Why Context Will Define Agentic AI at the Edge – with Iman Anvari

Welcome to the Data Movement. I'm Paul Langston and on this episode I'm talking to Iman and Vary, the Director of Advanced Technology here at SeaGay. We're going to be diving into context, memory, and storage for enterprise AI systems. Let's get into it. Iman, welcome to the data movements. You're kind of one of my go-to guys when I have like a question about what's going on in the industry around us, so I'm super excited to spend some time with you today. Why you kind of just give me a really quick recap around your journey at SeaGay and kind of where you've come from and what you're what you kind of focused on today. Yeah, I mean, SeaGay, it's, that has been amazing. Just I've been very lucky with my career at SeaGay. It couldn't ask for more. I've been here for more than a decade now. So many different roles and in each one of them I learned something new. I started as a solutions architect or a systems engineer, I guess back then and kind of like you know, work on many different projects just building exabyte scale storage for other companies. It's one of those challenges that you learned so much, not just on the technical side but also from a business side and everything around it. So it's been quite a bit of time there and then I built and managed the team for a while. But you know, for me, I'm a product guys and I'm very passionate about building and innovating products. So that kind of naturally got me into product management for a while. And kind of along the say we're out in the last couple of years, I've been running the advanced technology team which for me, it's my dream job. We're focusing on what's coming next three, five, seven years and really trying to figure out how do we innovate and position storage and SeaGay for what's coming and it's just so much thinking, so much researching, so much work in with smart folks like yourself. So I'm just happy to be here and learning. Yeah, that horizon is so interesting. Like the five to seven year kind of horizon as you're thinking about your technology and how things are evolving and the pace, right, that things are evolving, especially at the moment, right, it feels like we're truly like something seismic. Big is kind of happening in the technology space at the moment. Like how do you how do you think about that? Back when I had here and I was a grant school. You know, I was I was focusing on robots and autonomous vehicles, right, and I was no dabbling with neural networks back then. And you know, when you've been kind of I reflect on this every few days actually, which is it might be too much, but it's very interesting because I look back and back then all of that technology was being talked like I think neural nets actually has been around for more than maybe like 100 years now from a concepts perspective. But they were not production ready, right, they were used in some use cases, but they were not really something that you would think to put in a robot. And then now you get up every day and you know, I'm very passionate about robots and automation and autonomy and all of that. And I see that this is becoming a reality. It's something that I thought would be happening maybe 30 years from now. Now it's like happening and progressing every day. So just I feel I feel like we're at this inflection points that our lives are not going to be the same before and after this period of time. So yeah, I mean, for even for me, I'm trying to stay on top, right, and I think a lot of folks do the same, but just being able to see where the technology goes and being able to shape it, that's what gets me excited and nervous sometimes even. The robotics thing is really interesting at the moment, especially coming out of GTC, you were there, right? Yes. What were kind of some of the kind of, you know, the the emerging trends or things that you learned at that show about robotics? I usually think about these things in an architecture, right? So kind of like what's underneath them was kind of like the end goal. And you know, for me, I feel like the ultimate end goal is that to have a robot, they can actually fold laundry for us. And the dream passed on as well, especially when you have kids. There you go. I say that as a Jess, but at the same time, it's truly a hard goal for a robot, right? To be to have a robot in your home that you trust, and it can do things autonomously. And it can be flexible enough that can go on fold laundry, for example, right? And I'm giving like this example, but they're like there were many different things about robots or different use cases, I would say. But kind of like back to GTC, there were two tiers that really were focused on robotics and robots. And the two tiers are like one is the fundamental hardware, right? Like you need a hardware that can do things, right? And I think over the past 20 years, if you look at companies like Boston Dynamics, they've spent a lot of research time on making it to work, making the body of the robot a human artwork, right? And I think we're closer than ever to have a robot that can be programmed to do very meaningful things. But the other part of this is the autonomy. And I think that's where the whole agent to KI comes in, right? And when I reflect back to GTC, you could see that the focus was agent to KI. How do you deploy this at different scales and in different places, right? And