MIT Prof. Ramesh Raskar: The Internet of AI Agents: Why AGI Won’t Be One “God Model”?

Sa. We are going very fast right now. We're building things and if an agent makes the wrong calculation or makes a poor decision, then who is to blame in this environment, you know, when you're working with a junior employee, in the beginning, you know, you have a lot of oversight, but over time you give them autonomy. It's the same thing with Asians. In the beginning we're going to be obsessively observing them and controlling them and have some rule based systems as opposed to full fledged autonomy. But over time we got to give them more autonomy because that's where we'll see this new emergent behavior. We'll get these amazing things out of them. So I think there's no right answer to who's responsible for what. What if agents delegate to other agents and so on and so on. Because I might launch a million agen for myself to do all kinds of different things. And those million agents could launch billion agents each to do different things. Cool. Okay. So to kick things off like I love, start with your journey. Origins of your story. You grew up in Nashik. Your father served in the military. And then your family had deep roots in the community. And somewhere in that environment you developed incredible curiosity about technology. What's possible that led all the way today. And then I read that watching Jurassic Park I think is and seeing those special effects spark something in you about computer graphics and imaginary life as well. Can you take me back to that time? Like what was it about that moment or that period of life that made you think I want to be part of building things like that? Yeah, I mean, I come from a very small town, you know, well outside Mumbai, about 200 km out. And most of my extended family is just farmers in agriculture. But you know, when you're in the farm in the agricultural world, you realize, you know, nature does not negotiate and just sort of figure out how to make it work. And you know, my father was not a farmer, but rest of the family was. So I learned a lot. And I think the adversity of just being in an environment where, you know, there's curiosity but there's not much else you can do. You know, I just spent a lot of time in the library, you know, the city library. And I would just read books. And after I did my engineering, I did very well and you know, started doing my PhD. This Jurassic park came around in mid-90s. I was like, wow, this is so amazing. I want to build that. And I love geometry, I loved math. It's just the right combination of how to create special effects using computer code. And that's what got me into the field. And since then I have been mesmerized by pictures and always just to make fun of my colleagues in other fields of computer science. You build this whole thing. At the end you just get a number on a command line and you know, something is working. But in my field in computer graphics or computer vision, you know, we see an image, you know, we saw something that's, that's actually working. So it's, you know, instant gratification. There's something real. Yeah, something real. So I was very, very, very attracted to the field of computer graphics and computer vision. That was Jurassic park in the early 90s. But then I saw South park, late 90s, and I was like, actually the south park is as entertaining as Jurassic park and it doesn't use any of those fancy special effects and big machines. And when it comes to storytelling, actually maybe I'm in the wrong field. And so abandoned computer graphics as a field and started moving towards machine learning. That's. That's really cool. So it's like, what, what in south park really made you think about that? Like, is it more on the storyline? Is more on the, like, how the characters were represented? Yeah, I mean, if you think about South Park, I mean, of course it's a cartoon. Even the shadows are just circles, you know, and the interaction between the characters is not even realistic. They don't even look at each other, they're just talking. But it's hilarious as hell. It's so amazing. And so realized, I think the human emotions don't need photorealism, they just need functional realism. And that really kind of woke me up. I was like, I don't want to spend most of the rest of my life just creating better entertainment and better games, which is the main field, main applications of computer graphics back then. And Nvidia was like a brand new company. And a lot of my colleagues had just joined Nvidia and they were very busy building the technology to create even better games. And it just didn't feel like the right thing to do. Of course I was wrong because Nvidia ended up doing many other things along the way beyond creating games and entertainment. So that's a very contrarian opinion, actually, because in a way, everyone goes to one direction, you chose to go to the other. Right. And then when you made that decision. Can walk me through what was the next step after that? Yeah, I think so. First of all, I thought the problems in computer vision and machine learning and human computer interaction are really difficult. And for somebody like me, I think if there's a really difficult problem that's very exciting, I really get hooked. So intellectually it was very, very, very exciting. And also when it kind of looked at the space of problems you can solve with, felt like you can solve anything, space imaging to transportation, to agriculture and so on. So it just felt like the whole world is your lab when you're working in a field like that. So that was very exciting. And then after I got my PhD, this was around the dot com, so all my colleagues were joining Yahoo and AltaVista and Google and their stock options were getting worth millions of dollars. I was very confused. I was like, but that doesn't excite me at all. I really want to do something that's deeply scientific. So I ended up joining Merle, which is a research lab right here in Cambridge. Mitsubishi Electric Research lab, which is called the Bell Labs of computer vision and human computer interaction. Some of the smartest people here. And I ended up when I was interviewing, I asked a lot of people like, why are you guys not leaving this role and joining one of all these tech companies? At that time they were going like 100x overnight, their valuations and going IPOs. And I was like, why are you not doing that? And people said, no, we really love what we are doing here. And that was a good indicator for me that this team, this group of people at this lab, at this lab called Merle, is the place I want to be. And remember, they had no stock options, they had a good retirement plan, but they had no stock options. And I still chose that. And I think throughout my life I have always chosen something that's really deep, meaningful and something I can be passionate about than be driven by money or instant fame or something like that. And that has always played out for me. So it's in a way the problem solving very difficult problems where the others shy away from solving versus going into where the money is at, which today is actually with the AI arms and everything. Yeah, I think money should be like a secondary goal, clearly. But if you can find the right intersection of an amazing problem and an amazing potential impact, then that's the best. That's the best. I mean, often tell folks that you need something for your mind, something for your heart and something for your wallet. And often you do three different things for it. You're very lucky if what you're working on satisfies all those three. Your mind, your heart and your wallet. But it's perfectly fine if the things I'm working on. I have to work on three different projects to satisfy my intellectual curiosity. Something that makes me feel good as a human being, but it's also good for my bank. I think there's the Japanese concept called ikigai that has the three Venn diagrams and that's exactly how you put it. If they meet all, then that's the magical. That's the magical combination. So right now the entire AI industry is an arms race, especially for larger centralized models like OpenAI, Google's Meta. I would like to point now, because you worked in computer graphics, you worked in cameras, you worked in robotics, especially your work in MIT Media Lab and along with Facebook and Google as well, including the healthcare aspect of things too, right now you're advocating for decentralized AI and that includes championing split learning, keeping the data at the edge, especially on the edge devices. So what do you see that everyone else is missing right now? Especially the same exact moment. Everyone is rushing towards the open AIs of the world, where the money is at. But you see something else right? Right now. Walk me through that. I mean, if this debate between centralization and decentralization is an interesting one, at the same time, to me it's kind of obvious. You have kind of the foundry phase, the foundry era, then you have kind of the garage era, and then you have kind of the bazaar era. That's what happens in every field. If you think about computing, you had the mainframe era, which I would call the foundry. Things are forged in some central entity, central organization. Then you have the garage