.jpg)
Advantest Talks Semi
Dive into the world of semiconductors and Automatic Test Equipment with our educational podcast, Advantest Talks Semi, where we explore the power of knowledge in this dynamic field. Hosted by Keith Schaub, Vice President of Technology and Strategy and Don Ong, Director and Head of Innovation for Advantest Field Service Business Group, at Advantest, this series features insightful conversations with experts and thought leaders in the industry.
In today's fast-paced environment, continuous learning is essential for staying ahead. Join us in these thought-provoking discussions, where you can learn about the latest trends and cutting-edge strategies being used in the semiconductor industry. Explore how innovative technologies are revolutionizing testing processes and shaping the future.
Stay updated on the ever-evolving semiconductor industry with Advantest Talks Semi, and gain exclusive insights into the future of technology.
The views, information, or opinions expressed during the Advantest Talks Semi series are solely those of the individuals interviewed and do not necessarily represent those of Advantest.
Advantest Talks Semi
Inference is shaping a major role in AI's future
Bringing AI to the Edge: Dr. Bannon Bastani on Enterprise Computing's New Frontier
Thanks for tuning in to "Advantest Talks Semi"!
If you enjoyed this episode, we'd love to hear from you! Please take a moment to leave a rating on Apple Podcast. Your feedback helps us improve and reach new listeners.
Don't forget to subscribe and share with your friends. We appreciate your support!
Hello and welcome back to Avantest Talk Semi, the show where we unpack the breakthroughs shaping semiconductors and the world of power. I'm your host, Don Ong. Today we're zeroing in on a frontier that's shifting from password to business backbone, enterprise, edge computing, Picture data center horsepower running right on drones, robots or a task handler on the fab floor Decisions made locally, instantly, securely. Guiding us through this rapidly evolving landscape is Dr Bannon Bastani, CEO and co-founder of OpenInfer, a company that is redefining how we run AI in the world. Welcome, Dr Bastani. Thank you. Thanks for having me, Dr Bastani. For the listeners new to the concept, how would you define enterprise-age computing and why does it matter right now?
Speaker 2:Good question. Today we are seeing AI and AI being an inference usage of AI application being more adopted in our daily life. But we are constrained. Constrained by do I have to have connectivity? Do I need to share my private data? Is it reliable or not? And enterprise today is struggling with those three components. Do I need to share always my context, my data, in order to get access to AI? Why do I need to be paying for those inferences at such a hefty price if I have some of those computing, and can I always assume that infrastructure, that connection is going to be available if some of my applications are going to be fairly needing to operate?
Speaker 2:Real-life? Edge for us means things, surfaces, that are closer to us, closer to our hands, closer to our daily usage. It can be your phone, it can be your computer, it can be your companion, the robots, the car you're driving, the self-driving, the car infotainments. Or it can be the companion, the services, the assistant that exists around you. You go to retail, you shop, that's surface that you're actually doing the purchases. Edge for us means edge devices. The surface is close to you and for enterprises it's a significant space to address their data privacy, their reliability and cost of operation.
Speaker 1:I'd like to push this a little bit further. So H versus cloud versus endpoint, where does the line blur and why should enterprise care?
Speaker 2:Rather we like to say there are computing available everywhere. The world should move into a more continuous computing. Computing can have different capabilities. The farther out maybe you have more capable computing up in the cloud, the closer it comes to you maybe on your pram, maybe even on the, maybe on your pram, maybe even on the local machines you have in your telcos and closer to your devices, the things that you're holding in your hand, things that you're wearing. We should be able to use all those surfaces, all those computing, and each will serve different purposes. The more closer they are to you, they're going to be more responsive. They're going to you're not going to share as much data broadly with a lot more companion, but perhaps there's a lot more computing capability. So, more companion, but perhaps there's a lot more computing capability.
Speaker 2:So I'm looking to that future that there is a continuity between computing. Today there's a lot of focus on the cloud and we like to make it easier and easier to take advantage of other computing that are closer to us. And edge is a very unutilized edge. Particular edge devices are underutilized computing ecosystem that we like to see open up and that continuity gets built between the cloud all the way to edge devices A little bit personal question between the cloud all the way to edge devices A little bit personal question.
Speaker 1:So what personal aha moment convinced you that the real-time inference belongs at the edge and not in the cloud?