what do you need to get there? But you could see, you could already see so many robotic companies becoming more of a reality, showing demos that actually worked, even like surgery robots, right? Surgery robots today, like if you look at Da Vinci robots, it's very, you know, it's designed for precision, right? Because you don't want surgery to go wrong, if you will. But at the same time, the future that idea is that maybe you have robot or AI assisted surgeries and they were already showcasing some of this. So yeah, I mean, it's just mind blowing to see where we'd go and then what's happening. But at the same time, it just gets me so excited. I mean, for me, robotics is the center of all of this because it's hardware and software and infrastructure comes together in that one package. The robotics things really interesting because you have a bunch of different technologies converging in order to be able to execute that. And and one of them is like sensors, right? Video sensors is ingesting your sensor in order to understand the context of its environment. And the, you know, we talk about data on this show video is, you know, just a really big data stream that the computing architecture that underpins the robot has to ingest and process and make sense of and and store by how do you how do you think about about that? Any system that has to run autonomously, which robot is again like the best example of it, right? It needs as you mentioned, it needs all the inputs and we have video, we have like radar radar, right? Like different different data points that they all come together and a great example of this is an autonomous vehicle, right? An autonomous vehicle is technically a robot that's ingesting herabytes of data per hour just to be able to make decisions on the fly, right? So, you know, as as somebody who's been in the data industry for many years, video has always been at the forefront of data capacity and data streams that's always like the challenge you try to solve because it's like so big and the quality keeps going up and the number of cameras keeps going up. So I mean generally, when I think about robotics, I think that's actually one of the biggest problems that a robotics engineer now has to solve that you have all this video stream coming in. What do you do with it? Right? Where do you process it? Where do you make decisions? Where do you save it? So it's in all every time I try to in my mind at least, right? And this is maybe like I've been trained to do this, right? But when I think about these architectures, it all comes down to that data infrastructure and how do you approach it for that specific problem, right? Yeah, it's like the the mechanics of the robot, like the physical aspect to the robot is one engineering challenge. The other engineering orange to me seems to be the the data and the computing and storage staff underneath it. You know, it seems like a huge challenge when delivering at a scale, right? So I go back to our like sock pairing or laundry robot in Europe, in your home, like the core system at some point, which seems like an inevitability. Yeah, it's like how how does the computer computing architecture catch up to that level of scale or what's your perspective on what needs to evolve? I mean, thinking about how would robots work in an environment? And there's this new term being used called co-bots, collaborative robots, right? And it kind of makes sense, right? You have robots that maybe some of them are experts. certain things they do. And as you mentioned, that could be logical or that could be physical. Maybe you have a robot. Now, this might not be in your home necessarily, but in a factory floor that you have different needs, you have a robot that can deal with conveyor belt stuff and a robot that can actually go and grab something from a container, if you will. So there are two different use cases for one, maybe you need the arm for the other one, you need the suction cup. So that's kind of like the mechanical side of it. But then how do you connect all of this together, right? How do you make sure a robot is this entity that can think together and work together and kind of like achieve that goal? And I've been thinking about the architecture and the deployment model, and I think this is where the trade-offs become real. Like, do you fit everything in one all can-do robot or do you have many specialist robots with this like brain, which I call it like an edge micro data center that kind of like has your infrastructure, right? You have like GPUs in there, you have CPUs, you have compute storage. And then the data from all of this gets there, gets processed locally, is still pretty fast. And then you know, like gets into the robots. So you have a level of autonomy within those robots. And then you kind of like layer on a second level of autonomy on that edge micro data center. And I truly believe like that's going to be kind of like the architecture that's going to work at least in the next few years because there's like when you look at the trajectory of the hardware, I think more and more is being like kind of smaller, right? Like becoming smaller or shrinking down to fit inside a robot, but it doesn't necessarily mean that that's enough to achieve a certain goal. Again, it becomes like an infrastructure problem that you have to solve and figure out the trade-offs, the way it's power, all of that. Yeah. So tell me