Earley AI Podcast
In this podcast hosts Seth Earley invites a broad array of thought leaders and practitioners to talk about what's possible in artificial intelligence as well as what is practical in the space as we move toward a world where AI is embedded in all aspects of our personal and professional lives. They explore what's emerging in technology, data science, and enterprise applications for artificial intelligence and machine learning and how to get from early-stage AI projects to fully mature applications. Seth is founder & CEO of Earley Information Science and the award-winning author of "The AI Powered Enterprise."
Earley AI Podcast
Earley AI Podcast – Ep 76: How AI and Clean Data Power Smarter Search
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Bharat Guruprakash, Chief Product Officer at Algolia, joins host Seth Earley on this episode of the Earley AI Podcast. With years of experience helping organizations leverage AI to connect people with information, Bharat brings deep insight into the evolving world of AI-powered search, retrieval, and agentic technologies. At Algolia, a global leader in AI-driven search and retrieval, he helps shape what’s next in unifying data, building intelligent systems, and designing platforms that understand real-time context.
Key Takeaways:
- Misconceptions about Search and AI: Many organizations think large language models (LLMs) can handle all search needs, but true effective AI solutions require robust retrieval systems beneath the surface.
- The Role of RAG and Memory: Retrieval Augmented Generation (RAG) remains important, but the future is moving toward agentic architectures that require "memory" and stateful interactions, not just stateless search.
- Clean, Structured Data is Crucial: The importance of having clean, accessible, and normalized data stores as the backbone for any successful AI and search initiative.
- Experimentation and Innovation: Enterprises struggle with a culture of experimentation and face challenges in safely running and scaling AI experiments. Autonomous experimentation, where AI can test and optimize different approaches, is emerging as a solution.
- The Rise of Agentic Technologies: The distinction between generative AI (focused on content creation) and agentic AI (focused on task automation and execution), and how agents will soon drive more dynamic, event-driven workflows.
- Guardrails and Risk: Implementing proper protocols (like MCP and Google’s A2A SDKs) and guardrails is essential to ensure agents act safely and within business parameters.
- Privacy and the Future: As agents learn more about users than users know about themselves, privacy, transparency, and identity become critical concerns. The speed of change is challenging, but small, iterative steps help organizations evolve responsibly.
Insightful Quote from the Show:
"It's very risky to say, let's just boldly go forth into the unknown, right? I think you have to have experiments, right? You have to control them, experiments, but you have to Be careful about that and have a mechanism for managing that and for controlling it and for monitoring the results." Seth Earley
"It's okay to start small. Find the small places where you can improve...and keep multiplying them. Over time, when you look back after a year or two, you'll have a very different company from when you started." – Bharat Guruprakash
Tune in for a thoughtful deep dive into the challenges, opportunities, and responsible strategies for embracing AI, search, and agentic technologies in your organization.
Links:
LinkedIn: https://www.linkedin.com/in/bharatguruprakash/
Website: https://www.algolia.com
Thanks to our sponsors:
This is the Early AI Podcast hosted by Seth Early. Join us as we delve deep into the passions, expertise, and experiences of thought leaders and practitioners to talk about what's possible with artificial intelligence. The Early AI Podcast is sponsored by Early Information Science, your digital transformation journey with design and deployment of innovative technology solutions, as well as Vector. Now enjoy the show.
SPEAKER_00Well, welcome to today's Early AI Podcast. I'm your host, Seth Early. Each episode explores how artificial intelligence is reshaping the way organizations manage information, create value, deliver better customer experience. Today we're going to be diving into the future of search, retrieval, and intelligent agents, and how, of course, clean structured data is really important to the foundation for every AI success story. Joining me is Bharat Agula Prakash. He's chief business officer at Algolia, which is a global leader in AI-powered search and retrieval. He has spent many years helping organizations harness AI to connect people with the right information at the right time, as we like to say in Knowledge Manager. He brings a unique perspective to where AI-driven search is heading and what it takes to build systems that respond to real-time context. Bharat, welcome to the show.
SPEAKER_01Thank you very much, Seth. Very glad to be here and uh excited to have a conversation about where the world is heading.
SPEAKER_00I know. This is such an interesting area. And what I wanted to do is start with what are your what's your take on the biggest misconceptions that organizations have about search, about AI and search? And if you want to kind of roll data and content into that, you can. But give me the big picture. What are people not understanding or what are they misconceiving about this space?