Speaker 2:It started perhaps at the journey I had at Google, google X. We were working on a project that we were evaluating scalable computing, if that makes sense, and based on a smaller compute ecosystem. At that time, we realized that how computing is being used in 2013 time periods is very traditional. It's being used as a monolithic processor. Think of it as a simple calculator versus a computing. A calculator is you tell it what to do, it responds. Calculator is you tell it what to do? It responds, whereas you want it to actually be able to reason, do some operation.
Speaker 2:And it was at that time that I realized the capabilities exist. The silicons are getting quite a bit powerful, but the system is missing. The system to take advantage of this smaller computer missing and there is a lot of potential is left in the background. And it was the aha moment in my mind. It's like wow, if we could utilize these processors more efficiently, if we could think about how systematically we use this, we can actually build a supercomputer with these Lego designs, with these smaller pieces built together. It may not be the cheapest one, but it allows you to scale, it allows you to adjust or it allows you to take advantage of the computing that exists around you. So I started. I actually we worked on that project for a while and when I joined Meta, the Reality Labs division was called Oculus at the time, I really pushed for that vision. I realized still we were betting on PCs at the time to be the future of virtual reality rendering, whereas the headsets are being looked at as a very low powered, not capable ecosystem. If you looked at the capabilities, though these devices, those headsets at the time could do significantly more specialized operations and if we empower those specialized operations on the headset, then we may open up new capabilities. Open up new capabilities and that's, for me, actually started the vision that later on became Oculus Link. Link became the foundation that hey, we can bring PC type of quality VR rendering to the edge that is sitting on your head while meeting the VR requirements. And how we handle latency of streaming PC quality to headset can be mitigated with AI. Ai can run efficiently simply on these headsets, using all the little computing units that exist here to mitigate the latency that may exist on the PC being streamed to the headset.
Speaker 2:That journey was amazing. We actually demonstrated that you can actually have PC-level quality on your head being streamed from the PC but you're not connected to it. You can now walk around and these headsets, these little computing units, can actually enable you to have amazing compute capability things that we didn't expect before. But it needed to be thought from ground up. What should happen here, what should happen on the GPU, what should happen on little CPUs? What should get streamed? How do we compress what additional information comes in? And that foundational thinking was critical to unblock headset computing without needing to be attached to a PC.
Speaker 2:All VR headsets became mobile. Metal Oculus became mobile and I saw that journey when I left Metal to Roblox, and I saw that journey actually becoming more and more apparent on Roblox when you have more computings all over the hand, hands of all the users computing on the prem. So, yeah, starting in 2015, seeing what's possible and just cracking on that idea and landing product after product. And yeah, around a few months back, we said this is becoming the future. It's capabilities on the headsets, on the edge devices. If they're done correctly from ground up, now we can do things that are super amazing. People don't need to always be connected, we can bring privacy back to the hands of the users and we can make sure AI is available on all surfaces. For me, AI is knowledge. Knowledge should be available and everyone's have. So, yes, that's been a journey and we are pushing forward.
Speaker 1:That resonates a lot with me because at Avantest, each architecture underpins our RTDI real-time data infrastructure platform and our new Gemini, which allows our customer to push AI and their models directly to the test floor. So, from your vantage point, what will the next big breakthrough in age arise? Is it hardware, acceleration, orchestration software or smarter models? What do you think?
Speaker 2:It's going to be, in my mind, three layers and those three layers perhaps mature at the different phases First, we should be able to run computing. Second, we should feel safe to run this computing. And third, we should feel easy to use this computing. So today, on the first layer, we should be able to run this computing. We are seeing edge devices.
Speaker 2:The main limitation is memory, memory, memory memory. It's less and less about computing, it's about memory, memory bandwidth, memory footprint and I would say particularly memory bandwidth, memory footprint and I would say particularly memory bandwidth. Innovations around memory is going to be huge. These models, these AI applications, are going to reason deeper, are going to be smarter, are going to remember more. We have a lot of compute units on edge devices. They have small bus, they have small throughput. Sometimes these buses are tinier. They're being scheduled inefficiently. They're being scheduled inefficiently. Major innovation, if I put it, is memory From ground up, from silicon to the operating system, to layers of actually using memory in a streaming format, moving away from traditional usage of memory, are going to be fundamental in enabling AI workflows come to edge.
Speaker 2:Once that's unblocked, I'd say now you're able to do something. Do you feel safe about it? I'd say now you're able to do something, do you feel safe about it? That means these applications should be reliable and should be secure. I ask it to do something. It should consistently do that. I tell it to read this file. I tell it to open up X. It needs to consistently do that, not that sometimes interpret something differently. And it should be secure. No one else should be able to do that. No one should be able to hijack my conversation. No one should be able to even know if I have something personalized on that model.
