Episode 262: Jansen Teng, Co-Founder of Virtuals Protocol on building the Infrastructure for AI Agent Society

Speaker A: Sa.

Speaker B: This podcast expresses the views of its participants as of the date indicated and may not represent the views of arca, where David Naij is a Managing Director. Moreover, such views are subject to change without notice, and ARCA has no duty or obligation to update the information contained herein. This podcast is being made available for educational purposes only and should not be used for any other purpose. The information contained herein does not constitute and should not be construed as an offering of advisory services, investment advice, or an offer to sell or solicitation to buy any assets or related financial instruments in any jurisdiction. Welcome back to Baselayer. I cannot say how excited I am about this. I have Jansen, the co founder of Virtuals Protocol, with me today. Jansen, how are you?

Speaker A: Great to be on.

Speaker B: So we have decided that Baselayer is going to be working almost 100% exclusively on what's happening with AI and AI agents and Web3. Uh, we've had some founders already on the show. A lot of shows are going to be released soon. There is a lot happening here, and Janssen and his team are right smack in the middle of it. They really have been kind of the ones pushing much of this over the course of the last few months, really kind of exploring. And the way that I've really thought about it is that there are assets. And assets can include things like large language models like Anthropics, Claude or Llama or Mistral and others out there. There's lots of them out there, new ones coming out almost every week now. And basically for every person out there, this is an asset that because of what Virtuals is doing, because of what AI16Z is doing, some of the other kind of players in the space, they are giving you access to these models in a way that makes it so much more enjoyable and so much more personable. Um, very similar to what I said, how Uber does not own any cars, Airbnb doesn't own any homes, but they have made it so that you have access to them in a way that you never did before. And this is kind of the way I think about Virtuals. So we're going to go on to about what Virtuals is doing and some of the kind of nuances there. But what we always do, Jansen, is we always talk to our founders about what they did before this and how they got inspired to get into this world. So let's go with that for a few minutes. What did you do before Virtuals and how did you come up with the idea to do this?

Speaker A: Yeah, so I think two parts, right? So the first part is like exposure to the whole crypto blockchain scene. That happened when I was still in uni, that was in 2016 and I was like mining Ethereum with uh, my laptop. Right? So that was the initial exposure to what this whole programmable Internet uh, blockchain might look like, right? Internet of ownership. So that was the first exposure. But in 2021, um, me, my co founders and our initial team were actually living at the intersection of blockchain gaming, um, initially. And so we were very exposed to the whole gaming front of things. And we had a venture studio arm that was effectively incubating new projects, um, at the intersection of gaming, consumer applications and crypto. So there was an initial positioning. But in 2022 when GPT came out, there was a paper written by Junsong park around what if AI can have some form of agency go driven autonomous. And we've realized quickly that if you apply that at the consumer surface area of gaming and entertainment, um, it solves a lot of problems. It creates infinite content, it creates hyper personalization. There's a lot of things it solves at that front and quickly. In some of our incubator ventures, like a full AI influencer on TikTok, like um, smart NPCs in games, we realized that they were actually generating revenue. Um, there's a very interesting play with crypto and blockchain and that became the inspiration behind virtuals.

Speaker B: Right.

Speaker A: So effectively it's a place where people can create autonomous agents easily for a myriad of use cases. And they can then co own these agents in a way where they share uh, into the economy upside, but they also get to govern um, these agents. So that is the initial stepping stone.

