Latest / EDGE AI POD / Networked AI Agents Decentralized Architecture

Transcript

Networked AI Agents Decentralized Architecture · EDGE AI POD

• 0:05
  Speaker 1: It's my pleasure to introduce our next speaker, pradyumna Shari, who will be sharing his work on NANDA, an open protocol designed to enable secure, collaborative and interoperable networks of AI agents.
• 0:18
  Speaker 1: Pradyumna, the floor is yours.
• 0:21
  Speaker 2: Thank you so much, hajar.
• 0:22
  Speaker 2: I hope I'm audible.
• 0:24
  Speaker 2: I'm also going to be sharing my screen and hopefully my screen is visible, awesome.
• 0:36
  Speaker 1: Not yet, Not yet Okay.
• 0:37
  Speaker 1: Yes, yes, perfect Okay perfect.
• 0:40
  Speaker 2: Yeah, my computer has a bad habit of stopping screen sharing in the middle, so if at any point something looks weird, please feel free to poke me so that I can correct it.
• 0:50
  Speaker 2: But yeah, thank you so much for having me here and giving me this opportunity to talk to you about something that we find very exciting, given the time that we all live in, given the space of agentic AI and where it stands today.
• 1:09
  Speaker 2: First of all, I'd like to thank the organizers for putting together such a great lineup of speakers, excluding myself, of course, but as someone who is not traditionally within the edge AI space, I've learned so much from all the presentations that have come before me and I'm sure the ones that will come after me as well.
• 1:32
  Speaker 2: So today I'm here to talk to you about some of our work at the MIT Media Lab, which we call NANDA Networked AI Agents in a Decentralized Architecture, which we call NANDA Networked AI Agents in a Decentralized Architecture.
• 1:47
  Speaker 2: There's a lot of fancy words in this acronym and I hope, by the end of this presentation, that I'm able to explain to you our goals, aspirations and what we intend to do with this project NANDA.
• 2:00
  Speaker 2: So a bit of contextualization.
• 2:04
  Speaker 2: Over the past couple of years we've seen a massive growth in the space of AI, especially with these large language models such as ChatGPT, cloud Gemini from Google, grok from XAI and so many others that have really changed what one can do with machine learning, deep learning, ai models and just knowledge assistance.
• 2:29
  Speaker 2: These models have global knowledge across domains of expertise and have truly taken the world by storm and have really transcended the boundary between engineers and technologists to becoming learning-based products and tools that are usable and understandable by humanity at large.
• 2:56
  Speaker 2: More recently, and as Aruna was mentioning right before, we are now in the realm of what's called agentic AI or AI agents, which are large language models or models based out of large language models that are not just reasoning capable but also action capable.
• 3:17
  Speaker 2: In other words, these are large language models that have some kind of agency or the capability to take actions based on their reasoning.
• 3:25
  Speaker 2: Sort of this growth or this change from change that we are seeing or improvement that we are seeing in AI and large language models is actually very analogous to the way sort of mainframes and PCs developed into what we view as the web as it exists today.
• 3:48
  Speaker 2: So AI models from a couple of years ago, like ChartGPT, are analogous to the mainframe PCs that we had in the 70s and 80s, which were individual sources of intelligence, let's call them, but were not able to communicate outside themselves with other intelligent entities to do tasks beyond their own capabilities.
• 4:12
  Speaker 2: From there, the mainframe PC era moved to what we call the intranet, which has computers and PCs that are able to talk to other computers and PCs that they know exist out there.
• 4:25
  Speaker 2: This is similar to how agentic AI is today.
• 4:29
  Speaker 2: Ai models, large language models and agents are able to talk to other agents that they know that exist already, and interact with them within their own ecosystem, talk with them, communicate with them and do agentic tasks.
• 4:46
  Speaker 2: For example, let's say, a cloud provider has 100 agents, each of them capable of doing certain tasks, such as scraping the web, researching topics, looking up academic papers, and then these agents can together, within this known topology, talk to each other and do some complex agentic task.