Speaker 2:I should feel safe and once that's done, we're going to move to always on AI application. I'm not going to pay for it. Don't pay server side. It's private. Why turn it off? My assistant has to be always there. You should always listen to me. If it's my assistant here, if it's my assistant and if it's code assist, if it's a robot, if it's a security application, it's always on.
Speaker 2:And what's that say? We're going to see a lot more applications coming to edge. We're seeing 75% of ISV developers when they want to come to Edge. Actually they fail because the ecosystem of building workflows for Edge being able to monitor, observe them remotely, upgrade them it's missing. We don't have an end-to-end workflow, and that's where we are coming in. We're trying to remove not even the boundaries of technology, but making sure we pay attention to our enterprise partner so they can develop the application simply and they can actually maintain it once it's out there. So those three layers are very critical and there are a lot of technical depth technical innovation that needs to happen, but also very much system thinking how that pipeline should be structured.
Speaker 1:You mentioned about memory, so I want to go into that. So you recently slashed memory footprint. Can you walk us through the breakthrough using a travel packing analogy?
Speaker 2:Let's think about the memory as a multi-layer thing. Or maybe in our how our brain works, there are things that we remember, maybe from years back, something of that. I have to go think about it. Right, that's more or less resembling the five system, the disks on this. On these disks, on these edge devices, there are things that's a context, like conversation you and I are having. I just remembered that, right, it's a short memory and that type of memory plus my awareness, my consciousness. So how do I answer that question? Might be your system map. You can see a child, the system RAM, the context they can remember might be shorter when they're baby and they grow up it might be much longer. And there is something that you learn your awareness gets better. That means the model you're actually using on that consciousness becomes richer, becomes smarter, becomes fine-tuned, becomes trained. But then there is tasking. Right, hey, I'm listening to you and at the same time, maybe taking time. Can I actually do multitasking? Well, my bandwidth is limited. And similar thing is happening on the edge. You have edge when you're dealing about AI inference, especially large models. You have a footprint down on your system RAM. You're operating, it's your model, it's the context it remembers.
Speaker 2:There are things that are off the disk might be longer memory. We should be able to utilize those. And how we connect between system RAM and off disk Disk typically is not an issue. Ram there are limitations how big of the model we can do it. But then there is bandwidth. How are you going to actually run the operations? How many bits can you go back and forth so you can on the CPUs, gpus, npus of the world you can do processing? There is a lot of back and forth and how much you can hold on your processor unit cache, how many cache misses you have, how many operations you need to push back to the RAM. That bandit, that tunnel is very limited and we are seeing there is a lot of bottleneck. How you can expand that bottleneck, how you can fit, utilize that bottleneck smartly and how can you actually bring that bandwidth closer and closer between memory and processor units? And there are architectures being exploited by NVIDIA and others, which has been amazing, and I think that's where the future of Edge will go paying attention to those constraints.
Speaker 1:Interesting, because you also claim that your engine squeeze about 2 to 3x more performance from commodity silicon. So what elegant tricks make that possible in plain English, if that's possible?
Speaker 2:I'd say, when you look at this in the entire system, there's a lot of thinking that can be done. There's a lot of thinking that's for a given operation, what is the right workflow structure should be designed, what model should be talking to another model, to another model, how should they be talking, when should they be talking and what information should be actually sent back and forth during inference. A lot of approaches today may look at these operations very monolithic. I do one X figure out everything, do the computation go to the next one, go to the next one, so on and so forth. I may not pay attention when I'm using this workflow, maybe on my device, maybe I'm on the laptop, I have a browser running and I have camera running and I have XYZ. Maybe actually I need to do the inference slightly different. Maybe I should use the bandwidth slightly different. Maybe for that given task that I've been asked for, I don't need to use the entire band, the entire. Maybe there is a smaller reasoning, the processes that I could be using.
Speaker 2:So innovation is not just how a model runs.