more about the micro data center. It's like in my head, it conjures up a certain image, which is probably not accurate. You put it's like, but it's about proximity, right? It's proximity to where the data is being should streamed processed, right? As opposed to massive kind of centralized exabyte scale data center, unpack the kind of the problem with a really like hyper-local robotic, like the one we were just describing and then the infrastructure, needed to support it. I can actually maybe give you a silly example. This is something that I've been trying at my own house, right? And it's applicable to any of these architectures, but so I had an old robot and it doesn't have a lot of computer on it, right? So it's pretty old and I was thinking about, okay, how do I actually make this robot agentic where it can do things? And I didn't have a crazy end goal for me. The end goal was that can I keep my cats pre-intertained with this robot, but I wanted to be more autonomous. And it was a good way for me to try some of the new technologies we've been working on actually. So over the past couple of years, really, my focus has been AJI and bringing everything down closer to the edge, right? And as you mentioned, it becomes very important when you have to make decisions fast because physics is still physics, right? If you need to send something even through fiber to the cloud, you have to pay for that round trip and it needs to go there and come back, right? Which is fine. It works for certain use cases, but for something like robots, you probably don't want that round trip latency. So you want it to be close to the robot to the user and you want the decision to happen very fast. For example, in our own factory, we have a rule of like two seconds that if you need to make a decision within two seconds, you probably don't send it to cloud and back. You have to make that decision locally. So kind of like back to the example was that we've been working on AJI devices where you can bring Agent to KI to the edge and be able to kind of like load it and make it run without any internet connection. Now you could connect it to internet if you wanted to, right? We don't stop you from doing that. But the idea is that if power goes out, you have a backup power. If internet goes out, you're not reliant on that. So this idea of micro edge data center is that you have these building blocks, which I call AJI device, right? And you could use one of them, five of them, ten of them, right? You're not building a full data center, but you still have a small building block that you could host your agents on and you're just been in the home, by the way, or is it last year? Yeah, I mean, right now it's in my garage. I'm actually hosting it in my home, but it could be anyway, right? So I think I'm using one one building block because that's enough for my silly use case. But in a more factory environment or industrial environment, you might use five of them. So it's really in the spectrum, right? When I think about Edge is an aspect room where it could be your house or it could be a factory or it could be far, right? It's like it really depends, but the whole idea is that it's far away from a big data center and you need to make decisions fast. And this is where edge AJI and the edge micro data center idea comes into play where you have many sensors, you have many robots and they all talk to this mothership of sorts, right? And they can make decisions. But it's interesting because when you when you trend, you look at the trend of computing architectures and we talk about this a lot at SEGA, is you know, the cloud kind of took a lot of computing tasks and centralized them, right? And then what you're talking about is decentralizing computing to the extreme. Yes, exactly. Yeah, I mean, I see this as a cycle that we go through, right? Like server and clients, like we always kind of like, if you look at the past 20 years, we go through the cycle that's on from becomes powerful and you could do more with it. So you do because it's more deterministic, but I think, you know, generally the architecture is always been hybrid. It's just a matter of the availability of resources. So I think as every time you see a hardware innovation happen, you kind of like have more to do on prem and we take advantage of it. And there are times that that becomes not enough and you start pushing that workload to the cloud. In my mind, it's all going to be hybrid. It's going to be a hybrid edge and cloud and hybrid storage, right? And we talk about this a lot like how storage layers and tiers matter to the memory of those agents and the context of those agents. So there's no one size fits all at least that's the way I see it. I feel like all these use cases, they will have a sweet spot and then they will evolve in the next few years. But for the robotics use case specifically because of the data lifecycle and where the process it needs to happen. For that specific use case, then that hybrid model is so critical, right? In managing the data across the lifecycle. You just mentioned context. Contacts. Can you unpack what that means for us? Yeah, that's my favorite topic nowadays. Me and you actually, we talk about it quite a bit. Yeah, we talk about this a lot. Yes. Very relevant. Even at GTC, you could see that context becoming a first tier discussion now. Like last year, it was something that you needed to think about. And