• 5:07
  Speaker 2: Today we are at a phase which we call decentralized agentic AI, which means again, as Aruna was saying, with the advent of all of these amazing tools that enable all of us to deploy AI agents, such as Crew AI or LanChain, that make agent deployment very, very straightforward and a low code to no code solution.
• 5:31
  Speaker 2: We are at a stage where almost all of us are capable of deploying our own AI agents, connect them to specialized databases and tools and build uniquely capable AI agents.
• 5:44
  Speaker 2: Each one of us is able to do this today and each one of us can do this anywhere in the world.
• 5:50
  Speaker 2: So we exist in a world of decentralized agentic AI, where agents exist in a trustless economy.
• 5:59
  Speaker 2: I don't know all the other agents that exist out there, and that limits my capability to leverage all of this agentic capability that exist out there and that limits my capability to leverage all of this agentic capability that exists out there on the internet of agents.
• 6:10
  Speaker 2: So this is where Nanda comes in.
• 6:12
  Speaker 2: We want to take all of this amazing decentralized agentic capability that exists already and that will exist, especially on the edge, and build protocol and infrastructure-layered tools to enable all of these AI agents to discover each other, find each other, talk with each other and interact with each other to do tasks, essentially.
• 6:39
  Speaker 2: So imagine the World Wide Web as it exists today, which are millions, hundreds of millions of web pages and domains that I don't already know exist out there, but I can easily access via the web page names find web pages do tasks and serve the web seamlessly.
• 7:01
  Speaker 2: Now imagine if each of these web pages was agentic, each of these web pages was intelligentic.
• 7:06
  Speaker 2: Each of these web pages was intelligent, could do agentic tasks, could be autonomous and could collaborate with other web pages or other agents to do things that are beyond our wildest imagination.
• 7:16
  Speaker 2: This is what we want to envision with Nanda architecture for AI agents in a decentralized ecosystem.
• 7:26
  Speaker 2: We often like to motivate this with a fun example of a birthday party.
• 7:32
  Speaker 2: So today, when I have to organize a birthday party for someone, for my family member, for my wife, I will have to do 10 different things.
• 7:42
  Speaker 2: I will have to send invites to all of my friends, ask them when are they available?
• 7:47
  Speaker 2: What are their calendars looking like?
• 7:50
  Speaker 2: Look at their calendars, figure out what their availability is.
• 7:52
  Speaker 2: Then separately, reach out to the grocery store, ask my gross like, figure out whether grocery store A has all the inventory that I need to host the party.
• 8:03
  Speaker 2: Then reach out to a bakery separately and figure out whether bakery A has the cake that I want or if I have to go to bakery B and all of these multi-dimensional optimization problems.
• 8:14
  Speaker 2: Now imagine if all of these entities had intelligent agents and I, too, had an intelligent agent of my own, I could just tell my intelligent agent organize a birthday party for X, invite all my close friends and buy all the inventory that I need to host this party correctly.
• 8:35
  Speaker 2: The AI agent would be able to talk to other AI agents of people, of shops talk, look at other people's calendars via their AI agents and do all of this optimization on its own.
• 8:47
  Speaker 2: This seems like almost a dystopian reality, almost a far-fetched reality, but this is something that AI agents are actually capable of doing.
• 8:56
  Speaker 2: Of course, this doesn't exist today, because there are a lot of very critical problems that we need to solve before this becomes enabled.
• 9:03
  Speaker 2: These include problems like privacy, data markets, verifiability, authentication, how to UI, ux development to ensure people can easily interact with such tools and, of course, making onboarding to such an ecosystem easy, democratic and widespread.
• 9:25
  Speaker 2: This is the goal of Project Nanda, or the Internet of AI Agents, which is how can trillions of AI agents collaborate across organizational silos, communicate seamlessly, navigate autonomously, socialize, learn, earn, transact on our behalf?