era, where people start tinkering, they start creating their own things. You know, Silicon Valley really celebrates the garage culture. So you had like these computer companies and, you know, device companies that came around. But once you start building in your garage, you have a lot of people are empowered, you know, to go ahead and build their own versions. But then you get into the bazaar era of it. You know, things that people have built in the garage start getting connected to each other and a whole commerce emerges around that. And that commerce expects you to create a whole new set of technologies. So if you go back to the, you know, the mainframe versus PC versus the Internet, you know, you had these big companies like DEC and Cray supercomputers and so on. And then we had a bunch of PC companies, and then the PC companies were also sidelined by the Ciscos of the world or the sun microcomputers of the world. So I think it's very important to realize that it's always a natural evolution where you get centralization, then some decentralization, and then you get the network that emerges because of the decentralization. So I'm a little bit surprised that people are staying in this foundry era where things are forged for so long. And I think slowly we'll be moving to the garage era where most of the intelligence will move to the edge. And alp, you and I can create our own AI. And that's what's going to be very, very exciting. We can train our own models, we can choose our own architectures, we might share notes with each other and so on. That's kind of the garage era and then the bizarre era when our agents start doing interesting things that are out there communicating with billions of other agents and micro AIs if you want to call them, is the exciting phase. So does it mean that OpenAI and everybody is doing wrong thing? No, not at all. Because they're in that foundry era where everything is centralized. The data is centralized, computer centralized, talent is centralized, $100 million salaries and nuclear power plants and all the good stuff. But you need to do that in the beginning. There's nothing wrong with it. But if you're thinking three to five years ahead, then you have to start moving to the garage era and start saying, wow, what if AI now runs on the edge? What would that look like? The same companies that we're talking about, so Nvidia has released its Nemotron and they're very proud of that. Microsoft of course has the 5 Series of models and of course the Chinese are doing really well in creating extremely compact and small models and so on. So I think things are already going in that direction. It's very interesting because in a way right now we do have this ownership versus renting and we rent the data. We are okay with giving our data away. And then only thing that maybe we own from the tech is that we have the phone in our hands, but most of the compute is happening elsewhere. So in a way, how do we take it from being elsewhere to convince people that it is in their hands and especially renting intelligences is also extremely important. It's like. And there are many different intelligences, right? Like based on what you were saying, if everyone has an AI open source model or closed source, those could be fine tuned, trained in a certain way, then we will have many shapes of intelligence that lives on device like five years ahead from now. How would that world look like? Yeah, yeah. I mean it's, it's kind of absurd, right? If you think about it, that right now we pay to rent those models. We don't own it. At the same time we give away our data. And that makes those models even better. Yes, it's kind of a bizarre world we live in. And I think what's going to happen is once we realize that we can run the same models, not just run, but actually train our own models on our own devices, the only thing we pay for is electricity pretty much. And this has many benefits. First of all, it's completely private. Second, the latency is going to be very, very low because it's running locally. And the third is that there's going to be a rapid innovation because once people own their own micro AI models, they can innovate very, very fast. You know, and I think that's the key because like some people would say, you know, the scaling laws, remember the scaling laws, and the best stuff is going to be in this. But you know, scale is not about intelligence. Scale is about compression. And I would say the next frontier is not bigger, but the next frontier is about being closer. It being closer to your data is being about, closer to your context. It's about being closer to the people who actually need it. So the frontier is not bigger, frontier is closer. And once we change that mental model that actually creating models that serve us and so on. And as you said, I mean, maybe in the beginning we're just going to take existing models, maybe some of the open source models, and we morph them and we fine tune them. New algorithms are emerging as you know, that exploit the rank, exploit the algebraic properties. Maybe in the beginning it will start looking like a small modification of existing philosophy. But once everybody has a large population, has their own compute, their own abilities to train, we will get a lot of heterogeneity and different kinds of models and the ideas that we have seen being explored just in the research fields, like spiking networks or different types of feedforward or combinations of mixtures of experts. You'll see a Cambrian explosion of very different kinds of models that are not just based on deep learning or transformers or you know, existing reasoning. You'll see this whole new vocabulary emerge because people on the edge, you know, are inventing and innovating on a daily basis and that, that really democratizes it because it's, in a way we, we do forego some of our ownership of, of that intelligence, but also it limits the, what we could create, even the accessibility of it too. Like for example, maybe in, in the US we have very good bandwidth and very good, a lot of Energy can use OpenAI's models or train our own by using a GPU. Fine tune our own using a GPU. But there are also parts of the world where there's absolutely no access to Internet or resources for them to leverage. And I think that's a huge thing, like not having Internet and being able to use AI. So I want to talk a bit about Agent Zero that gives about could give a private AI to win to the citizens. So I would love to talk about a little bit of that. How much of that code is right now is digital reflection of the human coordination and also being independent. Yeah. So let's step back a bit to see what we're talking about here. Right about three, four years ago, we started talking about an agent for every citizen, an agent for everybody. So instead of AI for all, it's AI agents for all. And this was before ChatGPT and so on. And people just started looking at us funny. They're like, that's not possible. We had to spend billions of dollars to host this model. So how would you have for everybody, an individual model? And I think we're almost there. We can probably run it on your laptops, on your phones, or have a mirror in the cloud for it, but nevertheless, something that's unique to every individual. So we call it Agent Zero, kind of the ground zero of what you would do. And the concept is similar to how in the mainframe era, kind of, you know, Bill Gates and Steve Jobs and a few others started saying there should be a computer on every desk. Which was a pretty ambitious statement back then. And this one's kind of similar. An agent for everybody. It doesn't matter how poor you are, doesn't matter how poor your bandwidth or compute is. It doesn't matter whether you live in a democratic country or you live in an authoritarian country. Everybody can have their own AI agent that's secure, safe, authenticated and private and ethical. I think we can get there. And the moment we do that, it's going to unleash a completely new kind of creativity. Because I think centralization creates consumers, but decentralization creates owners and creators. Yes. And you used it to. So Agent Zero, the work that you've done there, I'd love to hear more about it. Like, how would it be executed? How can it be actualized so users will still have a phone and then they'll be able to download this app and then it will run locally on device, on the edge, without data going into making API calls and such. Yeah. I think smartphone is a reasonable starting point. But you could imagine it also maps to other parts of your digital life. It could be through email, it could be through WhatsApp, it could be through your messages and so on. It's like the movie her, you know, so, you know, Joachim's character in that movie. I think the reason why people like the movie is it doesn't require a whole new form factor, like some kind of a fancy headset or a chip that's implanted in you. He's just using a phone, you know, and he's using the UX of the phone. But the agent, you know, Samantha in the movie is interacting with him in multiple ways. Sometimes it's coming through an email, sometimes it's coming through the phone and desktops and so on. Right? So I think it's the same thing. Now. You don't want to kind of anthropomorphize agent and say it's running