of course, when it comes to large language models and agent AI is very important, but now it's actually becoming the bottleneck for AI agents and autonomy to become real. So context, if I want to simplify it, I mean, the best way is to use a human analogy, right? So at the end of the day, agents are trying to mimic real humans from a kind of thinking and acting perspective. And they need to have a good memory because even if you have the best engineer in the world or the best partner in the world, if it doesn't have a good memory, it's not going to do a good job. So that really comes down to that. The context and memory of agents is what enables them to go from just good to this amazing autonomy. And it's inherently storage problem because that context needs to be stored somewhere. It needs to be accessible and it needs to be searchable. And that agent should be able to look at that and find the information it needs in as little time as possible. So putting off that together, it means that if you have the right context, then the right amount of context, then your agent is going to do amazing things. But if you feed it garbage, it's just going to give you garbage, right? That's yeah. And it's it's so interesting thinking about this as like because we're more and more we're using kind of chat chatbots and agents and there's a high degree of human intervention like with the human is prompting the experience with E.A.I. And so you can see first hand when you're prompting and you know you've already told it something and it kind of forgets or there's an issue with the user experience, you can see it. And then you can can kind of course correct it because you know you're managing the interaction. And so that for me, that's how I kind of, you know, visualized this idea of context. But then when you think about agentic AI where you're trusting the AI, there's minimal, if no human intervention, then you think about doing that at scale and then you're handing over operational tasks in an enterprise, you know, potentially mission-critical enterprise tasks, then the idea of context and inherently data is so so critical, right? The way I think about it is that at the end of the day, large language models, which are the core of agentic AI, right? Like that's still the brain. They're stochastic, right? They're by design not deterministic. So as you mentioned, if you want to delegate a work to them, and even humans are probably stochastic, right? But if you want to delegate, you want to delegate it to somebody that, you know, they will do a good job nine out of 10 times, right? And of course, the more critical that task is, you want that percentage to go higher. Now, there are different ways to ground and control these agents without it's kind of like being there and chatting with the men. You mentioned, right? Like as you're using your chatbot, you could totally see if you give it bad information, it goes off the rails and sometimes you have to even start over, right? You have to remove everything. Or it's trying to synthesize, it's trying to fill in the gaps. Like where it's missing like context, it's trying to do its best as filling the gaps. Because I guess it doesn't want to go further back, you know, because it must have some, this is my assumption, there must be some parameters that have been set in the back end to stop it going because it has the data, right? Because I've fed it to the data before. So either it's just not going back further enough in far enough in time to go and retrieve that because of some, I don't know, cost parameters or something that's going on that's preventing. That's an interesting one. Yeah, yeah. So that's actually a known problem with LLM. So there's a benchmark called haystack. And the whole idea is that can you find that needle in a haystack, right? If you give it a lot of stuff, right? Can you go back and find that specific thing that you asked 20 minutes ago, two hours ago, right? And if you look at the new models, LLMs, there actually is quite a bit of focus on making this better and kind of like getting better at finding that. But even if that happens, that only applies to the active context, right? The hot one that's in memory and it's like sitting there, right? But what if that agent has to think about what happened two weeks ago? That's not going to be there, right? And that's where this whole idea of this tiered context where you have a short-term memory, just like humans and a long-term memory, the short-term memory is right there. You just have to get better at retrieving it. And then this long-term memory is not there. You have to dig for it. You have to spend energy even when I need to think about something from 20 years ago, right? I have to kind of like close my eyes. As old as I'm getting, it's getting harder. But I have to kind of like dig for it and go back and really try to find that information. So exact same architecture applies to agents. So the question becomes how do we, as an industry, make it easier for these agents to retrieve that data and make that data available to them, right? And I think that's like the baseline. As a storage company and a storage guy, I see so much data being thrown out that we always think about it, right? I get in past 10 years, I think, me and you probably talked about this more than 10 times that the data that you're throwing out today might be super beneficial in five years. You