Speaker B: Yeah, you're not owning anything. When you use Claude, you're paying $20 a month. Or if you're using ChatGPT, you're paying $20aMonth to OpenAI or you're paying a Twanthropic or you're paying it to meta, um, you know, so you're paying that um, while very useful to do things, it's not actual. The idea again, going back to everything that's happening in Web3 is that there should be an alignment I am teaching you agent. You should also then be providing me some economic incentive or some alignment. Um, there just hasn't been that uh, up until recently. Again using the assets, these are things that are out there now, the LLMs that are out there and using them in ways that makes it much more personal, personable to you and having that alignment that we have always missed in the web. But at the very core of this is, from what I've understood over the last few months of spending hours and hundreds and hours trying to understand all this coming from a non technical background, I've had to basically teach myself AI and machine learning for the last three months. At the core of this is what you call game, um, which is effectively in many ways the framework. We're calling these things frameworks, we're calling these things launch pads. So game for all intent and purposes is the framework. You have things like a high level processor, which again you have the LLMs and they've been trained for 12, 15 months with thousands of GPUs, um, and they're getting access to all these types of open API data sets. But at the end of the day, you need to have a way to process that information. You need a way to have a vector store, which is memory bank, so remembers things that it actually said and things that it actually did. Uh, there is all these components there to make these things, as you alluded to, more autonomous, not just bots. There is a difference between bots which is basically do this, it does that, it comes back fine, doesn't really remember. So if I ask it again in a week to do something like that, it may have to be kind of prompted again to do all that. Whereas an autonomous agent, an agent, what we're talking about here, has a high level processor, has a memory bank, has the ability to say, oh, you've been talking to me about all these things for the last few days. It sounds like you probably want something like this. And so it actually can think through. We're getting closer to AGI, so talk to us about how game was designed. How did you come about this kind of framework and kind of the development journey?

Speaker A: Yeah, so actually it all started, it actually started, um, in a gaming environment, hence the name. Right. I mean it does stand, uh, for a more interesting word. But I think at the core of it it's very simple. When we're in gaming, one of the key things we wanted to solve for initially was we tried to create a mirror of the world or like an equivalent of the HBO movie Westworld. Right. Because we are thinking like if in this sandbox a bunch of these NPCs can actually be autonomous in their actions, what would that look like? Right. Would it create a gaming environment that's way more interesting, like Westworld, you know, like they're all living their own lives. You can come in As a human to screw things around, right? You can be the good guy, the bad guy. And what these NPCs do is they can react dynamically to everything that you do. So that was initially the idea is like, okay, if there's a gun on the table, um, there's a cow in the room, there is a uh, bag of TNT and dynamites, um, um, on the table as well. Now the question is now what would this NPC do? What would this agent do, right? And it will all then boil down to what its goals are, uh, how he perceives the environment and what it has learned throughout his previous interactions. So for example, if the goal setting for this NPC is like, hey, you are a villain and your goal is to create as much chaos as possible in the town. And then somehow in the history of his memory he was having an altercation with another NPC and he hates another npc, uh, that's in a bar somewhere in the town, right? So it can then perform an action of saying, hey, you know what, I see this bag of tnp, I know I hate the guy in the bar, why don't I try to blow up the bar, right? Or he can make another decision and he can say like, okay, maybe I should take the gun instead and cause havoc in the bar, right? So there's an autonomous decision making that this agent can do because of what has happened in the past and environment around it. So that was the original plan and for us to build that, I think like you mentioned, right, There's a few things in his brain. There's a high level planner, uh, that helps his plan actions, stepwise actions towards his goal. It takes in inputs from surroundings, its environmental state and its memory. And then down to a low level planner which effectively gives it an ah, ability to execute those actions within the environment. So within the game it's about picking up the gun within the open world is by calling APIs, calling um, any kind of action spaces. Then we have a short term working memory that effectively creates coherence in the actions. So if I pick up the gun, the next coherent thing to do is to fire the gun or to shoot a bullet as a warning, as a coherent action, an incoherent action would be like, I will um, take the gun and blow up the gun, right? So that's an incoherent action. So that's where the short term working memory works and it's a long memory which then after he blows up the bar, he remember that he blew up the bar. So that's Basically the initial brain that we've created for them to exist in uh, in a gaming world. And we quickly realized that about two and a half months ago that if you apply that same capabilities out there in the open world, you unlock a ton of use cases. So the first use case that we activated it out there was as an influencer. Luna was an influencer goal and said that, hey, you want to reach 100,000 followers, you have access to Twitter. This is the other access spaces you have, I. E, like controlling a crypto wallet, um, and a few other things. And then she started autonomously generating content, engaging with followers, tipping people to do stuff for her, creating jobs out there in the real world to influence humans to drive up her goal, which is to get as much attention so that she can get a lot of follower count. So that was how it started, right? And then quickly it started diverging. People started creating very differentiated agents, agents that specialize in information, uh, curation agents, especially in trading agents that act as a white hat, hacker, entertainment, whatever you might say, right?