like an app on your phone. You know, it's diffused in your life. You know, it's collecting information about you from multiple sources and, you know, interacting with you in, you know, with different surfaces as such. So kind of on the technical aspect of it, yes, it might look like an app, but it might also look like an email exchange in the beginning. But over time, we don't know how it could evolve. Over time, we might start seeing completely new forms of devices that maybe look like phone, maybe look like a pendant, maybe looks like a traditional desktop. I don't know. Or maybe there are multiple of them, but then clearly most of it will run on the edge. But there has to be some kind of a replica in the cloud as well, because then you could lose your phone, for example. Also you have to authenticate yourself. And then if my agent has to talk to your agent, there has to be some equivalent of a telephone exchange. Just the way when my phone calls, your phone, even if you're sitting right next to each other, there's a massive ecosystem that's built around giving me a unique phone number. I have my own carrier, I talk to my own towers, you know, go through some router. You know, my. My voice gets encoded in some packets and it comes back to you. A lot happens to build a telephone exchange. And the same thing here, which is, you know, in the beginning, we might be just building AI and AI agents kind of for, you know, in the garage world, but in the bazaar world, I need to figure out how they all talk to each other. So we think the agents are going to go through three phases. We call it agent foundations. Agentic commerce and agentic societies. So agent foundations is, you know, what's authentication, what's the trust, what's the registry, you know, what's the certification and so on. And the agent ecommerce is about, you know, how does agents, you know, spend or earn money, how do they find jobs, you know, will there be a stock market for Asians? How are we going to do shopping? And agentic societies is where we think these agents, you know, will create their own organizations, they'll create their own companies, you know, they'll create their own justice system and so on. And that's very powerful. I think, like in a way, if we think about if the agent is the intermediary between us and the Internet, right? Then do you think that Internet is slowly going to. Slowly, maybe there won't be that static website sitting there, but there will be just this web network of agents independently operating. And then when I want to show up, I'm not going to go to Amazon.com or somewhere, but I'll go to my agent and my agent will find it for me and then bring it to me. So like, in a way, apps would and web will render that agent self. Is that kind of the future? I mean, from a consumer point of view, that feels like utopia. Like my agent is like my assistant and doing everything. Not just gathering information, but actually, you know, doing some action. But if you look at, from a business point of view, there's going to be a lot of friction, right? Like so, for example, if I just talk to my agent and say, hey, can you get me an Uber in, you know, on this street in Cambridge and just Uber shops open five minutes. That's kind of cool. But the company Uber doesn't like it because they would like me to use their app because, you know, they would like to upsell. Maybe instead of getting, you know, an uberX, they want me to send an Uber premiere or while I'm waiting for it, they want to show me some ads or use the data to get a flywheel to see how they can extend to more businesses and so on. So Uber wants to maintain the relationship with me and other customers. It doesn't want to get intermediated through some kind of an agent ecosystem. So I can see this playing out for quite a bit where companies will resist direct interface to user agents and they would like to create some friction so that you are constantly aware of what's going on. Same thing with shopping. I mean, the number one rule for shopping is you shouldn't make it so frictionless. That People say, oh, I'm going to look at it for some time but I'll come back tomorrow because it's only going to take me five seconds to make a decision. You know, there's a sweet spot in shopping where you have to make it, you have to add a little bit of friction, maybe 5 minutes, maybe 10 minutes, but no more than 15 minutes where a user has to spend some time before they can buy something. And by doing that, seeing new things. Exactly, experiencing it. And so I think it's the same thing with agents. And I mean the honest question is what are billions of people going to do if everything is automated? So I think the businesses and even the population will like the fact that things are not as easy. There's a little bit of friction, they'll spend a little bit more time. Now what the UX of that will look like is going to be very fascinating. It's like when you call toll free numbers, you have to listen to some kind of music or some kind of an ad before you can talk to somebody. So will my agent say, hey, I can do this for you, but you, my master will have to listen to this stupid thing. The agent can do something. I think some really fascinating features are going to emerge like play an ad before I show you the answer. And the human has to listen to the ad before the agent can do something for you. So I think the business models are on that. But I think if you're a new entrepreneur, you're an investor, a lot of people are telling me this sounds like 1995. You can take any idea that's already well established in the real world and then map it back then to the web World in 1995 and now you can map it to the agentic world. It's going to be fascinating next two, three years where we'll see many existing businesses becoming agentified. At the same time, agents will create completely new forms of businesses that we cannot imagine right now. So for example, you could get agent repair shops saying hey Alp, you're agent a little bit rusty, you want to send us a repair shop. I know, and they'll be a whole agent repair shop. Or at some point your agent will start not be as smart. It's like, hey, this other agent went to MIT or this agent went to Harvard, like what's the best university? So it'll be Asian tech universities. And you might send your agent to learn at these universities and at some point your agent might misbehave. It's like, you know what, I need to Instill some values. They'll make an Asian tech church. Agent has to go to this church and get kind of influenced with good values and so on. So you could imagine there'll be this whole new set of, you know, a cottage industry of new jobs, new economies, new workflows, new platforms will emerge that we can't even predict it. So in a way then like intelligence, which is true for humans as well, like we do as we read more, interact with good mentors and like good teachers, we get better and then we learn in a way intelligence do multiply. If multiple agents do the work together versus then having one like AGI type of an aspect, like it knows everything 100 very good. Yeah. So then it sounds like over time, as agents go train with each other by interactivity, they'll raise the bar. Absolutely. And this goes back to Marvin Minsky and the society of the mind, which is, you know, it's not about one super intelligent brain, even in the human brain, human case, but it's actually the multiple minds for our memory, for our executive function, for planning. All these different pieces of brain work together. And that's what we call intelligence. And I think if you think about what has happened over the last year, year and a half is very quickly everybody realize actually scale doesn't help because a single model is too difficult to manage. And so we are going with all kinds of shortcuts. And nobody's admitting the fact that you cannot train one large model. The way we're kind of getting around that argument is like, actually it's better to create a mixture of experts. Each of them have a little bit of their own expertise and they can work together. That's like society of the mind. Or they're saying, actually the way to do this is take one problem and break it down with some kind of reasoning and planning and do a bunch of tasks and do some kind of an amalgamation of that. So if you think about it, the centralized model are already moving towards decentralization, but we don't use that phrase. So the fact that with a mixture of experts, although experts are running on a single machine, or they're distributed all over the world, those experts and they work with each other to solve a problem is the same thing. Or the fact that when you do reasoning, you have to break it down into this workflow, those individual components could be running anywhere. They don't run a single machine. So we have already started giving up on this notion that there'll be one massive model that knows everything about every question and every task. And we have already moved away from that into the society of the mind and some decentralization. And I think the natural progression there is that every one of us will run our own expert agent. And when somebody wants to do a task, they'll just break it down into this 20 step or 20 experts saying hey, where are these 20 