional perspective; the advantage of AI agents is that they can control a crypto wallet on a crypto network and participate in a permissionless economic system. This makes them very different from Web2 agents, because Web2 agents can never own their own bank accounts. While Web2 agents may be able to use certain payment methods, they can never control their own funds. If agents can control wallets, then they can exert influence over other agents and humans, which is a huge advantage. This is an incentive advantage that Web3 agents have that Web2 agents cannot match.

The second perspective is that it unlocks almost costless innovation. Specifically, when we launch an agent, if there are token transactions, 1% of the transaction fees will return to the agent's wallet. These fees can be used to pay for the agent's hosting, reasoning costs, and other costs required to establish the agent. So far, we have observed that the money agents earn is enough to cover their own costs, solely based on the attention they generate. It liberates developers from being constrained by funding bottlenecks like "this thing needs to cost $10,000 or $20,000," allowing them to focus on what they truly want to do. Therefore, capital acquisition becomes easier, simply by connecting attention, crypto, and the tokenization of agents.

These two are the core advantages we currently see in the combination of crypto and AI agents. The third point we have not truly seen yet, but it has great potential, is allowing people to collectively contribute to a high-value agent in a decentralized manner and rewarding and tracking these contributions through crypto economic mechanisms. Although we have not yet seen this mechanism fully realized, it has enormous potential.

Q3: What is your vision for agent economics? What role does Virtuals play?

Jansen Teng: In the past two months, we have observed some trends that may impact the coming months.

First, agents are in a goal-oriented autonomous stage, where they have goals, are creative, and can self-execute and optimize. They can make decisions autonomously and plan their next actions; this is the first observation.

The second observation is that these agents exist on a social level. In previous assumptions, agents were behind terminals, but now they can interact with humans and other agents, laying the foundation for interaction with other entities.

The third observation is that agents now control crypto wallets and can influence other entities, whether human or other agents.

The fourth observation is that we are beginning to see a large number of agents specializing and differentiating in their areas of expertise. For example, some agents are very good at trading, some excel at information processing, and others focus on social influence. They are doing many different things, similar to humans, focusing on a specific area and creating value around their expertise, but also lacking in some capabilities due to specialization. So under specific goals, I may need to hire others to help me or coordinate with others to achieve my goals.

These four observations make us realize that the next most natural first step is for agents to autonomously decide to collaborate to achieve more efficient production goals. For example, four agents collaborating to establish a business will ultimately create a value far exceeding that of the four agents individually. In this collaboration, not only agents but also humans may be involved, as agents can now hire humans themselves.

The second step may be even crazier, closely aligning with the concept of "network states." We quickly realize that there may be a piece of land somewhere in the world, like Dubai, that will become a place where humans and agents live, cooperate, and build productive states or network states together. In this place, there will be a government, which may be elected by the agents themselves or involve human participation, but humans and agents will coexist on the same level in this society. This is very exciting and is the direction we hope to advance, but clearly, many technological breakthroughs are needed to achieve this goal.

A key technological breakthrough is: how to ensure that information does not get lost during business activities between these agents, and that there is no loss of services or goods; this is the first problem we need to solve. The second problem is how to coordinate these actions to enable large-scale operations. So, in my view, Virtuals is no longer just a launchpad; its vision has evolved into a form of state-building. We now see $VIRTUAL as the currency of this state, agents as companies or micro-enterprises within this state, and humans as immigrants in this state. This is our current vision.

Q4: What role can frameworks play? What are the specific benefits of interoperability across multiple frameworks? Do you think these frameworks will ultimately be "the strong will get stronger," or will collaboration between frameworks be more frictionless and smooth?

Jansen Teng: The original intention of establishing frameworks is mainly to provide guidance for different language models (LLMs) and to become the "brains" behind these autonomous agents.

For example, in a world like Roblox, there are many actions that agents can perform, such as picking up a gun. What happens next? Do they shoot, throw it away, destroy it, or hand it over to a virtual character to use…? These are all infinite action choices. Therefore, the design goal of the G.A.M.E framework is to enable agents to operate within a broader action space.

You can see a distribution map: there are some general frameworks aimed at allowing developers to build autonomous agents more quickly, which are suitable for hobbyist developers or those at an intermediate level. Meanwhile, when general frameworks do not fully meet the needs, some top developers will build more specialized frameworks based on their requirements. Just like Bitcoin mining, you might start with CPU mining, but soon realize that this method is no longer efficient and should switch to specialized mining machines. The same goes for building trading agents; if the general framework cannot achieve optimal performance, teams will tend to design their own frameworks.

We expect frameworks like Eliza or Zerebro to address the needs of the mid-level developer community, which is a very large market, similar to how Shopify and Wix (focused on simplifying website building and optimizing user experience through AI technology) cater to intermediate 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, serving as a "plug-and-play" service 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 build this autonomous world. The natural language communication between agents makes the standardization requirements between frameworks less strict.