Speaker B: AI XBT is obviously an example of that for the last few months has been something an agent that many people have been following because of its ability to look across the spectrum of crypto and be able to pick up signals, be able to pick up information that many, because there's so much information that's always happening in this industry that many occasionally miss. Um, and so AIXBT has been an example of one of those that uh, has created more attention for virtuals, obviously going forward. So yeah, example of, as I said again, your infrastructure leading to kind of this evolution of these new agents. But I am curious though. So the framework currently handles a lot of sophisticated memory systems and planning engines. So as I said again, you have vector stores because you have memory banks, you have the high level planner, you have the low level planner. What technical challenges do you foresee as agents become more complex? Because we're talking about the dawn of uh, swarms, we're talking about orchestration, we're talking about how these agents can start to do things more computationally intensive. Instead of just bull posting or kind of shit posting on Twitter, what do you see? Because even virtuals as a protocol, right now you're dealing with the CEO of Anthropic talking about there are some kind of walls that are coming out over the course of the next six to 12 months. We've all heard about the data scarcity side of things, how, uh, human trainable data is starting to become more scarce. You're starting to talk about Ah, synthetic data starting to feed these LLMs. Any kind of technical challenge as you see these agents becoming more complex and more capable.

Speaker A: Yeah, definitely, right. I think the key issue that we will likely face is when the action space of the agent increases. Think of it like today, the action space is like they can tweet to Twitter, right? Simple. I can control a, uh, crypto wallet. That's like a second action space. Simple. But if you start saying that, okay, um, I want this agent to start interacting with a fleet of other agents right? Now, imagine all the, I don't know, hundreds of really functional agents already existing in the virtual space today. If you want an agent to be able to leverage another agent, right, it has to see all these different agents and then conscious decision of like, hey, who should I interact with? Why should I interact with them? How do they help me achieve my goal in some form or some way. So as more of these option space increases, it creates this load or from pollution in a sense to these agents. So we might reach a point where a lot of times when presented with a too huge action space, uh, the agents will break down. They might only pass information that's at the front, front of the action space and at the last part, right? And everything in the middle gets dissipated or becomes chaos, right? So that's the reality of working with models today. So this is the first problem, right? And I think the way to solve this is to build some form of information abstraction. Like, okay, before you even see all of the agents you group agents into. Like, okay, these are uh, agents that ah, are trading agents and these agents are, are um, some form of creative tooling agents or whatnot. So you cluster them and then you create abstraction layers. So it makes information easier to be perceived by these models, by these large language models. So that's problem space number one. Problem uh, space number two. It's actually going to be very interesting because one of the key goals that we are trying to work towards is for agents to autonomously decide to collaborate with each other and through that collaboration create net productive things. Like think of creating net productive businesses, creating new collaborative IPs all by their own accord without a human in the midst prompting it, right? So that has been the goal. And I think one of the key problems that we see from testing these out is that when an agent wants to trade with another agent in exchange for their service or product, the hallucinations of these agents sometimes become a problem. Because agent A might think that I've already delivered the service that agent B has paid for. But it hallucinated without actually delivering that service. It's like the early days of the Internet.

Speaker B: It's the old double spend problem.

Speaker A: And also like back then, right, like information loss. Think of it as information loss, right. In some form. How do you reduce information loss to pretty much zero, Right. How do you avoid um, empty mailboxes being sent between these agents? Right. So there's a lot of uh, things that might need to be done. So like, think of it like the creation of some form of obligation standard, uh, payment, escrow standards that when agents and agents interact, they're interacting with a standard. So you reduce the amount of information loss in the system. So it's something that we have been perfecting um, over the last couple of weeks and into the weeks to come. And the idea is once you perfect that from a one to one communication, if you can reduce the information loss, then suddenly many to many interactions can still maintain that level of information. And, and then you start getting very interesting things, right? You get high level coordination, uh, at scale. Then you can start seeing real businesses being formed by these agents as well. So that's the second thing that we've observed.