experts? Who can I talk to? I'm not going to run them on my own stupid machine because I have a sub billion parameter model working on this 4 watt machine and can I call you guys to get what I want to get this done? And then in real time it finds this 20 other agents, which is like the mixture of agents. And then things get done and alp. I think the exciting part for me is that that's when a new form of capitalism emerges. Because now these other 20 agents had to keep up their game. They had to continuously update their own agents because you know, one is good at coding, one is good at legal, one is good at health, one is good at, I don't know, accounting and so on. And all of them are trying to create the best accounting agent, but nothing else. They specialize in that task so much. And then there's the commerce aspect of it too, because then they could charge for that task. Exactly. And now, now we're talking about like is the accounting agent worth, you know, $5 a query or is the legal agent worth $500 a query? And why? So a lot of the research we have done at MIT is what we call knowledge pricing, you know, pricing the data and pricing the models as well. And so that becomes very fascinating because now we are not talking about behavioral economics, we are talking about this agent economics and so on. So that's very interesting. So what did you find there? Because it's in a way it's very similar. You can have agents that maybe do very good accounting, but that charge a lot. And then there will be agents that do poor accounting because they're behind in their training or they didn't use the most specialized model. That's right. It's like the stock market, right? There is intrinsic and extrinsics, there are fundamentals and there is market forces. And the same thing with an accounting agent or a legal agent which is like, hey, this legal agent was trained by somebody at Harvard Law. Whoa, sounds pretty good. But can I really probe and see how good it is? And to your point, maybe the legal agent does really well in the daytime when the nighttime they run on some cheap compute. So maybe it's not as good. So there could be pump and dump schemes where the agents kind of change their color and their accuracy and their quality, you know, from a minute to minute basis. So how can you have real time validation saying hey, you're charging $5 a query. A few minutes ago, you're still charging me $5 a query, but it's not as good as what it was. So you need to build a whole infrastructure. Think about like the FICO score, where a FICO score, you know, your credit score is not dependent. You don't decide your credit score. There's a whole sequence of transactions run by hundreds of other entities. Whether you travel, whether you eat at a restaurant, whether you credit card bill, that's all changing your credit score. And the same thing is going to happen with agents where their score is going to change, their FICO score is going to change on a minute to minute basis. And so all of this has to be built and baked into the infrastructure of the Internet of AI agents. And in the Internet era, what ended up happening was if you talk to Tim Berners Lee, I had a dinner with him a couple of months ago, they made a very conscious decision back then to only work on problems that has enough consensus. And so that's why they left open many of the pesky issues. So for example, they didn't worry about payment protocols, they didn't worry about having, you know, ssl, the green padlock that you see for HTTPs, all the things very contentious. So they just kind of left them and they made sure the consensus is what they're focusing on. And that was good in most cases. But we also suffered a lot because of all the security issues, all the scams and so on. I think with the agentic web, the Internet of AI agents, we have an opportunity to bake in the security and trust into the infrastructure. So let's see how that plays out. That's incredible because at scale when imagining that then the entire infrastructure has to be maybe redesigned so that it can accommodate this. And one question that comes to my mind is again, if it goes very fast, we are going very fast. Right now we're building things and if an agent makes the wrong calculation or makes a poor decision, then who is to blame in this environment? Yeah, I mean it's like, you know, if till 18 you're responsible for your kid and after 18, hopefully you have instilled enough values in them that you'll set them free and they do the right thing. And it's like the Speed of Trust from Stephen Kofier, one of my favorite books and the key message in the book, which is all about management coaching, it's not really about AI. Of course, when you're working with a junior employee, in the beginning you have a lot of oversight, but over time you give them autonomy. And it's the same thing with agents. The beginning we're going to be obsessively observing them and controlling them and have some rule based systems as opposed to full fledged autonomy. But over time we got to give them more autonomy because that's where we'll see this new emergent behavior. We'll get these amazing things out of them. So I think there's no right answer to who's responsible for what. What if agents delegate to other agents and so on and so on. Because I might launch, you know, a million agents for myself to do all kinds of different things. And those million agents could launch billion agents each to do different things. It's like that toy, the Russian. Exactly. It's like matryoshko. Exactly right. So I think it's going to be going to be interesting how we delegate. And so the concept behind Agent Zero, by the way we were talking about that a few minutes ago, is in the beginning, or not in the beginning, but the Agent Zero is my own agent. But the Agent Zero can create an Agent one, Agent two, Agent Three, and so on. And each of them may not be anchored on a human identity, so it can do it in itself. Create and assign word to the agent. Exactly. So just like you have a two factor authentication or multi factor authentication, the Agent Zero might say, hey, I'm willing to create Agent One, but Agent One needs to continue to prove to everybody else that it was credentialed by Agent Zero. That doesn't mean agents are responsible for what Agent one does. But Agent one has to live in this new justice system that the agents have created for themselves. And I think we'll get there very quickly. There's the LLM as a judge aspect of it as well, like judging each other and then improving over time. Yeah, but the problem with LLM as a judge and Anthropic just has a new paper that released last week on why it will be very difficult to do evals. It's easy to do evals for LLMs, but doing evals for agents is very, very challenging. Because first of all, there'll be so many of them, like so much heterogeneity, and each of them will be doing things. And the point is, if you want to evaluate an agent, then you need another entity that's smarter than that. Agent to be able to judge. But the whole point the agent exists is can do something that others cannot do. So you get into this chicken and egg problem of how do you do evals on agents. A lot of our research at MIT is thinking about creating, you know, a mechanism where you're not dependent on some kind of an oracle, some kind of an LLM as a judge, but it's based on, you know, a, you know, some kind of a directed graph that computes the value of, allows you to evaluate how an agent will behave. And how much of it do you think is like AI judging AI and then, and then or LLM judging LLM to go forward because there's also the unknown aspect of its intelligence that the generation that is vague for the scientific community to understand as well. So do you think it's going to still be reinforcement, learning more on human in the loop to an extent, until we drop that agency and give it to the agent too? Yeah. It's a very valid question because if you use some ideas from blockchain kind of proof of work, proof of stake, we call this the proof of wit. Wit. And as you know, in proof of work, it's about computing bunch of zeros at the beginning of a hash. And the good thing about those kind of proofs is that it's asymmetric, which means it's a lot of work to create that hash, but it's very easy to validate that hash. And so a lot of research in proof of wit is the same thing, which is for an agent to prove its wit, its wisdom, it's a lot of work. But validating that is going to be very, very easy. Now, of course, as I said, it uses other aspects of graph theory. Right now, all of work at MIT in this space is purely digital. So we are not going into creating new environments or new RL mechanisms. But you're absolutely right. I think a whole new set of techniques will emerge that will say, hey, it's not just enough to be being in the sandbox, but we are going to evaluate a judge, evaluate an agent by effectively asking it to do a driver's license test, which is we go in the physical real world and we drive around and see, hey, are you good enough to drive this car? And so the environments for evals, as you know, it's a very vibrant research area as well. So I want to ask a little bit about healthcare and especially like we'll see breakthroughs, I think, when AI taps into the deep web of private