Speaker 2:So innovation is not just how a model runs, how the entire workflow as a system runs, when you have an operating system that runs on a device with limited bandwidth, limited RAM, handling multiple other applications beside what you're doing and, more importantly, paying attention to what at that moment, that user is asking for. Those combining them together and being able to adapt in together and being able to adapt in the operating system requirement brings so much more potential that was previously not exploited. So we need to think about AI as a system. Ai is not as in some basic operation. It's not like just a single language model. It's a complex workflow running on a complex operating system, running on a very constrained compute units, and there's a lot of smart that needs to happen there. Compute units and there's a lot of smart that needs to happen there in terms of how you architect end-to-end and benefits, can be significant and it can be huge. It may be two, it may be larger. Really really depends on the operations and that's what we're thriving at.
Speaker 1:One thing also each device spends ARM, risc-v and custom ASICs. So how does OpenInfero keep the promise of write once and run anywhere?
Speaker 2:I like the Android analogy. There are layers, there are abstraction layers that think of it as HAL layer. Those have to be written quite well for each processor units. But there's another layer that acts as an operating system Memory planning, memory management, memory layout, cache alignment. Those may not be very specific to a given hardware being able to architect that engine well enough that you can go deep, deep, optimized for a given instruction sets and instruction sets and memory are memory available, the cache architectures, but able to, but also be able to abstract it out so one orchestration can act instead of available, let's say, AI APIs enables us to adopt more and more processor units, more and more even custom processor units that are coming in, but have the main brain and orchestration and management sitting on top. When that's more agnostic of hardware operation, that more adaptive to the workflow architecture, I'm not saying it's a challenge.
Speaker 2:We started particularly from CPUs. Cpus are well standardized. A lot of our opportunities, a lot of instruction sets are well documented, so we started with those, especially CPUs. On shared memory architectures when you get to shared memory architecture and your memory bandwidth is bound primarily, then the choice of what processor units to use is less important and rather how you actually make sure that processor units uses the bandwidth efficiently so you can get the right throughputs at the right thermal, at the right power Once we get a hold of processor units from CPU perspective. Now the question is what other processor units you bring on the table? Are those of higher priority? Now the question is what other processor units you bring on the table Are those of higher priority? Or maybe there are other things we should bring on the priority? So architecture of the software is important, but also making sure that we don't necessarily say, hey, every single hardware unit on a given shared memory architecture, APUs or SoCs, is important, Rather paying attention to what technologies, what system technology is important to adopt to change the envelope of computing.
Speaker 1:It may not always be adding new processor unit, it may mean something different not always be adding new processor units, it may mean something different as you push the capabilities of each machine, and now you're making it more powerful and making it easier for enterprise to adopt this. So, in your experience, what single biggest barrier keeps enterprise from scaling to edge today?
Speaker 2:Complexity of development. You're dealing with heterogeneous compute units CPUs, gpus, igpus, npus that each vendor have a separate set of APIs and there are some attempts being done to unify those, but those applications are very difficult to use for developers, to use ISV developers very hard to manage, very hard to maintain. So it's a complexity, complexity of development, complexity of utilizing this. The tooling is missing. The existing compute units, the existing, the existing compute units, the existing bandwidth architecture, memory architecture can do a lot Of course. We have paid a lot of attention in that space, but we still see lack of adoption.
Speaker 2:It is just cumbersome to come to Edge and we like to be we like to be the company that's really on block set. It's a pain in there. It's a pain every enterprise have. We've gone through that pain many, many times. It's very expensive in terms of man hour, engineering hour that you need to put. This needs to get solved. It totally reminds me of the time that you know, each mobile headset had a different operating system and you just couldn't have apps, flip-flop, save one app you see on here. Each telpo had their own thing and each silicon maker, each phone maker has its own operating system. It just wasn't working and we are in that dilemma today for AI application and we need an inference operating system to solve that. Write abstraction there. Write AI inference management with the tooling. Think of Android Studio built on top of it. This will unblock it. The adoption is hard.
Speaker 1:You had a lot of customers already, and which industry surprised you by adopting on-device AI first and doing the always-on AI that you talk about? And what did it teach you?
Speaker 2:Financial and manufacturing was by far the one that taught me a lot and taught me a lot. You're dealing with humongous data sets that are being generated continuously. Whether you're on manufacturing sites or you're in the financial firms, a lot of data are coming in. There is an existing infrastructure that is working well. Systems are working. There is hesitations to change. There's hesitation to yet introduce another uncramped hardware or move my data somewhere else. But they're thirsty to use the latest of AI capability bringing reasoning and assistance. They're thirsty because the knowledge is growing faster. They're thirsty because they're spending a lot of money on analysis, predictability in manufacturing, seeing something correcting so there's a lot of pain. They're thirsty. The bottleneck is huge.