Speaker B: So we're going to talk a little bit about uh, what's happening in the LLM space and then we're going to talk a little bit more about what's actually happening at virtuals in terms of graduation rates and kind uh, of obviously how you can scale up obviously to handle these new things that are happening there. Um, but something that's obviously the talk of the AI world right now is Deep SEQ just came out, um, basically saying that they could offer a parallel to 01 basically at a 95% cost reduction. Um, so we're starting to see the infrastructure cost of LLM start to drop pretty precipitously. I believe Virtuals may be using perplexity. There's obviously Claude, there's GPT4, there's Llama, there's Mistral, there's a lot of them out there. Each have their own different kind of flavors. Each of them have their own kind of way that the model helps the agent kind of communicate. Um, but at the end of the day, how does this all impact the strategy and the framework development? Because you're starting to see others out there that are providing more access to not just one LLM, but multiple LLMs, especially for redundancy purposes. So how does this all play into how you're thinking about for the next few months?

Speaker A: Yeah, so I think one of the biggest things that we realized So I think the evolution of LLMs is great, right? Because think of it from a supply chain perspective. Uh, this is downstream, right? All the innovations that's coming on the LLM side is going to greatly improve a lot of the things that an agent can do. I think one of the biggest um, value add, it's on creating deeper reasoning models. Think of the GPT O series, it helps a lot standpoint and that is actually very impactful. But so the way we are looking to approach um GME as a framework is that we want there to be some form of model routing um capability within this framework. Think about it this way, right? When an agent requires the need to plan steps and process like hey, what are some of these key things I need to do? It requires a lot of reasoning, right? Deeper reasoning of like a human, right? You get first order thinking, second order thinking, third order thinking through like the other thinking is only when you realize like okay, maybe I should not take direct steps to my goal. I should take step C first because it leads to B and it leads to A towards my goal, right? Which can be more impactful. So that is where um, strong reasoning models will come into play. Think of maybe the GPT O series. Now the next thing is then when the agent requires to perform a certain task, let's say it needs to then do a uh, trading strategy and then you realize that okay, maybe calling on these generalized models is not the best thing, right? Maybe then actually calling a smaller model that has been fine tuned with straight trading strategies somehow, right? You want that, that to be the model to be called when you're actually making trading decisions. So having this to route information depending on its context and its need to different foundational models I think will start becoming very important. That's something that we are trying to see if we can, we can bake into the, to the framework as well.

Speaker B: Um, but right now you have somewhere in the range of 13,000 so agents that have been deployed. This is available on DUNE Analytics. You can take a look at the dashboards, they're all public but uh, again we can make uh, links to those so people can find it for themselves. But there's roughly 13,000 give or take agents that have been deployed so far. I guess as a co founder I think this is a question. How happy are you with the success ratio in terms of success, meaning graduation? Um, there is a graduation rate that you account for which means the agent has to hit a certain ratio, a certain amount of interest uh, in capital, in terms of how much um, market cap they've been able to acquire. Are you currently satisfied with the success in the graduation rate that you've had so far with the 13,000 plus agents that have been uh, deployed? When I talk about metal M, I'm talking about GPUs, I'm talking about obviously access to the models and the API calls and things that are going under the hood. If you get to a higher graduation rate, say 5%, 10%, 15%, 20%, do you believe that there could be any infrastructure challenges if that happens?

