Ramesh Raskar - The Agentic Web and the Bazaar Era of AI

0:02 Fantastic. So for some time we have been in the factory era of AI where some large companies would produce these products in you know large quantities and we would just consume them. but very soon thanks to edge AI and small models and things running on the edge things running on your own machines we're shifting to the bazar era. So a lot of our work at MIT is about thinking about when we'll have billions maybe trillions of AI agents on the open web what would the internet of AI agents would look like and to motivate that imagine a scenario where Sarah in Texas in rural Texas has a health condition she goes to her doctor John and says hey I have chest pain I have coorbidity and she's just not getting a good answer from from her doctors and Dr. John can go on the open hive on the open web and say hey listen for 100 bucks can somebody solve the problem for Sarah and when that happens billions of agents all over the world wake up and they say hey maybe I can take a piece of this 100 bucks and you know they can look at the data if it's related to corresponding health condition start training models also create a coalition start collaborating with each other and eventually maybe a handful of them can solve the for Dr. John and those let's say about 10 of them sol those 10 get paid out from that $100 and everybody else kind of disappears and then eventually Dr. John gets you know an Expedia like answer of how he could help you know Sarah the patient and so what's happening here is this not a it's not like agents for e-commerce right this is commerce that's possible because of AI agents and Dr. John is not accessing you know some kind of an amazing health model AI model out there but he's tapping into a network of intelligence and that network is available as a marketplace right this is what we mean by the internet of AI agents and the commerce for agents so again we are kind of in the mainframe era of AI kind of a factory era of AI but very quickly we're shifting to the PC era some companies already you know using agents within trusted boundaries, the internet of agents. But very soon we'll shift to the internet of agents where agents are, you know, collaborating and and negotiating with other agents that are well outside the trusted boundaries. but that's still just for geeks how we can connect.

2:42 That's like the internet. To go from the internet to the worldwide web, we needed, you know, HTML and browsers and URLs and all the good stuff. And project Nanda u that originally started at MIT stands for AI agents in decentralized architecture. Right? A mouthful. Nanda for short means Sanskrit means joy in Sanskrit but also name of my sister. All right. So what does it mean to have billions of agents out there? And if you take this example I showed you of commerce that agents will conduct. I think the closest analogy I can imagine is highspeed trading. High frequency trading. and the reason why high frequency trading works is we have a bunch of things right. We have you know there is a registry you know there is you know every merchant every stock you know is is is pre-qualified you know there is somewhat price transparency in how that works. you know there is a way to deal with fraud detection and so on and all these things also had to be built for the internet of AI agents right so this is a massive infrastructure that's going to come together come you know come get built for the billions of Asians that are out there so any anybody here like a openclaw fan here right but you know openclaw is great but it's only inbound it's only behaves as a tool I cannot call your open claw agent it's like having a telephone that can only make calls outside but can never receive calls right so it's a one-way agent not a birectional agent so the question is who's going to build this infrastructure but a more problematic way to think about this is whoever builds this infrastructure will also own it so the question for us is will we end up in a you know wall gardens of this internet of AI agents or will we get an an open and vibrant and neutral and safe and scalable web for AI Asians. So that's the mission for us for for Nanda. and we think that there are four main bottlenecks.

4:49 and if you can solve those four bottlenecks, there's a chance we can keep the agentic web as open as today's internet as opposed to how we see on the mobile phones like iOS and Android or Facebook and Twitter all the wall gardens. so the four bottlenecks are you know a a I can DNS- like structure second is passports and and certification third is obvious which is interoperability and fourth is at a stationation so I'll go through this very quickly and our group at MIT has been working in this space for you know last 11 or 12 years so we this journey of thinking about what happens after alexnet when we have billions of individual you know AI systems that are interacting with each other and now going forward within There are three main research problems we need to solve.

5:35 originally starting at MIT but now we have many collaborators. The first one as I said is like the DNS for Asians. The second is knowledge pricing. We think it's very critical that you know data models inference compute all needs to be priced you know just the way when you hire a candidate we can figure out their salary. and the same way you know we have to do knowledge pricing. It's a very beautiful mathematical problem. Come talk to us. And the other problem is co-learning. You know the last 14 years have been a fantastic progress in machine learning but going forward we think it's all about machine co-learning about how two agents learn from each other. and when you think about co-learning, you know, it's not just about, you know, agents buying services and transacting, but they're going to form coalitions. They're going to negotiate, they're going to compete, you know, for resources. and they have to achieve decisions and there'll be some alliances, there'll be some betrayal, you know, it'll be like the game of thrones among agents. so the other thing we need to do is we need to have a starting point for the for the web. There was something called Yahoo.

6:42 those of you old old those of us as old as me and Yahoo kind of created a safe space to explore what the web is and we need something similar to do that so we created something called Nanda Town so go check out Nanda Town off of MIT website it's kind of a sandbox that allows anybody to bring enterprise agents payments memory you know if you're working in any of the spaces you can come to another town and explore this extremely heterogeneous system and you can have some ethical hackers there and so on as well it's it's it's moving very very quickly. So I enjoy I encourage you to explore Nanda.

7:18 and then the research of course works in multiple directions you know Asian techch devices you know this knowledge pricing co-learning you know new forms of Asian marketplaces and so on and of course this is happening under a very large tent that involves large companies and startups and governments and so on. and more recently we have this series of hackathons called NandaHack that are sponsored by different organizations. So just go to nanda hack.mmedia.mmit.edu to explore it. Now when we talk about billions of agents we should be thinking about at nation scales. so what's happening now is thanks to rather work at nanda we're able to work with very large ecosystem. So for example in India we launched something called Duth which means messenger and 1.5 billion agents that are not owned by there's no corporate capture and that run on the identity and payment rails they already have city of Boston is also launching with us you know a Boston agent initiative and I I invite you to come and talk to us about how the civic agents are evolving as well and what you'll quickly realize is that today we are thinking about agents for X you agents for coding, agents for browsing, agents for, you know, debugging, agents for shopping and so on. But we need to flip that and start thinking about, you know, services for agents, X for agents, infrastructure for agents, compute for agents, insurance for agent, you know, healthcare for agents, education for agents, everything that we do in the society needs to be done for agents, right?

8:51 So I think the open agentic commerce is going to emerge because intelligence will be discovered intelligence will be priced and intelligence will be coordinated at scale. So come join us at nanda.edu. Thank you.