Speaker 2:Any AI says either move everything to the cloud, adopt XYZ, depend on my infrastructure, assume it's always reliable. Not all manufacturers are going to accept that. Not all retailers are going to accept that. Not all retailers are going to accept that. And they love to bring a system. They love to bring deeper reasoning. There's so much data they have being able to deploy those simply on their edge devices.
Speaker 2:No change the same system that I have, the same system that I'm using every day now, can do a lot more Because it is capable and that taught me a lot. That taught me a lot that there are players out there that are thirsty. They're not going to move their data somewhere else, they don't want to yet adopt another set of GPUs and all that, and they love a smooth change, a smooth transition and something that if they didn't like it, they can just flip it and turn it off. I just didn't like this system. Done no pain. We should come in with that angle and that was the biggest eye-opening that. How much demand exists in that space Retail, point-of-sales, robotics, code gen, all of those we anticipated, but we didn't anticipate so much on the manufacturing and financial side.
Speaker 1:That resonates a lot with us. That's why we're doing the age inferrings with ACS, rtdi and Gemini. A lot of data generated at our manufacturing floor and you don't want to share those data on the cloud and security and all this. So I like to move to a different angle. So among your pilot customers, which KPIs keeps them up at night and how does H move that needle for them?
Speaker 2:A lot of partners care about performance Tasks done per second, tokens done per second. Power is the secondary that is becoming an important topic. How much power is it going to cost me? Is it going to drain my battery? Can it run on this tiny device that sits out there? If it's on drone, can it just actually run for this many operations? If it's my handheld, can it just run all the time but just use tiny amounts of battery? Those two have been the primary initial asks and the observations I have is that everyone talks about compute, everyone talks about performance, but they may not have it connected well with the outcome that they're asking for.
Speaker 2:It may not always be how fast you respond.
Speaker 2:Sometimes it's about how you respond. Imagine in a conversation, in a voice conversation, right, when I'm saying something and you're like, oh, aha, aha, I'm not gonna pause, right, I'm seeing that you're confirming, so that means Ben, I'm continue. But if I was talking faster and faster and faster, and then you say aha, and then I stop and then I respond, overall experience is worse than if someone is talking. But then also understand what's the outcome, how the flow should actually be, how the interaction of these models should look like, when one should interrupt the other ones. Educating our partners to think about that aspect that may resonate in terms of the task completion rather than inference per second is something that we're putting a lot of effort on and requires a lot of product thinking and product education. So I'd say there are certain KPIs that customers care about and there are certain KPIs that we need to actually create terminology and education as these AI applications become true assistants in our daily life, and just linking this back to what you mentioned earlier, this is about the whole task flow, right?
Speaker 1:The whole flow of everything, the end-to-end.
Speaker 2:Exactly what you see your camera, to what it does, to what it tells to do next. That task is important, not that how this fastest one does an operation.
Speaker 1:That makes a lot of sense. Openinfo already powers some game-changing deployments. Could you share a success story that crystallized the business value of Enterprise Edge?
Speaker 2:It's good to think about what's actually worked already well and we are so used to it. Think about the assistant in the cars. Think about the assistant in the cars. Almost most of the cars today they have a type of assistant. They have lane assistant. You back up, it gives you an alarm and you may go as far as having a full autonomy in the car, driving right, it's a full workflow. It's a full workflow that is embedded in your life and you're trusting your life almost every day. You back out, you're expecting that it beeps if it's in something. Sometimes the car actually pauses, so it has the right to calling. It can control the lane.
Speaker 2:You unlock your phones today with base IDs right, you have iPhone or you have Android. You have the Android version of it. It already does a lot of reasoning on device, because the latency being always connected and sometimes data privacy, just can't allow you to stream back and forth, send your data, to unlock your phone or to make sure you're staying in line when you're driving. These applications already exist in our life. These applications are already changing our life and we should be seeing more of this. We should be building the full orchestration, the full operating system, the development tool so more enterprise and consumer end applications come to edge.
Speaker 2:Remove pain point of adoption. Face ID amazing application Took a trillion dollar company to build it. This should change. More and more people should take advantage of all the compute units you have in your hands, and then silicon makers are more and more excited to build the compute reference units. That is addressing not the daily tasks that traditionally we were thinking this device you should be doing rather daily tasks that AI workflow is asking us to do. Change the way memory bandits are operating. Change the way we are actually building the full systems so that influence will trickle down, and will trickle down all the way to silicon designs and we're hoping that we can be an influencer on that whole subject, from all the ISV developer ecosystem to influencing the silicon.