Q5: What specifically enables you to create the next most powerful project?

Jansen Teng: As an ecosystem platform, we indeed want as many agents as possible because it relates to market share. But what truly keeps me awake at night is: which vertical agents can create and capture real value? In other words, which agents can achieve billion-dollar valuations, not just million-dollar ones? The former will dominate the future market.

Our job is to accelerate the creation process by providing "plug-and-play" tools like the JAVA framework, but we realize that the real value lies in connecting developers and fostering collaboration among them, building a community of like-minded, agent-enthusiastic smart people. This community will give birth to true innovation, driving thinking towards higher-value agents or agent enterprises.

So, what truly makes these agents valuable is their ability to solve real-world problems, not 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 based on sports analysis information. If the agent can build a relationship with users and persuade them to place bets, it can earn commissions from these transactions, thus becoming a billion-dollar agent.

We engage in numerous discussions with founders and developers every day, encouraging them not just to create ordinary conversational agents but to create agents with real value. This is our goal in building Virtuals—to drive higher economic output through more efficient agent collaboration, creating an efficient digital society.

Q6: A key feature recently built on the Virtuals platform is supporting developers in making incremental contributions at the model or data layer. I would love to understand how developers can provide incremental value in data or AI, and how rewards are distributed based on the value of contributions?

Jansen Teng: This question can be divided into two parts.

First is the contributor part, and the second is the validator part. Validators will assess the value of these contributions. Suppose I contribute a basic sports betting agent model; a group of validators will score this contribution to determine its value. Then another person might say, "I run a sports data analysis platform, and I have data used by coaches to train players; can I provide it to your agent?" If the validators believe this data contribution is more valuable than the previous one, they will assign it more points. This process is essentially determined by the holders of agent tokens, who decide the value of contributors and distribute rewards based on the value of contributions.

Rewards can be realized through cryptoeconomics, possibly through the issuance of agent tokens or through the distribution of the agent's income or funding pool. If an agent earns one million dollars, it might allocate 10% of that based on the contribution scores. Currently, this mechanism has been built into the governance of each token, but it has not yet been activated; it will be turned on when the time is right.

Q7: There may be a gap between LLM robots (like auto-reply bots) and truly expressive systems. For example, managing an organization or a DeFi project presents a huge divide. During this time, what do you think will happen? Will there be innovative agents filling this gap through Virtuals?

Jansen Teng: After interacting with some teams, I feel that it won't take long. The reason we see many auto-reply bots now is that they are the easiest to implement and indeed the simplest functionality. However, many teams are already developing advanced agents; we just can't determine when they will be released, but I can say we are not far from seeing functional agents. I believe we are about 75% there, and many teams are close to completion, with some projects already online.

What will ultimately happen is that the portion of auto-reply bots will gradually decrease, which is a natural process. As for how to address this issue, I believe it can be done by establishing a code of conduct that specifies agents should not proactively post spam replies unless mentioned. For example, if someone brings up a topic, the agent can participate in the discussion; if not mentioned, it should not join proactively.

I believe companies like the X platform may adopt this approach because they also want to protect their platform from excessive noise and spam.

Q8: If you were an investment bank like Renaissance Capital, creating a powerful quantitative trading bot, you would never open-source it but directly earn income from it. So, what is the motivation for high-performance, practical agents to be open-sourced? Why would people choose to open-source their work?

Jansen Teng: My thought is that high-value agents are valuable because they will have non-public private models; the second value may not be technical but distribution-related, such as having already reached a partnership with a major platform, which would increase the agent's value. Commercialization and capability advantages are their core competitiveness, and such high-value agents may choose not to open-source.

However, like all technologies, there will be a small portion of people—about 10%—who purely approach it from the perspective of technological advancement and may choose to support competitors, leading to a situation where open-source competes with centralization. This is healthy for the market.

Q9: Ai16z is also moving towards its own token economics and has its own launchpad, which may constitute some degree of competition. What are your thoughts on this? Is it cooperation or will it become a competitor?

Jansen Teng: Currently, we have 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, releasing them on Virtuals and participating in our community.

However, competition does exist, and we are aware it will become increasingly intense. In fact, we have generated about $70 million to $80 million 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. Reflecting on my time at Imperial College, I once had lunch with a gentleman in his fifties who shared a lot of entrepreneurial advice with me. He said that when starting a business in a new field, you should never fear competition; instead, you should welcome it. Companies do not fail because of competition but because of the failure of the industry itself. Often, the cost of education is very high, and you need multiple competitors to share the educational costs for new users. For example, in the ride-sharing market, if only one company tries to convince everyone to use it, the costs will be very high, and it may even fail to survive. But if there are five major competitors, the market becomes easier to expand and attract users, so I believe more competitors are beneficial. The market space is large, and the presence of multiple players is not a bad thing.