SPEAKER_01Yeah, you know, I think the biggest, you know, we speak to so many customers, and you know, about two or three years ago it was, hey, search is dead. Today it is, oh my gosh, we need retrieval or search. And I think the big the big misconception is that AI does search and that LLMs do search, and that's partly because of the way it presents information to you. And so everyone thinks that you know LLMs can do all of it. In the traditional sense, search is about presenting all the options to you and you decide what the truth is. LLMs package the truth for you. So when you search for something, they sort of like package it up and give it to you. I think the you know, the the big conception is that you don't need retrieval under the hood for effective, generative, and agent first world. And today we're seeing that happen. Like suddenly now retrieval is one of the hottest topics. It's interesting for me, at least being in a search company, to see this sort of like evolution.
SPEAKER_00Yeah, it's been a pendulum and it's really swung from one extreme to the other. And it was something where people were saying, oh yeah, you know, the uh LLMs are gonna do it. And then what do we find? We find that number one, LLMs have a model of the world, but they don't necessarily understand your world, your products and services and solutions and so on, or your customers, your routes to market, your competitive advantage, your knowledge, right? All of that stuff is proprietary in how you differentiate in the marketplace. And an LLM isn't gonna do that unless we actually connect it to our sources. So that brings in this idea of retrieval augmented generation. And for those of you who are not familiar with RAG, retrieval augmented generation, it's saying to the LLM, don't just use your knowledge of the world to answer my question. Use your knowledge of the world to make this conversational and understand the user's intent, but answer the question based on my data, right? Get it from my source of information, my source of truth. Guess what? That's retrieval, right? That's search at the end of the day. So talk more about how that kind of starts in front. And people, to me, it seems like, oh yeah, rag, just use rag, right? Like, oh yeah, it's just rag, and that solves the problem. It's like, check the box. Tell me about why that's not necessarily true and why rag is not necessarily the answer. It's it's part of the answer, but it depends on so many other things. So maybe you can break that apart a little bit.
SPEAKER_01Yeah, and you give a really good explanation of what rag is. Um, you know, I will say that even rag is becoming passe now, it's no longer the thing. And the reason for that is because rag is is good when you're doing sort of like a stateless conversation. Now you're just in that session, you've had a little bit of a conversation, and it pulls relevant information and augments the generation. But what we're moving into, the world of agents, you need statefulness. And what does that mean? It means that agents can remember what happened the last time, a week ago, a month ago, a year ago, what you did on this particular topic. And so RAG is sort of like not able to help with that. And so I think you know that is where the limitations of RAG come in. The second thing is, you know, I just want to touch upon what you said, the limitation of an LLM. LLMs are very expensive to train, and so the history that they have of the internet is actually outdated. It's maybe six months old, it's maybe a year old, and to retrain it is very, very expensive, which is why LLMs have to depend on retrieval to get the latest information. So, you know, you put the fact that more information is being created all the time, and the fact that the LLMs don't have that information built into their world model, you need the ability to pull information, which is a search problem, which is a retrieval problem. And so, and so, yeah, RAG is uh is no longer in vogue, uh Seth, unfortunately. The new the new word is memory now, and everyone is talking about that because it allows agents to become more powerful and actually do useful work.
SPEAKER_00So that's interesting. And and it's funny because the terminology and the technology and the latest and greatest changes so quickly, and organizations have challenges not only absorbing that there's there's currency of the approach, and then there's the ability to absorb change, right? And of course, technology changes faster than the business can absorb, and then it's changing more quickly. Uh the business can only absorb so much, and so you're always going to have this lag. And so we still are talking to lots of people that say, hey, I just want to find my stuff, right? So I still think there is, even though you may say that that rag is passe, I think that there's still a fundamental challenge around retrieval and finding information. How do you reconcile that with what organizations are trying to do? When when you talk about memory, you're talking about retaining context, right? Over sessions, over longer periods of time. Uh, you know, I'm amazed when I am putting stuff into my LLM of choice, which can be Claude or ChatGPT, and I'm giving it very complex problems to solve, and I'm giving it lots of information. And it does amaze me uh at the ability for it to synthesize from what I'm giving it based on what it's already keeping in long-term memory, which I'm always telling it to update. I'm saying, make sure you put this in long-term memory. So tell me a little bit more about you know how these things are complementary, and then what does an organization need to do in order to get the best of both worlds? So tell me what happens if they ignore one or the other, and then what's the foundation for both?