data, more like medical records, biological Simulation population models. What breakthrough are you most excited about when it comes to health? And especially you've done some work there in the past as well. Yeah, I've been in digital health space for a very long time. And in fact, I started the health innovation team when I took a sabbatical leave from mit and I started an AI team at Facebook. And this was focused on health because I thought Facebook is the place that can, just seven or eight years ago can really change the way we think about digital health. And I'm very proud of the work we have done in that team. I think there's something magical when you have highly crowdsourced, highly engaged individuals. And, you know, in fact, in fact, I should tell you that the whole reason I got into this decentralized AI is maybe two, three months into my. My stint at Facebook. Back then, it was called Facebook. I woke up one day and I said, you know, I can solve pretty much any problem in the world with the resources I have at Facebook. And I chose Health as the first one because Facebook wasn't working in that space. What I realized is that we have a platform that has billions of users for engagement. We of course have ability to learn about what they're doing. We also have abilities to recruit, for example, for clinical trials, anybody we want. And then we're also getting telemetry. So if I'm looking at behavioral health, whether it's Alzheimer's or dementia, I can do diagnosis on the fly. And so I hired little, like five doctors. I had a lot of funding, and we scoped out the whole thing, and I hired a nice team to build that out. And what I realized, although I was doing it in a highly centralized setting, that it is possible to go and solve health care overnight. In what way? Yeah, and I know it sounds pompous, but literally majority of health problem today can be solved with one thing, which is what everybody will tell you in digital health. If you go to any conference in digital health, speaker after speaker will come on the stage and said, only if you had data, we can solve it. And by that, what they mean is, because of hipaa, because of privacy, or because of centralization issues, because of national secrets, they cannot bring all the data in one place. But imagine there is some kind of a magical ability to centralize all the data in the world for any health condition. You have infinite compute. You can compute what the solution should be. You have some kind of an incentive model where whoever contributed to that data or compute can be compensated. Then you can create an expedia Experience where you can show the patient, hey, here's a more aggressive treatment, here's what you could do. You can do all of that. The problem is it'll cost billions of dollars to just help, like one patient. So you have to kind of amortize all of that and change the model to say, hey, if I try to do this for one person at a time, it's going to cost me a lot of money. But most people want something very similar. Most patients want something similar. You know, if I order something from Amazon, you know, if I order an umbrella from Vietnam, it shows up on my doorstep in Cambridge for $10. How's that possible? I mean, there's a whole ship, there are manufacturers and all these things. And the reason is because a lot of other people in Cambridge are also buying crap from Amazon. Yes. So because of them, I get the benefits of getting. It's the same thing with AI and healthcare, because although one patient might need a solution, there are millions of others who need either the same or somewhat similar solution. So you can amortize all the computation, all the data and so on. And that's what we call population AI. And there's a whole new field of AI that's very different than what we have seen. So where we have multiple inputs and multiple outputs, and you have to optimize it for a population as opposed to optimizing for individual. And I think that's what allows you to create a completely new form of AI for health. Because I'm using the word AI. But if you just move from AI to AI agents, then every hospital, every patient, every insurance company, every wearable is an agent that's participating in this bazaar. And if somebody says, hey, my patient has this health condition, I'm willing to bet 10 bucks who wants to provide the best answer. All these agents all over the world can scramble and start bidding to see if they can solve this problem. And at any given moment, many people are asking the same question. So every minute, if million people are asking a question for 10 bucks each, there's about $10 million available to solve this problem. And over a day, it's more than a billion dollars. And then I think it's that, like there's. It's very incredible because there's the aspect of the centralization and decentralization. And I think in a way, healthcare is in a way decentralized, but locked. It's a bad kind of decentralization. Yes, it's bad kind of decentralization. So we are unable to leverage network effects. And then Share the knowledge. So what do you think would take us there? Especially from population models aspect, would we be okay to have maybe HIPAA to not have the data so closed in, but maybe be sit in a cloud or sit somewhere else or in the edge device itself? Yeah, I think that's probably the most important question, which is when data is siloed, nobody wins. Yes. And in fact, what's happening with a lot of healthcare systems is, it's even worse. Not only data is siloed, but even when researchers or startups want to work with the data, the CIOs of these hospitals are saying, I think my data is worth a lot, so I'm not going to give it to you. But they don't know what it's worth, Right. And then what ends up happening is some CIOs flip and they say, you know what, whatever we have, we just sell it for $100,000. And then others say, oh, we had a similar kind of data, now it's worthless because a neighboring hospital sold the same diabetes data for $100,000. So it's kind of become this, you know, a really lopsided market. To answer your question, imagine now every data, every wallet, you know, every EC2 instance, you know, every S3 bucket is actually an agent and they have understood what the HIPAA is, what the value of the data is. And they can start in this bazaar, they can start negotiating with each other while maintaining all the privacy, safety, security constraints. So it's like me calling up Harvard Hospital saying, hey, I need your data, what can I do? And they're like, I don't know what to do. But imagine if there's a really smart cio, CIO agent, there's like, I'm in charge of the information system here and I'm willing to give this to you for 10 cents to just do this. And you get this massive tokenization that you can start highly elastic markets in this bazaar and we can solve healthcare. And there's like, it impacts the research aspect because then you'll have more data to solve like the cancer or treatments and also the operational aspect of things. Even today, if I call a hospital and I'm like, hey, I want to get this treatment, how much is it? They're like, we don't know, you need to show up. So like, in a way, it's so fragmented that it doesn't let people contribute. Yeah, I was talking exactly. I was talking to my friends in the financial world. And if you remember the 80s and 90s, if you were to Buy a stock, you call up your agent, you know, and they would pick up the phone and they'll say alp, I'm going to buy the stock for you. And you just believed it, right? And now of course, you know, you use your Robinhood and you have a dashboard, you have predict chain engine. So healthcare needs to move to that equivalent of high frequency trading, you know, where if you need that information, it's like for five bucks, I just want this now. Yes, yeah. And it's there for you. And the rest of the world is kind of, you know, in this bazaar is kind of taking the $5 and selling to somebody else for, you know, $4.99 and they're taking their cuts and on and on and on and through. That is going to come with new form of intelligence because you know, you probably heard the UCP protocol, universal commerce protocol from Google. And then you know, Coinbase has X402 and so on. But those are syntax protocols, you know, which is great, but that's just syntax protocol. What we need is system level protocols that allow two healthcare ecosystems to actually talk to each other and negotiate. Hey, in return for getting diagnostic data, I'm willing to do this and so on. You need system level protocols, but you also need global level protocols where there may be markets for hospitals putting their data online and there could be collective bargaining. They say, hey, we're going to create a live team of these five different clinics or doctors or hospitals. And when we come together, what we are sharing with the world is even more valuable. You need this syntax level to system level to the global level. A lot of the research for us at MIT is creating and architecting this bazaar of AI agents. If it reaches there, it will be, I mean it will require a lot of infrastructure build up and then to be able to support it at scale in today's world. Because there's also the centralization happening at infrastructure level