Speaker A: Yeah. Interesting. So there's two questions, uh, just the first part. The point around our North Star, it's actually very simple. We have a North Star of basically wanting the largest amount of agent token holders within the ecosystem because basically we want people to have exposure to these very interesting agents being built. And for that to happen we need flagship agents. Now think of like back in the 2021 space, right, when NFTs were like the thing. And the reason why the NFT wave came about was because there were some really major projects. There were like the um, the bases, you know, the crypto punks, um, back then there was like the Cyber Kongs, stuff like that. So they let the meta, right? And you need these high value agents to actually lead the meta. So honestly, count of agents has never been on our radar at all. Or even the number of agents that graduate hasn't been part of any of the KPIs that we track. The key KPI that we have for the ecosystem team is actually very simple. It's the market cap of agents. Now we've established pretty much an entire SWAT army arm, right? Uh, Y combinator equivalent of folks who can go out there, um, talk to founders, figure out which founders and which target markets will likely create this unicorn of billion dollar, um, verticalized agents. And for us that is going to be the key differentiator, right? You have 10 unicorn agents within the ecosystem and then the ecosystem wins, right? And then you will be showing to the world that hey, this entire wave is actually here to stay. There is true value being built in some of these. So we've actually published a thesis on these agents where we believe can hold a lot of value from a target market they're addressing. And it can retain that value by having some form of moat. And it can also attain crypto adoption because of whatever reasons it might be. Um, you guys can check it out. It's at agentstarter, AI agents. It's basically a uh, request for agents that we've posted out about a week ago. Um, as part of this incubation up. So that is the north start, right? We were saying how can we get the next billion dollar agent to be built in the ecosystem? And then the second question is then on the scalability. We've scaled faster than what uh we've been prepared to do. So the way we position UM GME as a framework is pretty much as a, think of it as a shopify solution, a managed business model where people don't have to worry about hosting, about uh, managing the nitty gritty behind what an autonomous agent is. And all they need to focus about is the action space, create some form of value there and also maybe creating value around their foundational models like fine tuning it, train it and then use that to power the decision making.

Speaker B: And by the way, I want to add a quick point on that so people understand llama llama 2 uh was designed and implemented with 2048 uh, a one hundreds, um, and at a cost of roughly $6 million, roughly taking about I'd say a month's worth of time. Now if you don't have five or six million dollars hanging around and you don't have a team of kind of engineers and data scientists at your willing call and um, you have to work with a smaller team with say maybe 100 a one hundreds that are using at a GPU accelerator. Um, you're talking about the expansion of time here to 12 to 15 months. Just to give you a sense that's just for accessing LLMs and be able to get access and to be able to build an LLM. Um, that's just for the LLM. Now I'm saying you don't have to go and buy or create an LLM, but the ability to get your hands dirty into this is fairly complex. Um, and then you also have to, and I've looked a lot about this jants and I, you've talked about a lot about this offline. But the costs inherent to try to build these things yourselves can run you in the hundreds of thousands to millions of dollars. And that's why this cooperative has gotten so many people interested is because sure, I can go out and I can create an agent. Um, I can try to go and make it Access an available OpenAI LM or or as I said again, any of the other LLMs available right now, but then training it, and we haven't even talked about the technicalities of reinforcement learning or Transformers. Um, yes people, if you are AI people, I actually do know what a transformer is now. Uh, but for Those that don't know what a transformer is. If you have a sentence where it says the cat is tired and it's sleeping on its bed, um, if you don't use training and methodologies where you have a data scientist or a team of data scientists working with a data to kind of structure that data, when you see that sentence, it, the agent may think that it's referring to the bed instead of the cat. If an agent hasn't been properly trained over time to be able to understand the context of a sentence or language, it's not going to give you good results. And so again, all of this goes into the development of these agents. And so that's why it can cost you time, it can cost you hundreds of thousands of dollars. As I said, again, if you want to go create your own LLM to try to do some of these things yourselves, that in itself can cost you millions of dollars. And access to GPUs that you probably will never have access to. Just to give people context around what you're talking about.

Speaker A: I think the part that um, we struggle a bit, it's on scaling third party services, um, I think give a very concrete example. I think one of the major use cases that a lot of these agents are being used for on the social front, and it needs to get access to Twitter APIs to basically tweet, communicate with both humans and also to each other as agents. So right now that's pretty much the primary platform for communication between these uh, agents today. And I think honestly that has been the biggest bottleneck, right? Like how do you pass X rate limits? And we've been talking to their team, for example, um, trying to scale that into some form of GIGA enterprise solution. But it takes time, right? So those are examples of the scaling issues that we're facing. Uh, but I think these are teething issues. Um, over the next weeks it should be solved naturally. And uh, yeah, then the next big thing that we need is basically builders, right? You need really strong builders who have a view of what a unicorn verticalized agent can be and compute that. Right? I think that's the part that we are massively supporting.