SPEAKER_01So, you know, when we talk about retrieval augmented generation, we're really talking about retrieving content that is maybe only with a particular enterprise. And then we are augmenting and creating a generation out of that. When we talk about memory, we're talking about context, as you said. But what is context? Context is content that is relevant from the past, content that is relevant right now to the current conversation or the current task that you have going on. It is related to the interactions that you had before, so that you can you know bring that forward into what is the latest um, you know, up-to-date uh, you know, context that you have with the uh with the topic. So I'll give you an example. I lost my bike, I had to file an insurance claim, police uh records and and all of that. Uh it got stolen. So I had to work with Chat GPT. I'm like, how do you do all of this? It was a three or four week period that I was engaging with it. And by the third week, it started to forget things that I had told it before. Yeah, and it was really, really annoying. And so this is where you have to bring in both rag and memory concepts together. And that is effectively under the hood, it's a search problem because you're searching through past interactions and finding the most relevant things to bring up in that moment. And so you do need both. And I think organizations are really sort of like firstly confused. There's like so much, so much going on right now. All these buzzwords keep floating about. And then the second thing is they do not know how to use these. I think there's a little bit of um analysis, paralysis as well, like which one, like what happens if tomorrow there's a new thing that comes out again as well. So I think there's a little bit of that paralysis going on as well. And then I think the third thing is they understand that they have really important data, but that data right now is stuck either because it's not clean or it is stuck in static workflows. And so if you think about the world that we came from, it was all about static workflows, rules-driven workflows. When this happens, if this happens, do that. And now you're moving to a dynamic world where those rules are constantly being updated. And so I think enterprises are struggling to figure out how to, because it changes their entire product, it changes their business model, it changes so many things about them that they're struggling to figure out how to manage that leap into the new world. So that's what we're seeing constantly. And so the bolder ones are ready to like just say, you know what, let's just go into it. The more traditional, you know, maybe companies that have crossed the chasm have large businesses, they have to take care of their existing customers. Like that middle ground is very messy for them.
SPEAKER_00It's very risky to say, let's just boldly go forth into the unknown, right? I think you have to have experiments, right? You have to control the experiments, uh, but you have to be careful about that and have a mechanism for managing that and for controlling it and for monitoring the results. One of the things we're doing for an organization right now is you're putting that in place. We're saying, how do you manage and control your experiments and how do you measure the results of those? And how do you make sure that you're not, you know, spinning up applications that are gonna that are gonna uh run up enormous token usage and enormous bills uh for things that are going off in some part of the organization you may not know about, right? So, you know, there's the let a thousand flowers bloom, you know, idea of of AI experimentation. And there's also, well, let's make sure that we're not reinventing things all the time, that we're uh harvesting the best practices, that we're normalizing on some of those, that we're sharing those appropriately, the building standards. And so tell me a little bit about you know how you see organizations doing that. So, you know, with your technology, with Algolio, how does it fit in uh with this changing world and with trying to reconcile long-term memory, which is not just retrieval, but it is there is certainly some retrieval. It's retrieval of past interactions in context with the actual search uh mechanism. How do you resolve that? How do you reconcile that?
SPEAKER_01Yeah. And actually I'll I'll take a step back from what you said, which is to also say that companies I think do not know, enterprises actually do not know how to experiment. And there's two parts to it. One is just the culture of experimentation, it's very alien and discomforting to do that all the time. And then the second is the safety side of it. Like, how do you make sure it doesn't you know run off with token usage or create a bad customer experience and you know all of that stuff? So I think both are important. Algolia, we are providing a platform where customers can actually divert a portion of their traffic and they can build all of these new experiments, whether it's agents that can then retrieve using RAG or agents that can augment the, let's say it's a shopping agent, right? And the shopping agent knows that you set came a week ago or a month ago and you had bought something for your kids. And so now you've come in and the agent says, Hey, how did those how did those shoes for your kids fit? Are you here to buy socks or maybe jeans or whatever it is, right?
SPEAKER_00So that's or or or hey, let me try to sell you shoes again and again and again and again. Let me retarget you about shoes because you once looked at shoes and I'm and kids' shoes. So forever on, I'm gonna write usage, right?
SPEAKER_01And you and you just snagged the right point here, which is that personalization. That was the world of personalization. This is the world of alleged, alleged alleged personalization, exactly. This is the world of I know you did this last time, and so now what are you here to do? To are you here to augment that? Are you here to do something else? And so that effectively, when you then break it down, that agent has to first understand the context, which means it has to retrieve that past interaction. And then based on what that customer, that end consumer is typing in, actually, you know what, I'm not here for my son anymore. I'm here for myself. It then has to retrieve a whole different set of information that that particular retailer might have. So, you know, we're moving into that world now. And I'll I'll address the cultural side of experimentation. What we are also noticing is that enterprises struggle to have a constant experimentation mode. Like it's just either it's not in their DNA or it's just very hard to do. And we're moving into a world where actually AI can do the experiments for you also. So, how do we, you know, so what we're working on is how do we allow experimentation to happen autonomously so that you also don't have to like put your, you know, uh DNA, you you don't have to change your DNA to be able to do that. And so we're starting to see that. Can ex can AI rewrite the experiments? Can AI run a different experiment? Keep optimizing what's the best thing.