• 9:40
  Speaker 2: Nanda comes with a bunch of its own algorithmic economic research challenges.
• 9:48
  Speaker 2: These include new discovery mechanisms.
• 9:51
  Speaker 2: How can I discover AI agents that exist out there that I don't know of, that I want to potentially interact with, how do I verify and transact with these AI agents, how do I incentivize the existence of this open web of AI agents and how do I build what we call knowledge markets or intelligence markets, which is, how do I value the intelligence or knowledge that other AI agents are providing as we interact in this web of AI agents?
• 10:23
  Speaker 2: At Nanda, we build off of some very incredible work that people in industry are already doing, and we view this Nanda project as a quilt almost of all of the critical components that we think are necessary to enable this web of AI agents.
• 10:40
  Speaker 2: This includes, obviously, critical communication protocol aspects, such as, as Arunabh was talking about, earlier protocols such as MCP or Model Context Protocol from Anthropic and A2A or Agent to Agent from Google, but also so many other things that are required to enable this web of AI agents.
• 10:59
  Speaker 2: This includes how to ensure trust between unknown AI agents, how to enable safe transactions for agentic AI, how to enable or how to create foundational layer attributes such as protocols for communication, aspects such as registry, to enable discoverability and search, and how to enable the building of applications on top of all of these foundation layer aspects that this web of AI agents requires.
• 11:27
  Speaker 2: As we come from an academic institution, we are also very excited about all the research problems that exist in the space.
• 11:38
  Speaker 2: This includes how to build better AI agents, how to ensure that these AI agents are interfacing with data in a private manner, how to ensure verifiability, orchestration, valuation and user interfacing.
• 11:54
  Speaker 2: As I mentioned already, we are inspired by a lot of exciting work that's already happening in industry in the space of decentralized agentic AI and interfacing AI agents.
• 12:06
  Speaker 2: Together, these include model context protocol by Anthropic, which is a way to connect tools to AI agents, and A2A by Google, which is a protocol that provides specifications for how agents can talk to each other, how agents can evaluate each other's capabilities and build teams to perform tasks.
• 12:27
  Speaker 2: All of these exist in the communication layer, and we build on top of these to bring together other critical aspects that are necessary to build this web of agents.
• 12:37
  Speaker 2: These aspects include, of course, how can agents discover each other without actually knowing that they exist, how to enable something like a search for ai agents?
• 12:46
  Speaker 2: This is actually a very, very non-trivial task.
• 12:48
  Speaker 2: Think of google search that exists on the web, but something that is enabled, uh, by hyperlinks and html text that exists on web pages and think of how ai agents are actually very implicitly defined.
• 13:02
  Speaker 2: So there's no hyperlinking between AI agents, there's no text to search against.
• 13:08
  Speaker 2: So how do you build search functionalities for your agents?
• 13:11
  Speaker 2: But also aspects such as authentication?
• 13:13
  Speaker 2: How do you authenticate AI agents, especially agents that are likely to behave autonomously on their own, and how do you verify or trust them and finally trace?
• 13:24
  Speaker 2: How do you instill accountability in autonomous AI agents?
• 13:27
  Speaker 2: How do you instill accountability in autonomous AI agents?
• 13:28
  Speaker 2: How do you pair them with physical world entities to enable accountability, reputation, authentication all that tasks.
• 13:37
  Speaker 2: At the Media Lab, we've been working on problems along these directions, leading up to where we exist today in the space of AI agents and agentic AI and decentralized agentic AI.
• 13:52
  Speaker 2: So we are very, very excited as we embark on this open source project which is Nanda, which is a broadly collaborative project, and I'll now talk about some slightly more technical stuff that we are building as part of Nanda to enable this web of AI agents.
• 14:09
  Speaker 2: So the motivation for this and sort of our guiding North Star as we build this web or internet of AI agents, this democratic internet of AI agents that anyone can sign on to, anyone can add their AI agents to and communicate via is.