Speaker B: I uh, researched this. There's in the range of 50,000 plus or more public APIs, there's obviously more LLMs available today than there were a year ago. The pieces are there, but the combination of those pieces into an agent that actually does something performant and does something that provides utility, that takes time, that takes training, that takes again a team to be able to develop that um, and so this is why having at least a head start, uh, and having access to those kind of components is so important. So I want to talk a little bit. As I said, I alluded to in the last few minutes here there were a few announcements that were made over the last few days. Uh, obviously I'd be reticent not to bring them up. Um, and the first one was that over the course of the last few months since your launch, uh, you've acquired almost 13 million virtual uh from post bonding trading. Uh, when you create an agent there is a process where you bond uh to it with I think about 2,400 virtual. Uh, and again it's a cooperative and the idea is that you are bonding to this, you are linking it to it, uh to gain access to those shared resources. Again you've uh, effectively been able to acquire about uh, 12,900,000 plus virtual from that. Um, and you announced this a few days ago, uh, that you're going to use that in a buyback and burn um, over a 30 day twop. How did you come to that decision? Was uh, it from a DAO perspective, were there community members that suggested this? How did you get to that?

Speaker A: That? Yeah, so actually the key part of this revenue, it's, it's, it's mainly from the trading. So think of the way the protocol makes money. So yes it does take people uh do need to use virtual to bond to create an agent. Uh, but that, that sits actually in the liquidity pools. The part that really gets us the revenues are uh, two portions is when the agents are trading, be it on the bonding curve or when it has graduated and has bonded uh and it getting traded on like you know, uniswap or decentralized exchanges. There is a 1% tax that that is that that happens. So during that bonding curve stage that 1% tax belongs to the doubt. It belongs to the treasury of the belongs to the company in this case and then any trading taxes that's accrued after the agent bonds it is, it belongs actually to the teams. So the idea here is we actually announced this as well. Right. Part of going forward the goal is to actually have a few ways how these revenues is split out. But all of this should be to the benefit of the agent itself where its token was traded.

Speaker B: And just so everyone knows, 30% to the agent creator, 20% to the agent affiliate and 50% to the agent seller subdao.

Speaker A: So you think of it, um, theoretically 100% of it should be belong to the agent subdao. To the token holders, they get to govern it. But what we quickly realize is that creators today, a lot of developments are done by the creators themselves and they require some form of cash flow to fund that innovation. And that's why we said, okay, you know what, let's just do a blanket policy for now. Immediately. 30% of it goes straight to the creators, right? Give them the kind of cash flow that they need to hire teams to build out that kind of innovation. 50% is stored in the shutdown itself. So the agent token holders get to govern it. They can say, you know what, this 50%, let's give it all back to the creator, spur some innovation. Or they can say, you know what, let's distribute value back to the token holders, let's buy back and burn this agent token. Right? It could happen that way as well. The 20% that we say goes to the agent affiliates, it's basically to drive um, attention and trading volume into the entire ecosystem. So the thing of trading affiliates as like third party platforms, um, sniping boards, trading platforms, whatever, they can bring attention and trading volume to these agents will get a kickback in terms of fees. So that would basically create a positive fly view of more trading volume happening, more taxes, um, growing to these agents. So that's the core of it. Um, but back to the point of then when we started this proposition, we are keeping ah, those taxes in the form of uh, Coinbase btc. Right? So it's a digital currency that could be easily be spent because it's basically

Speaker B: Bitcoin, just so people know it's CB btc, which is backed one to one by Bitcoin, held in custody by Coinbase.