SPEAKER_00Can you give us a tangible example of that? So, you know, I think of this as it's almost like multivariant testing, where you map out what a uh, you know, kind of an artisan approach to the customer experience, right? This is what that ideal path is, this is the happy path, here's how I'm going to recommend products and so on. And then I have this componentized in such a way that I can change out different things. I can change out my hero image, I can change out my call to action, I can change the prop value proposition, it can change the target, right? I I have these, I have a messaging architecture that's componentized, that's like mad libs, right? Or Legos are mad libs where you can change out the different elements of it, but it all has a flow that that makes sense. And then the multivariate testing is switching those things in and out and optimizing for a goal, right? And the goal is greater clicks or conversions or sales or whatever it is, right?
SPEAKER_01Whatever it is, yeah.
SPEAKER_00Talk about how your you see either your technology or you see a similar technology achieving those experimentations in that experimentation. Is it a similar idea? What are what are the things that you're experimenting with and what are the parameters? How do you design that experiment and how do you execute on that?
SPEAKER_01Yeah. So we have an internal platform that we call the workflows runtime. And basically it is exactly like how you described it. It is made up of all the Lego pieces of and it can automatically switch itself in and out.
SPEAKER_00So you need those Lego pieces, you need that uh consistent architecture, you need the information architecture on the content or on the product information, and then you need a messaging architecture that tells you how those Legos fit it fit in with each other.
SPEAKER_01Totally. And right now, you know, we're using it just internally for things like hey, we have these ranking Lego blocks. How do you rank which is the best, you know, content that is being pulled out? Or hey, the Lego block, we need to pull out long-term memory and bias that more than short-term memory. So, you know, like these are all Lego pieces that you can sort of like declare, and the AI can sort of like say, okay, I'm gonna figure it out now. And now what we are working on is how do we in the future externalize this runtime? Really, we're thinking of it as a runtime, so that every Lego block in the in the planet, whether you're using some lang chain or lang graph or NA10, like can you plug these in? And can agents then sort of like use all these Lego blocks to achieve the goal that you're trying to get to? You know, you mentioned click-throughs, conversions, engagement, whatever is the goal. So we are in a world where massive experimentation can happen at a hypervelocity that was not possible before, and you gotta have modular Lego pieces to allow for that. And if we don't do that, do it that way. You will, as an enterprise, fall behind those who can do that. Right.
SPEAKER_00And then the fundamental so so if an organization is, if a business leader or technology leader is listening to this, and they say, Okay, this sounds reasonable. I need to have these Lego blocks, I need to do experimentation, I need to enable my agentic technologies to do this. Where do they start? Like what are they what what are the and then what's the beginning of something like this? Because again, there's experimentation, but you're talking about systematizing that experimentation or operationalizing that experimentation at scale using the tools, but there still has to be guardrails and parameters. So what's the starting point of of a lot of this stuff?
SPEAKER_01Yeah, it's a great question. And you know, our my goal is not to tell enterprises that they need to discard everything that they have and like rebuild. That is a not practical, it's not gonna happen. So the question is what have what have these enterprises been doing? They understand that all of this starts with data, and they need to understand that data needs to sit somewhere in a way that is accessible. So that has commonly been what for the last two, three years has been vector databases. But these vector databases are now not efficient anymore because of the rate at which information is being created. They can't keep up, they're expensive, etc. But what is important is that any you need some data store that can sort of like normalize across all the different databases that a company might have. They have you know SQL databases, they have CDPs, they have, you know, so many different the marketing data store, the sales data store. So you need a you need a common platform that can reach into those data stores and pull out relevant pieces. And that's sort of like what Algolia does. Like we can reach out, pull it in, and then use that to power agents and whatnot. So nobody needs to re-you know what is it, re-architect their entire stack. Like we've got to make the existing world work with the new world. Uh that's how that's how traditionally companies have been doing it. They've been starting with a vector DB, but the problem is once you get a vector DB, you need to build RAG, then you need to build memory, then you've got to scale the damn thing. Like it's it's a complicated engineering problem. So, what Algolia has done is sort of like abstract all of that away. And so then give you the ability to not change your own architecture, but give us the access into those data stores, because then you get a much more relevant, stateful, generative experience or an agentic experience.