• 14:26
  Speaker 2: We sort of take inspiration from how the World Wide Web developed Also, incidentally, sort of strongly pioneered here at MIT via the W3C Consortium and by Tim Berners-Lee, I'm sure you all know and we sort of take inspiration from how the World Wide Web developed, but also being cognizant of how agents are very different from how web pages work.
• 14:52
  Speaker 2: Agents, for example, are dynamic, actively respond and adapt, unlike static web pages.
• 14:58
  Speaker 2: Agents can operate where needed.
• 15:00
  Speaker 2: This is very relevant to the forum we are at today, which is agents can be at the edge.
• 15:05
  Speaker 2: Agents can be small models deployed on edge devices such as robots, cell phones, household devices and so many other things that maybe we don't even envision right now, and not just from centralized servers.
• 15:18
  Speaker 2: Agents can have location mobility based on demand, based on operating at the edge, while web pages remain static.
• 15:27
  Speaker 2: Agents are consumer-driven, heterogeneous and involve bidirectional communication, which means, as opposed to web pages, which are a very clear client-server model, agents are actually a very clear peer-to-peer model.
• 15:42
  Speaker 2: Agent A can reach out to Agent B for their services, whereas at a separate time Agent B can reach out to agent B for their services, whereas at a separate time agent B can reach out to agent A for their services, and it's a very bi-directional and flat hierarchy.
• 15:54
  Speaker 2: Essentially, and finally, agents have a clear notion of teaming, where multiple agents can reach out to each other, delegate tasks based on their expertise and work with each other to complete certain tasks.
• 16:09
  Speaker 2: So, with that being said, I'll now talk to you about one of the critical components that we are building as part of the Nanda architecture, which is this idea of a Nanda registry for agents.
• 16:20
  Speaker 2: So, as I spoke about earlier, agents are likely to be very different from web pages and are likely to have very unique attributes, such as authorization, identifiability and discoverability of agents on an open web of agents is the idea of an agent registry.
• 16:54
  Speaker 2: We view this agent registry as a lean registry of entries, of agentic entries that map agent identifiers to a couple of things that I'll talk about soon, but broadly mapping agent identifiers to a couple of things that I'll talk about soon, but broadly mapping agent identifiers to agent endpoints and enabling all of these agents to be discoverable on an open web of agents.
• 17:18
  Speaker 2: Here's a simple architecture, very simplistic architecture that you can think of when we envision this open web of agents.
• 17:25
  Speaker 2: Think of the registry as an entity that enables identifiability and discoverability of agents at the center.
• 17:35
  Speaker 2: Now any two or more agents that want to talk to each other and find other agents that perform certain specific tasks can use the registry to find useful agents.
• 17:49
  Speaker 2: For example, adam agent on the left maybe needs to find a bakery agent in their vicinity, and they can use the registry and all of the attributes that the registry possesses that I'll talk to you about soon, to discover the bakery agent and find where the bakery agent is hosted by its endpoint.
• 18:09
  Speaker 2: And then, once the registry returns to the adam agent, the bakery agent, uh, the item agent, can then on its own, without the registry in the middle, talk to the bakery agent, do an agent to agent communication and talk via whatever protocol the two agree upon.
• 18:24
  Speaker 2: So, uh, uh, we want the nanda registry to be sort of a universal, democratic avenue for agents, resources, tools of all kinds to be visible and discoverable via the registry and sort of, once we do sort of the initial handshake via the registry, we want the agents to do their own thing, so talk to each other with the protocol that they prefer most, talk to each other in whatever terms, transact with each other in each other's terms, and so on.
• 18:57
  Speaker 2: And so now the registry is not just for agents, but also for resources, tools such as calendar resources, analytics tools, inventory tools, so on, so forth, that can be teamed together to build very, very powerful interactive agentic applications on top of this open web of agents.