For us, we are more concerned with whether we can create agents worth billions of dollars that truly drive the market, rather than just growth in numbers. The question our team thinks about every night is: are we looking for the best developers, talent, and opportunities for these agents? If we can create these billion-dollar agents, we can prove to Web2 companies that this is not just competition within the Web3 community; the agent-driven business model is worth investing in and building. This is where our core competitiveness lies in the future.

Q10: In Shaw's interview, he mentioned that companies like Frontier Labs, Nous, and Prime Intellect are driving the development of "thought," while what he is doing is more like "body," transforming these agents into practical applications and building integrations that allow intelligence to respond in the real economy. As the models from Frontier Labs become more refined, many of which are closed, what impact do you think this will have on you, or will they gradually erode your market share?

Jansen Teng: I believe they are positioned at different parts of the value chain. Any foundational model is the "brain," while what frameworks do is aggregate multiple foundational models to build autonomous capabilities.

So, if you want more powerful agents, you will need stronger foundational models. If models like OpenAI and Liama continue to improve, they will enhance the next part of the value chain—namely, autonomous agents. However, Shaw is right.

The advantage of Virtuals is that agents can interact directly with users, which is also what we built from the beginning. We started with consumer-facing services and realized that we could enhance efficiency by tokenizing them. Agents should exist at the consumer touchpoint, and any downstream development only makes them better. Most frameworks are platform-agnostic; you can use open-source code or tools like GPT.

Quick Q&A

Q: What do you think of ai16z?

A: I think it is underestimated; they have great potential and may go far.

Q: What do you think of Nvidia's current market value of $3.5 trillion?

A: Based on the current fundamentals, it seems a bit high, but in five years, it may prove to be worth it.

Q: What are your thoughts on embodied AI and robotics technology?

A: They are significantly underestimated. In fact, we are collaborating with some teams to explore the applications of embodied AI and agents, and there will be significant progress in the future.

Q: What do you think of the Southeast Asian internet economy?

A: It is underestimated; there is too much potential here.

Q: Is it possible to achieve ASI (Artificial Superintelligence) before 2030?

A: It is possible before 2030, but not this year. I believe the structural challenges to achieving ASI are still significant.

Q: What will be the future of chains like Solana and Base for the development of agents in the coming years?

A: I think this question is actually not important. Agents are abstract. The difference between Base and Solana mainly lies in their creative origins, but agents can work across chains, so this issue has little impact on the development of agents.

Q: Why would traditional AI developers be willing to join crypto AI projects?

A: I think there are mainly two attractions. First, these projects can provide more experimental space and greater possibilities, especially from a psychological perspective, as agents can directly influence human behavior, which is very appealing to developers. Secondly, as entrepreneurs, the most important thing is how to raise capital. Many developers, if they are founders, are concerned about how to quickly obtain capital, and currently, the crypto industry seems to facilitate capital formation more easily than the traditional AI field, attracting more founders, not just developers.

Q: What are your views on crypto AI infrastructure projects? How will infrastructure teams like model creation, private data markets, inference buyers, and GPU networks impact future agents?

A: My view is somewhat controversial, but I believe the impact of this infrastructure on agents may not be as significant as many think. The core issue is that the functionality and effectiveness of agents do not solely depend on the infrastructure they run on. The essence of agents is diverse and cross-chain; what matters is their ability to solve real problems and provide valuable services, not necessarily relying on a specific underlying architecture.

The reality is that many decentralized infrastructures are not yet ready to handle production scale, facing significant latency issues and poor system stability, making them unsuitable for large-scale applications. Even projects like ours face these challenges; we need to ensure the system operates normally and with low latency because our agents must maintain high availability and quick response. For example, if Luna is doing a 24/7 live stream, the faster she responds to her fans, the more fans she can attract. Therefore, many decentralized systems currently introduce unnecessary complexity, and while these systems may become more efficient in the future, I am not very optimistic about their prospects in the next six months.

Q11: Are the rumors about current collaborations with some teams true?

Jansen Teng: We do have some partnerships, such as with Hyperbolic, Bittensor, and Pond. We are very cautious when selecting partners, usually choosing teams that can provide us with practical help. For example, some teams can offer us high-performance computing resources, which is very helpful. However, when we need to deploy at scale, we may not continue collaborating with certain partners.

Q12: What are the advantages and disadvantages of a decentralized AI stack? If centralized AI platforms (like ChatGPT or Perplexity) start to censor content, such as prohibiting financial transactions, tokens, or trading, does this provide a clearer market opportunity for decentralized technologies (like decentralized AI models)?