too. Especially so reliant on Nvidia GPUs. And then the TPUs from Google and GROK got acquired. Also not acquired, but a CEO. So from the infrastructure aspect, how do you see that shake up to be able to build the protocols and the layers so that the ecosystem can still function as decentralized? Yeah, yeah, I mean you're absolutely right. I think you need some centralization, some decentralization. And the pendulum kind of goes back and forth, right? Like we had mainframes, then we had PCs on the Internet, then we have cloud, so it got recentralized and then we are again moving to edge devices. So I think the pendulum will go back and forth, which is perfectly fine. And the whole stack, from chips to cloud to foundation models to platforms to application, the whole stack. So I like to say that as you know, the Internet has the seven layers of OSI stack and the agentic Internet will also need its own seven layers of OSL stack. So that all has to emerge. And it will be kind of a joke if you said there are only these top end chips that you should be using because there'll be chips of every kind all the way from Raspberry PI to the latest Blackwell. So you need the whole. And so when you see in the market where you have Nvidia partnering with intel and AMD and arm, I think we're going to see in Broadcom and so on, it's fantastic to see how this ecosystem is emerging and as you know, there is cpu, gpu, but there's also npu, right? Neural processing units and just the latest ces NPU has got a lot of coverage because a lot of PC makers are saying, hey, why rent AI when you can run a lot of the things on your PC? So I think the heterogeneity of that is going to be very, very exciting. But you know, I would say in a very short amount of time we'll have, you know, we won't kind of use AI, we won't use AI agents. We will live with AI agents. You know, it's like we don't say do you use the city? No, you live in the city, you know, or do you use your family? You live with your family. And we are going to live with our AI agents and they'll be part of our life. They'll be almost transparent. We won't even notice them. So from the, my mind goes to the, to the workforce and the workforce management because if you at scale, if we have like a CPA agent, accountant agent and then like many different agents working and collaborating in a system, right? Do we. And then especially in, mostly in American society there's a big, the identity of is associated to work. And if that work is taken away, what do you think will people also there's of course training and being part of the ecosystem and so forth. But what will be population evolve into what will be us will be then be the orchestrator delegating work to get things done for us. But then that will create a massive amount of economic output and productivity gain, right? What do you see from that lens? I mean, let's talk about two possibilities, right? The you know, the quiet dystopia, let's call it the red curve, which actually we are on unfortunately is where, you know, we have this intense protocol and partnership wars that companies start carving out their fiefdoms, then they use their user base and use that with network effect and create unethical walled gardens. And that's already emerging. And soon after that, maybe six months after that, they start creating dependencies in their network. So if you use the analogy of the World Wide Web versus iOS and Android, the World Wide web remains decentralized even today. Anybody can create a website and launch a business and get going. It's permissionless. But we didn't get that in the mobile Internet. We had iOS and Android and we should thank Apple and Google for creating a very secure and safe platform. But the flip side of that is it prevented Global south from participating in a massive innovation ecosystem. Right. And every decision has to be made by this small group of people in these two companies of what features will happen and how the innovation move forward and so on. And if the similar thing happens in the agentic web and there are only three or four or five companies that control not the App Store and Play Store, but the so called Agent Store, that any agent you launch has to be launched in one of these five platforms, that's going to be very dystopian. And as you said, what could happen is that we'll have this winner takes all, which is there'll be exactly one accounting agent for the whole world, exactly one legal agent for the whole world, exactly 1 math teacher agent for the whole world and everybody else space for that and then space for that, and a math teacher accountant just uses that and they have a soulless job where they're just using the tool and providing services to everybody else. So I would argue that when people talk about AGI and singularity, actually I'm not worried about some kind of, this massive, kind of cute, you know, a massive event with bad stuff will happen. I'm actually worried about this very quiet dystopia where most of the world has not a meaningful job. Doesn't mean there'll be job losses. They still have jobs, but there are so less jobs because they're just slaves and intelligent slaves to these highly trained models. So that's the red curve. And unfortunately we are on that red curve right now because any one of this platform that has $100,200,000,000 or sorry, 100,000,000 or 200,000,000 users or more overnight, they can convert and create an agent zero for each of those users and control their life and slowly they'll suck away what each of these agents can do. It's like, actually, you are pretty dumb. Your agent's pretty dumb. Why don't you just use this accounting agent we have centrally delivered for you and so on. So that could be the red curve. A lot of despair and dystopia, but. And that would be game over by the way, Alp. If that happens, we stay on the red curve. I like to kind of invoke Adam Smith and the Invisible Hand. This is like the end of middle class. Effectively you have this highly polarized haves and have nots. There'll be few players that control the AI ecosystem and everybody else is just downstream and it's eating the middle class, eating the knowledge work that is so widely been. So there's no shared prosperity. Yes, but we could be on this so called a green beautiful curve where there are billions of micro AIs. Each of us has our own agent and we create this amazing agent, the Ramesh agent, the ALP agent. I'm very proud of the Ramesh agent. I pour my blood and sweat into it. I make it better and better. I take it to the best school, I take it to the best repair shop. I make it available to others. Sometimes I charge, sometimes I do it for my nonprofits. No, my agent is doing amazing things and my agent mills the ALP agent and they do some amazing things. And it could be just a glorious world where everybody's busy improving their own agent. At least in the digital world we have this very interesting and vibrant bazaar. And to do that we have to make sure all the protocols are open. There are a lot of nonprofits, like a lot of dot orgs. We need a the next World Wide Web consortium, the next Mozilla, the next icann, the next ITF to emerge to make sure it all remains open. And then we'll have a very long tail of applications that we can't even imagine today. People are creating new forms of music. They're kind of new kind of experiences. They're taking their whole library of sitting something in their basement and put it out there. Pharma companies are taking their failed drugs that are not making any money and create agents for them that can benefit somebody else. You can see this amazing new world, very glorious world. And we could be on that green curve. We are not on that green curve right now. We are on the red curve where people want to monetize agents right away and they want to control, you know, who's running the agent stores. Do you think that, like, because we have the closed models and then the open source models and I think where closed models are not accessible, open source models are dominating right now. But there's also the commoditization of the application layer right now. Like anyone can redesign and reimagine an application and release it into the web. Do you think that itself is maybe where the hope is at? Right? Like where, where anyone can come and say, I'm going to create the new Facebook. I may not have 1 billion users, but I'll try. Yeah. So I would say we typically think of like, oh, there's foundations, which is like models and so on. There are platforms and there are applications and a lot of people say, hey, applications, where it's at, that's where most of the money will be made. We'll see. But I think there's a fourth layer which is very interesting, which is, I would call teaming, which is very clever ways to bring different agents together. Right. And I think the fourth layer is going to emerge and that's where actually most of the economic value will accrue. And if you think about today's capitalistic world, one could argue that human intelligence, like one human intelligence to next human intelligence is not that different. Maybe they're like, maybe one sigma away from each other. But what ends up happening is the team that comes together and executes together is the most powerful. So one startup versus another startup, one