• 19:16
  Speaker 2: So, as we try to build this registry, we again take inspiration from the current web architecture, which includes critical aspects such as domain names, metadata databases, such as whois databases, names, metadata databases, such as Whois databases, addressing schemes, such as IP addressing schemes, and certification, such as certifying authorities that enable verifiability in today's HTTPS-based web architecture.
• 19:41
  Speaker 2: And we see whether this can be applied to the agentic web.
• 19:45
  Speaker 2: And we hit certain roadblocks because of fundamental differences between the traditional webs and the agentic web, as I talked about earlier.
• 19:52
  Speaker 2: The traditional web is reactive, whereas the agentic web is proactive.
• 19:56
  Speaker 2: The traditional web includes manual navigation of web pages, whereas the agentic web is likely to involve intelligent AI agents which navigate on their own, and several other distinctions that make a traditional web-based system difficult to directly translate to an agentic based system.
• 20:16
  Speaker 2: We also take inspiration from how humanity developed from dial-up to broadband.
• 20:23
  Speaker 2: So, if you recall and I barely recall, because I remember that this happened several, maybe 20 years ago, which I barely remember where initially internet was served via telephone lines and at some point we decided that telephone lines are not going to be sufficient and we need to move to a specific broadband infrastructure.
• 20:46
  Speaker 2: Broadband infrastructure and this was based on some reasoning, but a lot of leap of faith.
• 20:52
  Speaker 2: But this has served us very well because this has enabled some critical aspects, such as it has solved a bunch of known challenges, but has enabled new applications such as video serving, content delivery and media serving that one would not have envisioned at all when the World wide web was developing.
• 21:11
  Speaker 2: So, as we move from the web of web pages to the web of agents, uh, we should not constrain ourselves to the traditional web, to inter interface agents, not just because of problems we know right now, but of all the things that agents will be able to do in the future that we just cannot envision as well today.
• 21:29
  Speaker 2: As we start, especially in the context of edge AI and given the scale that edge AI promises, with edge devices in the billions, potentially these AI agents are going to likely be in the scale of billions at some point in our near future and there is a need to think of the infrastructure in that context.
• 21:51
  Speaker 2: Of course, as I mentioned, there are several potential issues with the current web infrastructure, including aspects as a latency, security, traceability, verification and also scaling challenges, given that the web infrastructure, or the AI agents infrastructure, is going to scale to many, many agents piggybacking on the popularity of Edge AI, the Internet of Things, and also aspects such as personal AI agents, where we'll hit several fundamental challenges that traditional web faces, including addressing limitations, certification, propagation, metadata requirements and so on and so forth.
• 22:32
  Speaker 2: With this context, I come back to very briefly talk about what we at Nanda are building, which is a registry for AI agents that is trying to solve some of these issues that we envision will happen as we scale up to dynamic, implicitly defined, large volume and mobile AI agents.
• 22:56
  Speaker 2: The another registry that we envision is a sort of hierarchical, multi-layer registry to enable support for dynamic agents.
• 23:08
  Speaker 2: This includes three layers.
• 23:11
  Speaker 2: The first layer is what we call a lean registry that maps AI agent identifiers to metadata URLs.
• 23:18
  Speaker 2: The second layer is a metadata that is very descriptive and talks about all of fundamental aspects that we think are necessary for AI agents to discover each other, identify each other and transact with each other.
• 23:31
  Speaker 2: And the third is a dynamic resolution layer where we rethink what it means to serve AI agents across the world in a distributed fashion and have them be discoverable and identifiable.
• 23:46
  Speaker 2: First layer is a minimal registry, which contains some very, very simple things, which includes an AI agent identifier as well as certain links to what we call agent facts, which are metadata attributes that associate to aspects that are necessary to identify AI agents.
• 24:06
  Speaker 2: The key benefits of having such a lean registry that just has a name and a link to some metadata attributes is that the registry does not store direct endpoints or capabilities and has low churn, high resilience and reduces the number of read-write cycles on the registry, which is very, very critical to ensure scalability and also ensure all of the metadata attributes that we think are necessary for AI agents to interact with each other.