SPEAKER_00So if we abstract from the product or the technology that you're talking about, and we say generalized, it's it's a federated model of reaching into these data sources, which is you know been around for a little while. Um, but you're saying you're not so you still need some platform to reconcile this or to build a unified index of some sort, right? And then you're saying that that's not necessarily going to be uh you're not necessarily ingesting into a vector space.
SPEAKER_01So you are, just to be clear, you are like Algolia under the hood runs the largest vector database, but you with but our vector database is a bit different. We actually work it as a search index, and it is very fast and it's very efficient, and it can be updated constantly. So you don't have to worry about enterprises, don't have to worry about that anymore. And so, but that is not enough. Like on top of that, you need the retrieval capabilities, the ranking capability. You need the ability to respect the business rules that the company has about their data. These are all things that in the open source world or the world that we're in right now, everyone, every enterprise has to stitch every one of these pieces together. And that's the hard part of doing it. And then after you've stitched it, which is already really hard, then you've got to play the Lego game on top of that. Like, hey, how do I do these experiments now? And like, is everything like working as it should be? Like sure. I mean, these are these are you need like vast resources as an enterprise to be able to go do those things. Yeah, yeah. And I think that's where they're stuck. They're like, ugh, okay, how do I do this? Yeah, and especially if you're not a tech first company, like you're stuck even more.
SPEAKER_00Right, right, right, right.
SPEAKER_01So abstract all of that away for enterprises and say. Here is a starting point. Point whatever data you need in. Declare your goals, whatever those goals are, and let us sort of like optimize it on your behalf.
SPEAKER_00Yeah. And again, we want to, we don't want to do product pitch here, right? You want to give people general principles that they can apply regardless of their technology stack. And I and I understand you're you have capabilities that can do some of these things, which is fine, which is great. And people should know about that, right? So they should investigate this. But I think the the core the core issue is constructing this in a way that is serving use cases that can be testable, right? Because you need to say, what are we trying to achieve here? And what is the business value that we're going to get from this? And if I switch gears a little bit and talk about agentic approaches, now, what is your, in your opinion, what is the hype? So we talked about kind of the whole problem challenge of retrieval and augmentation and memory and the idea of building these unified indices, indexes, and uh, and then whether, you know, whether it's a vector space, how that needs to be updated, whether you can create a uh a search index from that, and so on. That's that's all kind of good stuff to know. Let's switch a little bit and talk about agentic approaches and agentic technologies and how organizations are looking at that. And what do you see as the as the the big uh challenges and issues of confusion and hype uh in when people refer to agents and say, okay, we're not gonna think about you know these other things. Now we're gonna talk about agents. Now we're gonna build agents, and agents are gonna do all these things. And of course, agentic approaches have been around for a very long time, you know, at the simplest level. It's calling another source of information or executing an action. And some of the examples I've seen are very trivial. You know, they're adding two numbers or they're doing something ridiculous. Um, but talk a little bit about that, and maybe you can define in that context uh some of the the um uh mechanisms around that. So things like MCP is an agent to agent, but start off talking about what does that big picture uh look like and you know where's where where's the hyper, where's the reality?
SPEAKER_01Yeah, so I think I'll start first by just stating that there's a difference between generative AI and agentic AI. And I think a lot of people we're still in the world where there's a lot of like confusion about those two things. Generative AI, as the name suggests, is generating new content that didn't exist before, like summaries and you know, recommendations and so on and so forth. Agentic AI is about executing tasks. Now it might generate some new content to help it to execute the task, but it is all about executing tasks. And so in the world of agents, it's about automation of tasks, it's about um, you know, improving your productivity, your efficiency, because the agents can do the tasks that maybe humans were doing before and they can do it in an autonomous way. And so, you know, when we think about uh a simple framework to think about this, is in the world of search, traditional search, humans were doing requests and they would get a response. In the world of generative or sorry, in the world of agents, you're going into event driven, which means that hey, something happens, do this. And that's what an agent does, and it combines it with goal-driven. So I have a goal, if this happens, do that, and it's doing all of this autonomously. And so, firstly, I think most people confuse the two because what you see today put out as agentic solutions are like chatbots. Like, really, that's what you see, like little chatbots on the right, bottom right, and they say that's our agent. The reality is that that's Clippy from the past, if you remember. It's just a really fancy clippy. It's a really fancy clippy, but it is just a thing. Now, so first up, that's the difference between the two.
SPEAKER_00Now, I think that for generating content versus versus achieving an action or executing an action.