just don't know about it, right? And that's kind of like the baseline. And then how do you make that searchable and indexable and easy for that agent to retrieve? That's the problem everybody's trying to solve. Yeah. I even think like five years. Yes, true. I agree. But also like two weeks or like a month. Yes, for sure. If you have this kind of long context window with lots of hands and you know that you're taking in the workflow that you're interacting with the AI, even a piece of even a piece of data or piece of context from a month ago, if the underlying architecture isn't built to enable retrieval of that data, then that's going to have an impact on outcome, right? It strikes me that this is fundamentally shifting the way data gets teared or needs to get teared. But it strikes me that storage is becoming more active because of this workflow and the lines between like the way the industry defines memory components of that tier and the way that the industry is defined storage and that kind of tier within the architecture of becoming increasingly blood. They are. Yeah, absolutely. Yeah. I mean, I feel like the general storage pyramid which starts with the fast and small storage to the, you know, to the bottom of like at the top and then very slow and big storage to the bottom still applies. But as you mentioned, the workflows are changing. We're going from just these applications that they would query the storage to these long living agents. They're they're running. They're thinking they're deciding what they need in the moment. We don't tell them go fetch the state. I mean, we might if it's like human in the loop, right? But if you're doing a full talk loop, they decide they might even talk to each other agent to agent and decide what information they need, which is fascinating. And the other part of this is that nobody really 100% understands how LLM still work because by design, they're like neural nets with an attention layer. So they're black boxes. And you see even anthropic is doing research on how do they behave? How do they decide to respond to you? Right? So we don't know. And when we don't know, we don't know how to design for it. So it becomes this chicken and egg problem of how to give them the memory they need. And when I was thinking about this a couple of days ago, Paul are like, it's a simple problem. You have to give them the context they need when they want it. Right? Like that's it. But you achieve that. Right? And that's that's that's really the big problem. Like how do you tear it up and down? And how do you save everything? And how do you index it? What formats do you need? Right? All of that. I think those are the constraints for storage and memory design for agents. Strikes me that the way that things have played out with large language model development is there are a few there are a few companies that have you let the charge right of at the front ear of, you know, large or true large language models. And so you you mentioned earlier is a like when it comes to agentic AI agentic systems automated systems in the enterprise, it strikes me that those language models will be licensed rather than developed may license for adaptation for specific use cases and an industry workflows in the enterprise. And so how those get deployed at scale? Right? In those very specific agent use cases, IT service, for example, service depths. A lot of it a lot of the kind of uniqueness of how that agent gets activated and deployed is based on the context, right? The proprietary data that's the agent and the LLM adaptation is fed. I would argue that the customization of the personality of agents primarily comes from the context. So if you think about LLMs, so nowadays we talk about harness engineering. What is a harness? A harness is everything that goes around the LLM to make it behave a certain way, right? So it could be it's the system prompt. It's the tools that it can use. It's the guard rails that it has. All the like skills and clawed code, for example, these are all part of a harness that somebody builds to make an LLM an agent that actually achieves a goal for you, right? Because that's what we want. I mean, everything we do today is cool. But unless it's solving a problem for you and it's delivering that return on investment long term, it's just a toy or a demo, right? So to turn it into a product, you really need to show that ROI and measure it. And in order to get there, this agent has to be customized. And context is the first thing you start with. You give it the system prompt. You tell it what tools it has access to where it lives, right? Is you're in a factory or in a home? What are you trying to achieve? What are your goals? So it's really the core that memory is what makes that LLM an agent and a personality. And in my mind, that becomes one of the most fundamental, not, I mean, it is a problem, right? I get a sense that you want it to make it as easy as possible for that agent to access that context and make it as relevant as possible, so that it actually does what you want. So solving that problem becomes like a core focus for a lot of folks. Especially if you're not just building LLMs, right? So the big frontier models, they're going to keep building amazing AI, and we're going to use them. But if you want to use it in a factory setting, how do we go from a generic model to a very vertical or specific agents? If I work for an enterprise