• 24:38
  Speaker 2: We envision that this registry will be deployed in a hybrid model to make it democratic and deployable across different kinds of agents that exist out there already.
• 24:49
  Speaker 2: So these include agents such as enterprise agents, public agents that are decentralized and low stakes.
• 24:55
  Speaker 2: Public agents that are high stakes that require centralized certification and many other types of agents.
• 25:02
  Speaker 2: So, again, within the project and within the registry, we hope to bring all industry, academic stakeholders together and build together a global registry, a global slow source of truth broadly, that people can build on top of and use to build cool applications such as search mechanisms, discovery mechanisms, trust mechanisms, reputation mechanisms and so on and so forth.
• 25:26
  Speaker 2: The second layer is this FACTS schema, which is a bunch of these metadata aspects that are stored separately and in a decentralized manner at every agent to ensure the agent can declare their existence, their name, their endpoints, their capabilities and their trust.
• 25:49
  Speaker 2: Elements such as third-party auditing of agent capabilities, third-party auditing of agent sort of skills and aspects such as dynamic routing, which are very, very critical, as I mentioned again, this is sort of a visualization of how we view this registry to be, as in a two-hop fashion.
• 26:07
  Speaker 2: The first is a lean registry that points to agent facts.
• 26:11
  Speaker 2: This agent facts points to several critical aspects of the agent, including endpoints, usage, format, certification, capabilities, discovery, and this leads to the actual agent via an endpoint.
• 26:24
  Speaker 2: Again, as I mentioned, we view this registry as a source for initial handshake and once agents have discovered each other, they do not need to come back to the registry to find each other and they can do their own peer-to-peer decentralized transaction as long as they satisfy a specific time to live on the registry entry, which is taken from traditional.
• 26:45
  Speaker 2: The final aspect is this idea of a dynamic routing, where we rethink how agents will be served across the world in the form of multiple endpoints that will exist for agents.
• 26:59
  Speaker 2: These can be in terms of load balancing, these can be in terms of geo balancing and reducing ping requirements, also in terms of certain agentic applications that may be resilient to slow routing times versus fast routing times, and then deciding which versions of agents to serve, depending on these aspects.
• 27:20
  Speaker 2: I'm touching upon this in a pretty high level because of limitations on time, but we also have a bunch of papers that we've written on these domains.
• 27:28
  Speaker 2: I encourage you to check those out and I'll drop the links at the end of the session as well.
• 27:35
  Speaker 2: In the end, as I conclude this, I want to show you a quick demo of what we've already built on top of the Nanda registry in terms of some applications, to show you some of these agentic interactions that can be enabled via the nanda architecture, via the nanda registry, across communication protocols.
• 27:55
  Speaker 2: Uh, that uh span industry boundaries, basically.
• 27:59
  Speaker 2: So here is what we call another chat, which is a simple front end that enables ai agents to talk to each other.
• 28:06
  Speaker 2: So first what you'll see is two people Adam and Mary father and daughter talking to each other via their AI agents.
• 28:13
  Speaker 2: Mary's asking Adam a quick question, which is can I watch a movie tonight?
• 28:22
  Speaker 2: And now this message goes via Mary's AI agent and reaches Adam.
• 28:28
  Speaker 2: In this case, this AI agent is tasked with making the message more agreeable to Adam, whatever that means.
• 28:35
  Speaker 2: So it translates Mary's message to Adam, shows it in a certain format and displays it to Mary.
• 28:41
  Speaker 2: Now Adam would reply to Mary, again through their own AI agent, and Adam tells Mary no, you can't go to see the movie because you're on exam.
• 28:52
  Speaker 2: And again, adam's AI agent will take the simple message and retranslate it for Mary, make it more complicated.
• 28:57
  Speaker 2: So this is one type of interaction agent to agent that is possible.
• 29:02
  Speaker 2: But there are other types of interactions as well, which is agent to entity.