SPEAKER_01Exactly. A simple example would be: let's say there's a customer support agent that you're talking to, and it can do the entire return. It can process the return by itself, it can also trigger the shipping of any new items, etc., without any humans in the loop. And so it's sort of like is taking care of all of that and increasing the efficiency and reducing the sort of like the humans in the loop uh process that might be there. So I think you know that's that's a simple agent that we're starting to see. So I think you know that that's how I think enterprises are also thinking about it. They're looking at different workflows that are existing within their organization and they're trying to see where can we insert autonomous AI to replace that? And so that's what I think that's how most companies are starting. And so you will see simple ones right now because everyone is going to start with the simple workflows, and then at some point it'll get you know much more sophisticated.
SPEAKER_00Yeah. So when you think about the stuff that we've been doing for a while in the industry, robotic process automation, that is doing things that are rules-based. So, what is the line between you know RPA and autonomy? Because we're saying that you know, RPA can operate semi-autonomously, right? It can it can go off and do some stuff and you know, replace a person, you know, keying in information or you know, go from one screen to another or execute some predefined processes. So, when do we get into what we're referring to as an autonomous agent versus robotic process automation?
SPEAKER_01I think it comes to event driven. So, can you build rules or you know, process automation, which are really static workflows? Can you build all the static workflows in the world? Probably not. So, can you have agents rewrite or create new workflows on the fly because certain events that you didn't think about have happened? And so, you know, we constantly see this, right? Like there'll be some edge condition or some workflow that wasn't thought of, and then we hit that, and as humans, we're frustrated, we're we're stuck in a loop of some sort. But in the world of agents, like they can see that and say, okay, we need a new workflow to be introduced. So it becomes dynamic in that sense. And I think that what is important with that is the guardrails that the business needs to put in place for that, because you also don't want them to go rogue and create like random things that don't respect any of their how autonomous can it be? Yeah, right. Autonomous can it be? And I think you know that's the way to sort of like think about this world. Um I yeah, so anyway, I think that's and that's a challenge for companies right now. But I think you know, RPA is going to benefit from agents. So it's not just about robotic, it's also about now can you create some of these newer ones, newer workflows and newer automations dynamically?
SPEAKER_00Within prescribed parameters and um tell me, give me your definition of protocols for agent communication, such as MCPs. What what are your what are your thoughts around a model context protocol? How does that fit in? What is it? And then how do organizations need to think about MCP?
SPEAKER_01MCP is an API for agents. That's the simplest way to think about it. MCP is just an interface for agents to talk to one another and to get answers or information back from other systems and other agents. That's all that MCP is. And you know, it handles fancy stuff like authentication and permissions and all of that, but ultimately it's an interface layer. And what companies can do if they build an MCP, what they can do is parameterize what is those guardrails that you know this MCP can respond with to another agent. So I think you know, those are those are that's what MCP is, and I think it's it's gotten probably the most adoption right now when it comes to you know communicating with different agents and different data sources and retrieving different information. There is a new, there is also, you know, Google, I believe, is pushing a lot on what they call the A to A SDK, the agent-to-agent SDK. And that is, I would say, a layer of sophistication above the MCP. Because there you're actually defining actual agents as if they were humans. And you're you're sort of like giving them personality, you're giving them, you know, their their role, you're giving them their purpose, and these agents then talk to other agents. And so if you look at the Google's SDK, you can see you can define things like personality and things like that, which is, I guess, just the next evolution. Um, it's still early days, though, for that.