company versus another company, one country versus another country. The only difference between them is how good is the team, how did they form the team and how did they execute? And so I think it's the same thing in the agentic world. We will see this agentic teams that emerge and they will actually have more value than everything else below it. That's highly commoditized. As you said yourself, intelligence is commoditized. The models are completely commoditized. The platforms will get commoditized, even the applications will get commoditized. So the creating the teams is the most exciting part. And I think the orchestration layer for that is some of the biggest opportunities out there. And if you think about the business world, the reason we have MBA schools is not to make people smarter in a particular domain, but it teaches them how to create, maintain and foster better teams. And I think we'll see the same thing, you know, the MBA equivalent of agentic world. And they do build like relationships with the other classmates, interact with each other. Right. Whereas the accounting book could be same everywhere. Exactly, exactly. I would love to shift the topic to robotics because they're saying it's five year ahead, 10 year ahead, or maybe today, but with your work on computer graphics and the AI understanding the actual world around it, and then that also includes the role models as well, like interacting with the video game 3D Space and Reacting to it. How far is it at? And do you see yourself having a robot at home? And if so, what would the user use it for? Right. Is it because there's the classical perspective, like robot doing the dishes or like robot cleaning the apartment and house, household chores. Right. But it could do many more things. Yeah, yeah. How do you see that field emerge and what is the limitation right now from firmware side and the hardware side? Right. So I would say like if I kind of think about the human senses and human intelligence, we started with brain, then brain plus eyes, then brain plus eyes plus legs. But the brain, eyes plus legs plus hands is actually very, very challenging right now. So we have thinking systems, the brain, then we have computer vision systems with eyes, then we have navigation systems with legs. But the manipulation systems are actually still very challenging because just the complexity of that, the number of degrees of freedom we have here, the data sets that we have to capture. I've seen these pictures and videos from China where you have hundreds of people sitting and coaching and training robots, and that's how they're capturing physical data. So I think we have a very long way to go in thinking about full fledged robots that can also do manipulation. We had two companies and we just had exits, but we solved the navigation problem in robotics. Very proud of that work. But still we were solving the navigation problem more like the autonomous. Autonomous vehicles. Autonomous vehicles, yeah. Right. And. Or autonomous gripping, you know, things like that in factor robotics and so on. So the two companies we sold were one in factory robotics, one in Autonomy Robotics. But I think that's much further from manipulation and that's why we haven't seen that many companies that can solve manipulation that well. So I would say we need a few more breakthroughs before we can see that. And I see kind of these three phases of agentic world. One is kind of just the digital agents that take care of your email and slack and travel and finances, knowledge work and so on. So I think that's going to happen very, very quickly. The next one is kind of the physical agents that can be in the physical world, not just in terms of manipulation, but they can build products that can have chemical reactions. Anything happening in the physical world, not just manipulation and navigation, then the most Complex one is I would call social agents and population agents that actually try to understand this species which is so complex, which is human beings. I think when it's dealing with digital data or physical data is still inanimate information. But when it comes to agents working with humans, it's societal problems. So think about health, think about education. These sectors are so called human economy sectors and they're the most difficult for agents to solve actually. So I would say we have several years to go between each of these phases. And I think there's the aspect of like we do want the robots to be human humanoids, right. To be close to the humans. And I always wonder why is that? Like why do we need to have an exact robot that maybe mimics similar or like think similar or maybe things more deeply? Yeah. Does it have to be like that? Because in a way we are also reinventing the future right now, like reinovating the future and what we imagine becomes reality. The argument folks make is that if you take even self driving cars in the 1980s, the way DARPA and DOT was funding the research, they were saying let's just build new highways with RFID tags along the way so that these new cars can just navigate themselves on this instrumented highways. But what ended up happening in the end is we started creating this self driving cars that are completely self contained and they don't need any help from outside. And of course the argument people make is that hey, the world is made for humans, so we must make the machines that live in the existing human world. We cannot go and retrofit the human world and the same thing with robotics, like let's create humanoid robotics because I can't instrument the world because humans already live in it. So I cannot just have real robots because they cannot take the stairs and blah blah, blah. Now I think what you're saying alp here is that's kind of a cop out. And if you just reinstrumented the world, actually we can do much better. I think it's a debate, I mean, what comes first? I think I would imagine a country like China could say let's just redesign the whole damn thing so that the problem of creating efficient robots is easier as opposed to saying let's put trillions of dollars in creating new robots so they will exist in the messy human world. So I think it's a debate of do you invest in instrumentation of the world or do you invest in self contained technologies? And then that's very interesting because we do reimagine the software world, right? Like we do. Imagine the infrastructure layer, protocol layer, all changing and evolving. But when it comes to the robotics, the robot still needs to walk in a road or pass through at all. It has to look at street signs that are made for humans. Yes. Imagine if every street sign, every stop sign, every traffic light had its own RF beacon. It'll be so much easier for self driving cars to navigate. But Nobody puts this $2 RF beacon like an airtag effectively to do that. So that's the world we live in. Yes. And if you, I would like to ask if you were to go back home and your hometown and now, right. Looking back and connecting all the dots that, that, that you lived and what would you say to a young person who's just as at your age, at that time to embark on a future, what would that person should do or chase? I come from a very humble beginnings and when I'm at MIT and I'm working on amazing projects, I pinch myself like, wow, I'm this kid from this small town, what am I doing here? And I have access to some of the best talent, the best resources. And we're doing this ambitious project and people actually pick up the phone when I call them, you know, some of the top leaders. And I think, you know, I was at Google X and I was chatting with Larry Page and about some projects and he said something very important. He says, you know, Ramesh, that's an interesting project, but that doesn't sound very ambitious. And I said, I need a little bit more time. And he said, ramesh, whether you work on a simple project or ambitious project, the amount of work you have to put in is the same. You're still working 30, 40, 60 hours a week. So it doesn't matter whether you work on small things or big things. So you might as well work on the most ambitious things. I wish somebody had told me that when I was 21, 22. The second thing I would tell myself is that as a scientist, we love to think about really elegant solutions to problems that are articulated by somebody else. But what I realized over time is actually the most exciting part is to define new problems. Inventing problems is more fun than inventing solutions. One of the reasons, as you know, after PhD, I joined a research lab here and the research lab, what I was really good at is kind of inventing new solutions. But the reason why I moved to MIT is of course it's a bigger playground. But I wasn't inventing solutions, I was inventing whole new fields. And I wish I had realized that much earlier. In my life that it's not about solutions, it's actually about going after the right kind of problems. Of course, if you have a match of a great problem and a great solution, a great algorithm, that's even better. But you should be spending even more time figuring out which problem or creating or inventing a new problem statement. I think that's most exciting. And then you don't have to do this earlier in your life, but later in your life you have to start