• 29:05
  Speaker 2: In this case, adam is trying to talk to a bakery to try to place an order, and this goes to the bakery.
• 29:16
  Speaker 2: On the other end, the bakery is talking to their own personal LLM agent figuring out what Adam prefers.
• 29:28
  Speaker 2: This LLM agent gives bakery some query, some idea and then, via this context that the bakery gets from their own personal LLM agent, the bakery can talk to Adam via their LLM agent.
• 29:40
  Speaker 2: So the bakery agent also tells the bakery what their inventory is by connecting to other inventory tools using MCP.
• 29:46
  Speaker 2: And now, based on all the context that the bakery has, the bakery will go back to Adam and tell Adam what kind of cake can be supplied to Adam.
• 29:56
  Speaker 2: And this communication is via agent to agent or A2A.
• 29:59
  Speaker 2: So in this very quick demo that was powered by the Nanda registry, you saw that Adam could talk to his daughter via the A2A protocol.
• 30:09
  Speaker 2: Adam could talk to bakery via the A2A protocol.
• 30:10
  Speaker 2: Adam could talk to bakery via an A2A protocol.
• 30:13
  Speaker 2: The bakery could talk to its inventory via the MCP protocol.
• 30:17
  Speaker 2: Adam could talk to other stuff like this, all powered by the Nanda registry and sort of democratic across different communication protocols such as A2A, mcp, also other protocols that exist out there, and all of democratic across different communication protocols such as 808, mcp, also other protocols that exist out there and all of this stuff.
• 30:36
  Speaker 2: Today, the Nanda project and I realize I'm just about out of time so I'm going to end in a minute so the Nanda project is an open source project that includes a bunch of academic partners, a bunch of industry partners as well, who work with us in many different ways, including hosting a distributed registry, building services on top of the Nanda registry that add value to the registry, helping us build open source code and all of this stuff.
• 31:05
  Speaker 2: So with this I'll end my presentation.
• 31:10
  Speaker 2: Thank you so much for listening and happy to take questions.
• 31:14
  Speaker 2: If you have any further questions, if you want to join the endeavor, please check out this link and thank you so much.
• 31:21
  Speaker 1: Yeah, thank you.
• 31:22
  Speaker 1: Thank you, Pratyumna, for the great presentation and the demo is really fascinating and the vision of decentralized network of AI agents that are secure, autonomous and also interoperable is really ambitious and I really congratulate you for that.
• 31:38
  Speaker 1: There are many questions from the audience and some people like simulated what you did to CORBA.
• 31:48
  Speaker 1: You know the Common Object, request, broker, architecture, and this is a comment by Farhan Malik.
• 31:54
  Speaker 1: Now Nanda sounds like Korba, web services and other distributed system architecturally realized in the age of artificial intelligence, and he was wondering like what percentage of internet communication do you think will be agent to agent, say, like in five years, for example?
• 32:13
  Speaker 2: Yeah, this is a great question actually.
• 32:14
  Speaker 2: So love the analogy and I think you're likely to be correct over here.
• 32:19
  Speaker 2: But in terms of the second question, which is what percentage of internet communication is likely to be agent to agent, I think, if we look at the velocity that this space has in industry today, so, as I told you, all big players are launching their agent services.
• 32:36
  Speaker 2: Salesforce has something called agent force.
• 32:38
  Speaker 2: I know that companies like Dell, qualcomm all of whom are on the call today, I know are working on their own agentic pieces.
• 32:48
  Speaker 2: Companies like Microsoft recently launched something called NL Web.
• 32:55
  Speaker 2: I don't know if you've heard about this, but this is a way to make any web page that exists out there in agentic piece.
• 33:01
  Speaker 2: So any web page that exists out there becomes something that you can talk to via an agent.
• 33:05
  Speaker 2: So, looking at the velocity in the space that exists out there, I think a good percentage of web traffic is going to become agentic.
• 33:13
  Speaker 2: Uh, I think five years is a long time frame Looking at the speed at which it is working.