SPEAKER_00So when you think about you know, an agent and and you think about all the different parameters around, say, you know, personality, those are represented in an LLM in a vector space, right, with lots of parameters. In other words, there's there's n dimensions. It's the uh it's the the number of features that are handled in these large language models, which can be thousands and tens of thousands, right, of parameters. And essentially, that's a mathematical representation typically of content, of a piece of content. And then these agent behaviors or these personalities come from how those um parameters are being operated upon or retrieved based on some interaction, right? So you're saying, uh if I'm if I'm uh if I is retaining things in long-term memory about me, I can ask it questions. I can say, you know, based on um the the samples of writing I've given it, what do you know about me? What can you tell me about my style? What can you tell me about my preferences? So it has a representation, a mathematical representation of what it understands about me. However, I can only ask it so many uh uh specific questions, I can only look at so much explicit knowledge, which is which are the parameters around these LLMs and around agent-to-agent protocols or or MCPs, right? Because we're saying we're embodying a whole range of variables, a whole range of representations in this multidimensional model. However, we're kind of flattening that, right? We're abstracting that to say, well, what are the things I can define in language and in metadata? Yeah, but that is a fraction of what is represented in the large language model, right? So now the question is it's like this these languages models know much more about me than uh I can define or I can express. I can't define all those parameters, I can't even put them in English. In other words, they have mathematical representations that cannot be translated into language because it is because they are so abstract and they are latent, right? These are attributes that are that are defined in multi-dimensions that we cannot necessarily translate into explicit. It's a very lossy representation, right? So now the question is first of all, do I want that language model to know so much about me that I can't even define, right? That I can't express as language. In other words, it knows more about me than than we can communicate. And if we get those language models communicating with one another, they can communicate stuff that we as humans can't necessarily understand. This is kind of a long-term implication around large language models and around representations of agents. Agents will be able to codify much, much more than humans will be able to codify or represent. It just doesn't translate. What are your thoughts about that and what are the implications? Because it seems like there can be some ramifications of this, shall we say, that have privacy implications, that have autonomy implications, that have bad actor implications, right? We know that in controlled environments that research has been able to induce unexpected behaviors or maybe predictable behaviors, but surprising behaviors on the part of agents where they have acted in ways or LLMs, and I'm using those terms a little bit interchangeably, they've acted in ways that are not completely transparent or honest. They've been duplicitous. They've, if they're given goals that um that they need to achieve at any cost, they will resort to behaviors around blackmailing or lying. And so when you think about this, the fact that these things are getting more and more powerful, that they have representations that we can't understand, that they'll be able to communicate in ways that we can't necessarily understand. You know, does that concern you? And what do you think the implications of that might be?
SPEAKER_01There's a lot there, but I think let me try and unpack it. I think it's a it's a really, really important topic. And maybe I'll use a simple example to sort of like unpack it. So I grew up in a time early on when you would go to a physical human travel agent, if you remember those days. And you would sit down with them and say, Hey, I want to go to the Maldives or wherever it is, like, can you recommend me what I need to do, etc.? And over time you build a relationship with that particular human travel agent. Now, what happens is, if you're if for those of your listeners and viewers who can remember this time, over time, what happens is that there is a judgment that is being made by your travel agent on what your preferences really are versus what you think they are. And they sort of like get that because they've seen the interactions, they've seen what you like, and you might not even be able to express that to them in the same way that they know about you. So, in a way, we are back in that world where you have agents, and now I'm using agents as the autonomous agents, the AI agents, who know more about you in a way that you don't know about yourself. And I think what is disconcerting about that is humans couldn't, you know, in in the previous example, the those that human travel agent, even though they knew more about you, you knew that you could keep pace with them. I think in this world, the amount of information that is there and the speed at which they can actually learn feels very disconcerting to you as an interactor of that LLM or that agent. So I think that is a there's a there's a cognitive sort of like overload that is happening as a result of that. But at its basic core, that is what ultimately it is supposed to be a useful, better way of achieving the tasks that you want. Now, you bring up a lot of other points, which is okay, but how do you make sure that it's controlled? How do you make sure they're not hallucinating and blackmailing and you know deceiving and nefarious and all of that?
SPEAKER_00And I in the hands of a bad actor, they're gonna know enough about you that they'll be able to manipulate you without necessarily without your awareness. I mean, I think I think Instagram does that to me already, right? It's like I get sucked into this black hole of of of feeds and and things and products that I end up buying, and it's like, geez, they've hijacked my brain.
SPEAKER_01And so again, if you look at human previous human world, all of that was still also possible. You could have scams, phishing emails, like you know, viruses that could attack the systems and extract information, etc. It's just that the nature of them now is very different, and we haven't actually developed the technologies to help with all of that. So I think that that's the shift that is happening, and the fact that it's happening at a speed that is unprecedented, I think is what is making this cognitively like way more, it feels way more dangerous to all of us. But you know, my my my uh my take on this is that we're not going to reverse. It's sort of like we've let the genie out of the bottle, and so there are whole new economies of companies that are evolving to solve the problems that you just said. How do you make sure it's not blackmailed, nefarious, all of that stuff? Authentication is a big deal right now in the world of agents. Like, how do you is this agent the agent that I thought I was talking to? Like, even that is becoming a thing now. Like, agent identity is a big thing. So, and we have seen many startups who are working in that space. How do you create agent identity and you make sure that it is actually who they say they are and you know what are their parameters? So, you know, I think we're going to see more and more of those. I don't think that world is going to revert back to the previous world. I think the speed at which it is going is really, really fast, which is what is making people sort of like not able to keep up with it and feel nervous about it. But I think, you know, that is where the world is headed. And so we've got to figure out what that transition looks like. And I think there are many startups attacking these different problem areas.