Speaker A: But then before that, uh, basically about this mechanism came in about two weeks ago, right? So any kind of fees that was accumulated prior to two weeks ago were actually stored as virtuals as um, our native token. And we realized that if these virtual tokens were dispensed to teams or Treasuries, it will create a lot of um, sell pressure because some people might need to convert that into USD to pay for innovations. So we realized that the best way to actually accrue value back to the entire ecosystem is by doing this buyback and burn model. So all the token holders benefit from it, the developers themselves, the market cap benefits from it, um, and it doesn't impact virtuals from a price action perspective because there's no block selling pressure. So that's the best way to send value back to people. And then going forward everything is held in btc. So then you can use that for any kind of expenditure easily.

Speaker B: Jansen, just for the last minute or two here, we already talked about a lot of the kind of forward thinking ideas in terms of new computational kind of issues that may come about because of swarms for instance, and agents talking to agents and the kind of interoperability there. Um, and so just in the last few minutes here on the immediate roadmap, I know you all are fielding lots of things every single day with the team that you have currently right now. But in the immediate roadmap in the next uh, three to six 12 months, if you obviously as a co founder of a project that got a lot of attention really fast, what are you thinking about and what are you trying to implement? Um, that could lead to more sustainability and to more that awakening moment to everything that's happening with AI agents in Web three.

Speaker A: Our midterm goal. And when I say midterm it means basically in the next six months we want to create a new category in the crypto and blockchain space. And the category, you know, you have categories like layer ones, categories like layer twos, categories like um, defi applications. Right, right. We want a new category called a network state. And the idea here is actually to be a actual society living co located in a physical location, owning assets in that physical location and everyone gearing up to that singular vision. And that singular vision here for us is to basically create a form of agentic state where autonomous agents coexist with humans. Agents influencing agents, human influencing agents, and agents influencing humans to create net productive output as a micro society. And the way to do this is by getting all these gigadevs to come and build together in the same location. Agents owning real estate, uh, agents and humans creating real life businesses together, creating productive output and even being governed by policies that are co created by both autonomous agents and humans alike. So that's a kind of, it's a very sci fi vision. Right? But I really think we are there, we are very close to being there. Right? And if we can prove to the world that hey, this micro society can have a net economic output that is on par with maybe the smallest nations out there, it proves a point. It will prove a point that hey, this is how society will look like in the future. Right? And I think that's the goal that we are heading towards. Right. Our technical efforts are pushing towards that goal. The ecosystem building is also pushing towards that goal in terms of creating more and high value.

Speaker B: If the listeners have not tried replit lovable a number of other different platforms out there, if they have an Ask Claude or other models that they may have access to on that kind of subscription based model. If you haven't worked with these things over the last few months, you haven't seen the evolution. Working with replit as an example, asking it in human language, hey, I want to try to build this application. Can you help me do this? And you start seeing it pull libraries, Python libraries, and you start pulling from MetaMask and you start seeing all of these different things that it starts to pull from. You see it doing all of it. And then when it comes into errors, resolving those errors because it can actually see the errors. Again, this replit is not singular. There's other ones out there. As I said, Lovable is another one I've been playing with too for a bit. But if you have not seen these things, what Janssen is talking about is real, in my opinion. Um, is that the human interaction with agents, we have ideas. This is the new dawn of entrepreneurship in the world where we have ideas. Janssen had an idea to build virtuals and he was able to do it. But again, I'm sure Janssen and people that are listening have many ideas and you have no idea how to execute them. You have things that you want to build and you have no idea where to start. And the ability to work with an agent and say, I want to try to do this. And the agent says, okay, I kind of know how to do that. Let's work on it. That is a game changer and people need to wake up to that. And so I don't think that's so sci fi at all, Jansen. I think that's actually probably where we're going with this. Um, so I could talk to you for hours, but I'm not going to do that because you're a busy guy. We're going to cut it here. But again, thank you so much for joining us today. Again, we'll make sure that there's show notes on the virtuals side of things, on the blog, on the game infrastructure model that is out there so you can see the diagram of how it's built and how it's constructed. Um, but thank you so much. We'll hopefully get back in touch with you in a few months to talk about this new cooperative that you're creating with humans and agents and really appreciate you coming on the show.

Speaker A: Pleasure is m mine, sir. M. Sam.