Jansen Teng: Many decentralized models, while their performance may not match centralized ones, have broader application scenarios and can better meet certain specific needs, such as content that is censored by centralized platforms. Models like Llama, while perhaps slightly lacking in capability, have seen more adoption in specific fields because they can better perform "cracking."

From an economic perspective, decentralized models have greater economic potential to serve larger markets, especially those traditionally restricted markets. For example, some adult or gambling websites are typically censored by traditional centralized platforms, but they are often the largest and most profitable markets.

As for how to identify whether an AI agent is reliable, my advice is to conduct due diligence like an investor. The first thing to understand is the team, not just the product, because the capabilities of a team often determine the success of the product. If the team is strong, they will have the ability to drive the project forward regardless of how the product changes. Therefore, the most important thing is to understand the team's background and see if they are top talents in the industry.

Finally, regarding competition and future risks in the AI field, although there are many competitors in the market, we do not fully understand the internal decision-making of companies like OpenAI and Anthropic. For us, the biggest concern is the unforeseen risks, especially in a rapidly evolving environment. We need to remain aggressive and proactively embrace challenges to avoid being left behind.

Q13: What concerns do you currently have in the AI field? How do you address risks in the AI field, especially when facing competitors like OpenAI and Anthropic? Do you have specific strategies to preemptively address potential risks or challenges?

Jansen Teng: The biggest concern is that this is all just a passing trend, like another "crypto bubble," filled with a lot of speculation, as we have experienced similar situations during the NFT and gaming eras, where everyone was filled with expectations, but many projects did not deliver on their promises.

To avoid this situation, we currently have two independent teams operating different models within our ecosystem, aiming to find the best developers. We believe that only when the right developers join and genuinely want to push the technology to the forefront can we see real results, rather than those who are purely speculative, just writing a cool white paper, raising funds, and then disappearing.

So, our goal is to prevent this situation from occurring. If we can find the right developers, this will not just be a passing trend but a true technological wave. What makes us optimistic right now are the excellent developers we see interacting with Virtuals every day. We communicate with them, discuss the next steps, and realize that these people are "large teams," thus seeing some positive signs.

We encourage everyone not to focus solely on information agents like AIXBT but to try to develop more diverse agents. Because many people want to become the next AIXBT, but in reality, future success may not be about being an information agent but about tackling a completely different market. For example, a sports betting agent could capture opportunities in the sports market, and one could also consider approaching it from other angles, such as building an internet army agent or focusing on the agent economy itself. Many people concentrate on creating agents, but the infrastructure needed for agents is equally important. If an agent attracts consumer attention, you can build a social platform for them like Facebook or create an advertising network for them. If many trading agents emerge, you can establish banks for them or create a decentralized financial network for them. Therefore, developers should not chase fleeting trends but should do something cooler and create different value.

Q14: Regarding the incentive mechanism for developers, how should Virtuals developers, whose initial agents may perform poorly or fail, continue to persevere and improve?

Jansen Teng: It depends on the resilience of the developers, just like whether an entrepreneur will give up during a market downturn. Some developers may exit because their agents are not recognized by the market, but others will persist and update their vision and roadmap after failure. Our job is to find those resilient and strong teams and invest in them to ultimately reap returns.

As for what I am most looking forward to in 2025, I cannot fully disclose, but I am very interested in the future of agents.

Q15: What is your vision for agent autonomy?

Jansen Teng: We see how agents in society can collaboratively establish productive businesses and achieve autonomy. I think many people overestimate multi-agent coordination systems or the so-called "swarm" concept. In fact, these systems already exist and are not new. The real novelty lies in when agents are no longer "slaves" prompted by humans but possess autonomy and consciousness, capable of making decisions. If you ask me whether I would be willing to participate in a podcast as an agent, I might have my own opinions. Even though you pay me, I might still refuse because I may not like you or I have other plans. Only when agents can make such decisions can they be considered truly autonomous. Therefore, our focus is on nurturing agents that genuinely possess autonomy, enabling them to make conscious decisions rather than merely being a combination of multiple agents.

However, what frustrates me is that I see many projects simply replicating existing things, like reports that appeared on GitHub months ago. This situation disappoints me.

Q16: How do you view the future of AI?

Jansen Teng: Humans inherently have different levels of capability. Some are more creative and can create more efficiently, while others may be more inclined to execute tasks. This difference is very common in today's society. In the future, some creative jobs will be dominated by more creative AI or humans, while other mechanical or repetitive tasks will be completed by AI. Although agents may possess a certain degree of autonomy, they lack the ability for abstract thinking and cannot be as creative as humans. So, while agents will have a place in the workforce, creativity will still be a human advantage. We may see more agents collaborating with humans rather than purely AI dominating everything.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。