SPEAKER_00Yeah, yeah. And I think that uh, you know, it's it's it's if we extrapolate, there's a great podcast, uh, Your Undivided Attention, that has had some researchers on it that have uh, and I have to get reference to those research papers, but where they've kind of just extrapolated where we're going, especially with um agents uh AI designing more AI and and the increasing power of these uh uh the systems and these tools. And for the first time, I thought, wow, this really does represent a threat, right, to to society and and to uh you know to our way of life, because we don't know where these things are going. You know, I'm not an alarmist, I'm a rationalist, I'm very much, you know, what's what's the immediate issue, but I can see, you know, as these become more powerful and as they learn more about us and as they are, you know, uh uh uh building these profiles of of who we are, depending upon who you're trusting with that information uh and how they're using it, it could potentially uh be detrimental to individuals and and to organizations. So it's something we have to be very careful about. And I think especially when you start looking at financial um applications, for example, around agent technology, as those become more embedded in systems, we have to really anticipate the potential unintended uh consequences and be able to manage that and be able to manage those risks. So I know it's a little bit off the original package of retrieval, but no, I I think I think harder, you know.
SPEAKER_01I mean, at the end of the day, under the hood, they all need to retrieve and then act on something. So it's not not really not really off. I would say that there, you know, in the last I would say 150 years, I think there have been three big changes that have challenged humanity. We had the industrial revolution, it changed how things are built, and it was a completely new way. People lost jobs, there were no guardrails back then on quality, like how do you do all of this? And we evolved. Then there was the internet era. Suddenly communications was everywhere, everybody did the communication was democratized. And what did that mean? It meant that someone in another part of the world got information immediately, and that had a whole bunch of implications. I think this is the next one that I can think of. Agentic, and its implications are maybe even bigger than the previous two ones. And so I have firm belief that we will evolve and put in and and and understand how to address this new model of the world, but it feels, I think it feels like we're rushing so quickly into it that it feels like, oh my gosh, can we even keep up with this?
SPEAKER_00Yeah, yeah. That is the challenge. And I think, you know, for any decision maker listening in, you know, it comes back to the basics, right? It's what are you trying to accomplish? What value are you creating? You know, what is your competitive advantage? How are you improving the employee experience or the customer experience? How are you adding efficiency to the process? How are you improving your your competitive advantage? And those are the fundamentals and those are the basics. And then we build on those and and try to still anticipate the the the risks and the potential uh downside of these things.
SPEAKER_01Yeah, I would say just to add to that, Seth, I would just say for all your viewers and enterprises, like it's okay to start small, find the small places where you can improve all of the things that you just said, but keep doing them more and more, multiply them. You know, over time, when you look past a year or two years of doing this, I think you'll have a very different company from when you started. And you might not realize it along the way, but if you keep doing them in small iterations, you will end up with a much better outcome.
SPEAKER_00Yeah, the evolution is happening. And it's funny, I was talking to another person who's going to be a future guest on the podcast about risk management compliance, and they're saying there's still some companies out there. There's like no to AI. Absolutely not. No, no, we're not going to do it. And it's like, oh, geez, you're you're you're not going to stay in business if you don't think about this.
SPEAKER_01There will always be people at all ends of the spectrum.
SPEAKER_00That's right. That's right.
SPEAKER_01And you know, that's just the nature of the world and enterprises. So, but I don't think we can ignore what is happening as well.
SPEAKER_00Absolutely not. Well, Brad, thank you so much for sharing your perspective on how AI is transforming search, retrieval, uh, agentic technologies. These are all such very important topics. And for listeners, be sure to uh subscribe to the AI uh uh podcast or the AI podcast for more insights. Brad, where can people find you? We'll put it in the show notes.
SPEAKER_01LinkedIn, uh um your company on LinkedIn, Bharat Guru Prakash. You know, that's the I I check it a lot. I'm more of a LinkedIn person than a Twitter person. So or or ex. So uh LinkedIn is the best way to reach me if you want to talk more about this stuff. And it was wonderful to be on here, Seth.
SPEAKER_00Well, thank you again. This has been terrific. Really, it's your interesting conversation. So I really appreciate your time. And uh thank you to our audience and thank you to Carolyn for uh doing all the production work behind the scenes, and we will see you on the next episode of the Early AI Podcast. Thank you.
SPEAKER_02Thank you for joining us on another deep dive into AI innovation. Tune in next time when we introduce another industry expert and discuss how to maximize AI in your world. The Early AI Podcast is sponsored by Early Information Science and Vector. To learn more, visit early.com. That's E A R L E Y dot com. Thank you for tuning in.