organization today, I'm listening to this and I'm thinking, okay, what, like, I'm thinking about my AI strategy long term. What are some of the things I should be thinking about as I kind of nap out that, that red knot. So the way I would approach that or answer that is that we've got to go back to fundamentals. We know things are going to change. We know LLMs are going to get better. Hardware is probably going to get better. But the core problem remains the same that if you don't have the right data and the right context, you're still going to have a problem. So if I was building an enterprise or just like thinking about my IT strategy or enterprise strategy in the next few years, my focus would be, how do I collect relevant data? Keep it clean. I mean, clean data is very important. Actually, it's like one of the hardest things, right? It's something that people don't love talking about it because it's not glamorous. But like making sure that your data is clean has the right metadata. It's searchable, right? And easily accessible. And then you kind of like have this strategy of infrastructure that you could modularly plug in new types of databases and stories. So for example, today, we focus a lot on graph database and vector database. Two things that the agents use to grab information into their context. We didn't talk about these two years ago, right? So two years from now, it could be completely different. But if you have good data that can be transformed and you have a modular architecture and infrastructure, then it's easy to adapt and enable your agents and the future of agent to AI in your environment. And I truly think that is the future. I mean, we are at a point that, you know, even personally, if I have, I don't have five different agents doing something at any single time, I feel nervous. I'm like, I'm not, I'm underutilizing my agentic platforms. What are your agentic workflows that you have working for you today? Yeah, I have one agent that I mentioned that I kind of like attached to my robot. That's kind of like for the home stuff. And yeah, it walks around. And whatever sees one of my cats takes a picture and sends it to me. That's just completely goofy. But you know, it's a good way of like playing with robots. But for work, I have a research agent that tries to keep up with everything that's going on with AJI. So I continuously run it. It looks for new sources, things that are going to the industry. And it will create reports and update me every day and tell me if there's something that is very important. I need to reach out. I use agents for my, all my scheduling now, which is great. That was like one of the pain points we would get on a call and you're like, okay, can I do this? Now I use co-pilot. And I'm like, okay, just go and find these like block some calendar and it does it every day. It's amazing. I could see in our own organization, like how you go from just a chat bot to something that can actually do things for you. And for me, that was interesting where with co-pilot initially, you just have chat bots and you're like, wow, this doesn't really do much. But now it has access to all my data. And it's amazing. I can, I can tell it, I basically tell it to go, like if an email comes in and they're looking for and a specific time frame to meet, I don't want them to wait for me to check my email and figure it out. I have certain times allocated and it goes and makes sure that the topic is relevant. I kind of like said, some, something in its context to say, okay, for these things, these are hot topics, schedule is right away. And it'll go and find a time that works for me and my team and it will schedule it. And it just lets me know. And it's just it's amazing. I used to spend so much time on this. And I absolutely hated it. And I always felt like I could, I could have a personal admin do this, right? And now, you know, my agents can't do this. Yeah. Yeah. But yeah. And then I think the most common one nowadays actually unlinked in I put a post asking people what they use agents for today, right? That was like, before coming here, I wanted to have that answer. I did your call, by the way. I did. Great. Great. Yeah. We got a few, few answers, but what was interesting was that it was 50/50 split between research and coding. So you could see that where we are today. And I think that's very like the two areas that most people are using it for. But I want to do the same poll maybe six months from now. I see how that involved. It's relatively new. So I'm still like refining, refining it and curating exactly what I want to see and filter out some of the stuff I don't. Because if it's too much, then I just it's just there's too much to pass through. Right. Right. Two verbose sometimes, right? Yeah. Yeah. Yeah. And you still have to steer them, right? I feel like that. But that's where again, building context, a personal context for you, for example, as you kind of like look at these behaviors and say, okay, next time, don't do this. And if that works well, right? And it's actually the LLM grounds itself into that context. Then suddenly you have this amazing agent that does a really good job for you. Yeah. Very cool. Iman, it's been such a fun, interesting conversation. Do you have any, what question didn't I ask you that you wish I'd ask you? I want to plug and do a shameless plug about the work we're doing. So we just showed our proof