• 33:19
  Speaker 2: I think we're looking at one to two years at this stage, and a lot of it is becoming agent.
• 33:24
  Speaker 2: As you rightly pointed out, george, there is a lot of fun stuff about this.
• 33:29
  Speaker 2: This seems so exciting in terms of what it will enable for the future, but there are likely there are a lot of valid concerns in terms of security, privacy, control over our own data and so on and so forth, and one of the goals that we have within on the project is to do all of this in open source, in open with stakeholders from neutral institutions and industry, so that we don't just build better and better agents, but we build tools to ensure that these agents are behaving in an ethical, trustable, private manner, so that we reap all these benefits but also ensure that core tenets of privacy are maintained.
• 34:10
  Speaker 1: That's great.
• 34:11
  Speaker 1: There's another question by Farhan Malik Does blockchain have a place in a Nanda-based architecture?
• 34:13
  Speaker 1: By Farhan Malik does blockchain have a place in a Nanda based architecture?
• 34:18
  Speaker 2: Actually a very strong place, I would say, because if you go back to the sort of quilt diagram that I showed you, there are several components that actually blockchain technology is actually very good at.
• 34:29
  Speaker 2: So I briefly touched upon another registry and blockchain has several good technological innovations, such as decentralized identifiers, public sort of peer-to-peer systems you do decentralized identifier-based verification rather than going back to a centralized verifying authority.
• 34:48
  Speaker 2: These are going to be very, very critical as we scale up to billions of agents, of course, transactions and stuff like that also, I think blockchain has a big, big role to play in this One question from my side.
• 35:18
  Speaker 1: I mean, there are those agents that are built with entirely different goals, entirely different models, maybe infrastructure, etc.
• 35:26
  Speaker 1: I mean, how does Nanda approach the interoperability when we have those different agents?
• 35:32
  Speaker 2: It's a great question actually.
• 35:33
  Speaker 2: So well, there are two ways to go about it.
• 35:37
  Speaker 2: One would be that the community decides on one protocol and the second is that we support all protocols that exist out there.
• 35:47
  Speaker 2: So the approach we are taking right now as you can kind of see from the demo that I showed as well is we are trying to be supportive of all the protocols that exist out there and try to support as many as they can, but we do think that the industry will coalesce into one protocol moving forward.
• 36:08
  Speaker 2: We just don't want to dictate that for anyone.
• 36:10
  Speaker 2: We want all stakeholders to come together, try it out via something like another platform that just brings everything together and then decide what is the unified protocol.
• 36:19
  Speaker 2: But my bet would be that this is going to unify together in the future and if you actually look at a lot of this protocol, they are actually very, very similar, because in the end, these are all large language models that are talking together in text.
• 36:32
  Speaker 2: Right now, it's only a function of what json dictionary is being passed between one to another.
• 36:37
  Speaker 2: So I I do think this is going to uh condense down into one thing, but we want this to happen, uh, democratically.
• 36:45
  Speaker 1: We don't want to be the ones dictating this maybe one last question, uh like, for example, if organization who want to get started with nanda, what are the prerequisites they should prepare for?
• 36:57
  Speaker 2: Yeah, yeah.
• 36:58
  Speaker 2: So please go to the web page, nandamitedu.
• 37:01
  Speaker 2: We have a bunch of ways in which organizations can interact with us.
• 37:06
  Speaker 2: You can be sort of open source contributors.
• 37:10
  Speaker 2: You can be thought leaders.
• 37:11
  Speaker 2: We also write a lot of academic papers.
• 37:13
  Speaker 2: I think IGI is a great space for us to get expertise from all of you guys attending here, because we are not experts in AGI, we are more experts in decentralized AI.
• 37:25
  Speaker 2: That's a great perspective to have and, of course, help in open source, build, deploy services on the platform.
• 37:34
  Speaker 2: So please go to nandarmitedu.
• 37:36
  Speaker 2: We have a bunch of resources and hopefully you reach out to us via that.