Speaker 1:I was talking about what you mentioned just now, about change the way how we do things, change the way how we think what outcome make even your own team say, wow, we didn't see that coming.
Speaker 2:I've spent a lot of a good portion of my career on the system architecture, particularly memory side, and the whole team is an amazing team that they put each maybe each of them 20 odd years on this space. But still, the nuances of this space, of being able to utilize the whole memory ecosystem. Well, think about everything should be streaming, everything should be flowing, everything should be models should be streaming, compute should be streaming back and smoothly, operating system should be able to orchestrate that. And every day we are baffled by how much we are learning on that space, even though we thought we have a good grasp of it. It, and how much opportunity is opening up. And it's been an amazing ride.
Speaker 2:Sometimes you think you know well exactly what's ahead and you're still surprised. Positively, it's like, wow, I did not think I could do x, y, now like imagine if I could influence in it this way, um, so yeah, those that that would be, I think, one, one area that baffled, um, baffled us technology, technically, I'd say, I'd say technically and even educationally, trying to actually set that as the framing as we talk to the partners, have been an interesting aspect that by far overrules the other angles.
Speaker 1:Now let's untravel. Paint us a bold picture. How will enterprise edge computing transform by 2030?
Speaker 2:We've seen always computing to be in pendulum. Sometimes we go heavy on the cloud, sometimes we go heavy on the edge. We've seen that in the past for non-AI application. We shouldn't expect that everything is going to come to edge. We shouldn't expect everything is going to be on the cloud, but we're going to see inference is going to play a major role in future of AI.
Speaker 2:Now we have a lot more trained models. People are becoming creative in building agents. Usage of those agents are limited, whether it's cost, data privacy or maybe reliability. It's just not always available for you. The 2030, in a matter of like the next five years we're going to see new innovations in silicons, new innovations in the hardware and a lot of creativity and seeing assistant living closer and closer, always on in your daily life, whether it's for your enterprise application or your personal life.
Speaker 2:I believe you're going to see it everywhere. We may even see it on our glasses constantly as a companion to our watches and everywhere. Or we're going to see in our manufacturing sites Part of our integration, part of our integration, part of our creation. We're going to see more of that and we're going to see more verticals, from ground up silicon to all the way, innovative applications being built. Robots are going to be everywhere. Those robots are I'm looking at them as hardened assistants. They may be doing one thing in manufacturing, at home, at retail. There's going to be everywhere. It's going to be everywhere and edge is going to be very huge. Inference is going to be a huge market, a very big market Doing those robustly, and security is critical.
Speaker 1:I believe that. So, for a team that are just beginning their edge journey, what first small win should they chase? What's your recommendation?
Speaker 2:Really think about what the task is asking for. Not always the do it on cloud or do it on edge or do it hyper efficiently. That may not be the right thing, identifying what's the out and what's the experience that assistant, that AI, that workflow is going to bring to you. And walk backward and figure out where what operation can run, where what operation can run. A lot of times you may realize this operation can run closer to you on edge. Don't force yourself. Be on edge, but start with that workflow. Have a very clear experience in mind and then walk backward.
Speaker 2:And if things can be done on devices, there's significant advantages to those on edge devices. But then sometimes it's just might be easier, faster, to prototype it on the cloud. Don't underestimate that. Look at that as a surface of computing. Decide when, in what phase, it makes sense to use what. But those all should be your toolbox. Use them fluently and keep checking us out. We love to hear from you. We are building the full ground up. We will be open sourcing. We will be definitely helping the open source community and we love to hear which aspects, how we should open source. We want to make sure, once we open source, we have the right community behind it. So I really invite the folks to reach out to us. We're looking forward to it.
Speaker 1:I'm sure a lot of people will be interested. I know for sure we are. I'd like to do a very quick lightning round, a quick fire, just to close us out. So answer in one phrase or two what is the most overrated password in tech right now in your mind?
Speaker 2:I think everyone talks about it's AI. Okay, you hear it all the time and it means a lot. It means many different things.
Speaker 1:Dr Baseni, thank you for the insights and stories. It's clear that edge isn't just where data lives, it's where action happens To our listeners. Thanks for joining Avantest Talk Semi. If you found value today, subscribe and leave a review. They'll help us bring you more forward-looking conversations. Until next time, keep pushing the boundaries at the edge.