of concept for our DJI box. We call it DJI Pro. And we've been working on this for a couple of years. And I say we, I probably have done least amount of the work, the most of the work happens in our amazing research group. And we've been really focusing and watching over the agents are going. And we finally got the proof of concept out in public at GTC and showed it to people. And it's always nerve wracking, right? When you're kind of bringing a product like that out and you're talking about it, it's very, you're worried about bad feedback. But of course, you want to have it, you want to understand it. But I was so surprised because we had so many people coming in with interest and different use cases, right? Some some robotic, some smart cities, some were just doing home automation that they wanted something similar. I'm not going to say exactly that, but like they would probably buy and take that today if they could. So the interesting thing about it is that we actually have deployed it internally in our own factories. And we already have seen really good results. But now we want to go after a wider audience and more agentic approach. And what's a novel about it is that we're bringing everything together in a very small portable box, right? So you have the AI, you have the Nvidia chip. But we married it with multi-tier storage. And this multi-tier storage in a small box is really hard to do. It's easy to do it if you're in a data center because you get scale and do different things that get very creative. But in a small scale, you only have a few knobs you can turn, right? And you're dealing with a lot of trade-offs, power, cooling, eating, all of that. So I think we finally hit a sweet spot where you could have hundreds of terabytes of storage and context memory plus thousands of terroflops of compute married together. And what that really means is that you could run fully autonomous agents. And actually that's where I'm running my robot agents now, right? So fully autonomous agents. This is what you're describing is the micro data center, right? Yes, absolutely. Yeah, it's like a tiny, I mean, yeah, it's even smaller than a PC, but it's technically a full data center that you could run an agent on, right? And it connects to cloud if you need it to, but you don't have to. It's just completely secure. So I'm super passionate and excited about it, so I had to do a shameless plug here. Yeah. I love the cat feeding. Yeah. Application for as well. And this is like, you can see some of this on your LinkedIn, right? Yes, yes, actually, I'll be probably posting some pictures of that, but yeah, it's amazing. I mean, I was able to revive a very old robot. This is an Ankyvector. It was a startup many years ago. Some smart folks started it, but it doesn't have a lot of local compute, but now I can offload this intelligence of the agent to the micro data center. So somehow my robot is alive now and it can think and it's just fascinating to see how easy it was to get here. All right, and on on the data movement on our show, we do a five-question lightning round. So let's just get through these really, really fast. I'm really curious about your responses. So question number one, what is the one AI tool or technology that you wouldn't or couldn't do your job without? Codex. Codex CLI, that's where I live in. That's the one. Okay, question two, are we heading towards a world with fewer, more powerful language models or lots of smaller, more specialized ones? I think it's going to be a mix of both, both more powerful centralized one and a lot of more specialized ones. Question number three is the edge becoming more important because of AI or is AI becoming more important because of the edge? I think the edge is becoming more important because of AI, because now you can actually do a ton of things at the edge, which you weren't able to do even two years ago. Question number four, what is the biggest unlock for AI, better models, or richer context? Oh, controversial, I'll go with richer context. Nice. Yeah. Yeah, I'll give some heat on this. Yeah. Yeah, the answer is obviously both. Right. Yes. I like how you went there. Yes. Okay. Um, and then question five, final question, complete this sentence. The future of data is mind blowing, amazing, um, iman, thank you so much. I learned a ton as I always do from talking to you, but what I love talking to, the reason I love talking to you is like you have, you have all of these ideas and, and understanding of what's going on, but then you're also embodying it and doing it in real life, like feeding your cats with an AI agent, but it's, it's, it's very, very cool and inspiring and interesting, and I appreciate your time this morning. Thanks, Paul. Thank you so much for having me. Always good to talk to you and, uh, yeah, you know, me, I have to, I have to get hands on to learn. So for me, it's always, uh, um, but it's, that's the beauty of AI. I think nowadays that it's easier than ever for everybody to get hands on, right? So even robots is something that anybody can get into now. It's not just for engineers anymore, and that's, that's just fascinating to see. Awesome. Thank you, sir. Appreciate you. Thank you, Paul. That's it for this episode of the Data Movement, a podcast from Seagate. Thanks to Aman for joining us and thank you for listening. Subscribe for more conversations about how data is moving the world forward. [Music]