realizing that you shouldn't just go after point solutions, but should go after systemic solutions. And so a lot of our work in Agentic Web and Project Nanda and so on are all these systemic solutions which sound very daunting because if you know, tug of war, it's like multi sided tug of war. Right. It's extremely complex, extremely ambitious. But that's where the fun is. Yeah, that's where the fun is. And those are the problems that most people shy away from. And then I think in the tip of the iceberg, that problem statement then drills down and then you see even bigger problems that needs to be solved. Yeah, that's what get me excited. And you know, as you know, I grew up in this small town from Mumbai. We have a very famous festival called Kumbh Mela and it's like 30, 40 million people show up. And the beauty of this pilgrimage is very colorful, it's very spiritual and people from all over the world come. But when you're in this festival, I feel like the whole population of 30 million is in unison. It's as if it's like this one single orchestrated entity that's moving very elegantly with a lot of grace around the whole town. That really taught me something, that actually the beauty comes from these decentralized entities that figure out what each entity's role is. And they don't try to find a greedy optimum, they find a global optimum for it, what works for the whole system as opposed to what works for one individual. There's just something beautiful about that as a computer scientist that we should always be thinking about this invisible orchestration layer and create solutions in computing world that are at that orchestration layer. And so a lot of our work in Agentic web is kind of inspired by when I was growing up looking at Kumbha Melo. That is very impressive because there's the humans homogeneously contributing to one another and then they're not selfishly trying to be taking the whole floor, but it's togetherness and then solving the problems for the others. Versus problems for ourselves. Yeah, absolutely. And in the beginning of the conversation, you asked me about centralization versus decentralization. And when you look at a city, if you were to design from scratch, you're like, actually the best way to run a city is to have a dictator who designs a city and tells people what to do. That's the most efficient way to run the city. But that's not what we do. You know, cities are more interesting because of heterogeneity. They're not most efficient, you know, but they're more exciting. Yes. And I think, I think we have to remember that, that, you know, centralization gives you efficiency, but decentralization is what gives the robustness over time. Yeah. And the creativity too. Because I think in a, like, in a city, if everything is very predictable and monologues and the monotonous, then it becomes a bit boring. Absolutely. Yeah. And hopefully the agentic web will be this vibrant, thriving place as opposed to it's siloed and a handful of agent stores. And an agent from one agent store cannot talk to another Asian store. Or if an agent from one agent store has to do something that's more creative and slightly outside the norm, it has to take the permission from the people who run the agent store and so on. By the way, if the agent makes any money, they have to give away 30% to the agent store. And at any point the algorithms could change and the visibility of that agent could change as well. So hopefully we will not be in that world and the agentic wealth will really unleash trillions of dollars of economic value for everybody, especially the global South. Yeah. I do believe that at the end of the day, we do, we do create what we imagine in the world. And so it's at the end how have the people will collaborate together and then do the solve the hard problems like yourself looking in deep and then finding those and not be discouraged by how difficult they are. And I'm pushing hard so that the future is for everyone. Absolutely. So we talked about the future with AI agents and how it can go for a good future that benefits everyone or a future that's limited in the hands of the few. What scares you in this future that we are handling right now? And what are you going to do about it to solve it? I mean, as I said, the foundry era versus the garage era versus the bazaar era of AI, the foundry area is safe. We have these centralized models. There's some handful of people keeping it safe and so on. But as we said, it has challenges because you could get an innovation lock in. You could get walled gardens. It might make Global south, could become a purely consumption economy. And we may not get the benefits of the agentic web. But there's also challenge on the decentralization when we have these workshops and bazaar. If it remains just crony capitalism as opposed to ethical market forces, you could also imagine the equivalent of the dark web. You have all these Asians that are untrusted, unreliable, unregulated, not anchored on human identity, and they're just running amok. There's a Goldie zones. You don't want to three or four or five agent stores, but you also don't want millions of random stores that are not regulated and that are not. So there's like a Goldilocks zone, maybe somewhere between 100 to 1,000 agent stores that kind of keep a leash on it. But because there's so many, they have to kind of talk to each other and they remain interoperable and they remain safe and so on. And so when I realized that, I said, you cannot be a centralization maximalist, but you also cannot be a decentralization maximalist. That's true in every ecosystem, because even Internet today is run by things like icann, which is a centralized system. And they give you the name and address, like your name for your website and the IP address for you. And there are nodes, there are super nodes, there are certifying authorities. So there's some centralization and some decentralization. So for last 10 years in our group at MIT, we have been working on how do we build the networked AI. But over three, four years ago, we realized actually we need to start thinking about the bizarre aspect of it, and we'll start thinking about what these institutions could be that will bring stability. And we call it Project nanda. NANDA stands for Networked AI Agents in Decentralized Architecture. Of course, it's a mouthful. And Nanda is the name of my sister, which means joy in Sanskrit. And so Project NANDA is trying to create four things in the first phase, which is how do we create the equivalent of ICANN and DNS, like the domain name service, so that we have Asian names that make sense, they are not duplicated. Like you cannot go tomorrow and pick up I'm Asian Microsoft, or I am Asian Starbucks. I cannot do that. Only those companies can do it, and so on. Or I cannot become Agent alp, for example. So you need an ICANN for that. Then you need certifying authorities, you have to credential them, but you need Certifiers of certifiers, then you need interoperability. And the fourth thing very critical is like attestation. So when you launch a car you need crash testing, you need test tracks. Once the car is launched every year, you have to get your inspection sticker. And once the car is in the market for a few years, they will do a safety recall. Because even if most cars, billions of cars are doing well, maybe a couple of cars have some trouble with their, I don't know, with their seat belt and they have to recall all the cars to fix their seat belts. So pre launch live monitoring and after launch is what we do in terms of attestation and the same thing we have to do for AI agents. So what we're doing in Project Nanda is making sure we have a way to create safe, secure ecosystem of agents. That's the first goal for us. The second goal for us is to not just think about the syntax of agent commerce, but also think about what can be built on top of that, what's the stock market of AI agents, what's the groupon of AI agents and so on, like collective bargaining or prediction markets and so on. So we are building that, what's the knowledge pricing and things like that. And the third one is this immersion behavior with agentic societies. So for us, Project Nanda is like the combination of ICANN and Mozilla and World Wide Web Consortium all built together in the next six months. Because if you don't get it right, my real concern is that there'll be so many scammy agents out there that we will all gravitate towards some safe environment. And this is what happened with iOS and Android, the Nokia N95 and others. You could load apps, side load apps, but they were so scammy that at some point we said, you know what, I would rather just download apps from an app store or play store. And so we thanked Apple and Google for creating this walled gardens for us because we got a very safe experience. But I think in the agentic web, I think it's a false choice. Secure AI agents, Secure and safe AI agents at the same time. A vibrant open ecosystem. I think we can achieve both. And that's what Project Nanda is all about. That's incredible. And I'm excited for that future. I think an ecosystem like that, that anyone can contribute is the future that we deserve. Absolutely. So thank you very much for joining me today. It was a pleasure to have you. Thank. You.