NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hello and welcome, everyone, and thank you all for joining us. We have an exciting panel lined up for you here today and a great set of panelists with whom I am [ amicably ] looking forward to have an interesting conversation. So the way we are going to do this panel is that I'm going to give you an overview of what the panel is going to be charged with, and then I'm going to hand it over to the panelists for their vision talks. Following that, we're going to have an exchange of Q&A. So let me share my screen and let's get started so that at least you get to see who the panelists are today. So I'm very pleased to have and really honored to have a number of experts from industry, from nonprofits, from the federal government and also from academia. And really I think this is a collaborative environment that is needed to solve this very pressing and tricky problem of will 6G be ready for native AI and machine learning. And this is clearly a hot topic today. I mean there's hardly a conference where you don't see tens or perhaps even dozens or hundreds of papers on just this topic, and yet we are still far away. So the goal of this panel is to really demystify the gaps, identify barriers and collaboratively figure out what do we need to get there. So just to frame this panel, I just have a few slides that I'm going to point out some key questions and key pointers that the panel is being charged to really address and answer. So the first one is, we do speak a lot about machine learning in 5G, and there's already a considerable amount of effort that has begun in incorporating machine learning into 5G. Now are we ready for 5G before we even go to 6G? Are we there yet already for 5G so that we can take the next big jump? And when we go to that next big jump for 6G, what would be the extreme performance metrics and what sort of applications would emerge for which there is no way out? We have to do -- we have to get a robust machine learning pipeline into the architecture design. So what would be those metrics and problems? And finally, is there a grand vision, a grand -- a moon-landing scenario that the panel could perhaps charge the audience with and say, "Can you get this? Can you land this flag on the moon?" And if you can identify a couple of these use cases, that would be wonderful. Now when we talk about 6G standard and what does it mean to be ML ready, there's clearly questions on where should the computation happen. There are questions on -- let's say, you have intelligence in the devices. And how do you even certify those devices? How -- are these devices are going to learn and adapt? And how will you benchmark them against the common frame of reference? There are other challenges of privacy and security concerns that will be brought up multiple times by the panel. And along with that, as wireless engineers, we love deterministic performance. We love guarantees. We love safeguards. How much of it is feasible? And in the absence of it, can we still get a meaningful utility out of the network, even if it can't give you absolute guarantees? So when we talk about machine learning, and we can't divorce the network from the machine learning itself. So at one hand, machine learning is aiding the network. On the other hand, the network must provide resources for the machine learning to operate. And so the panel is going to bring this up. And this is a really catchy title, is it ML for networks or networks for ML? And the panel will have some very interesting insights on what is needed and how do we bridge that gap, how can the network support complex and really meaningful real-time machine learning. And a couple of thoughts from just our recent experiences. So I'm glad to be a participant of the NSF-funded AI-Edge Institute that was just announced. It's led by Ohio State. And so we did a survey among the PIs who are participating in that institute. And the question was that, what is it that is stopping you from doing great machine learning in context of data sets and availability? And so the 2 pointers that stand out here are: A, that there's a consumer learning curve towards using test bed platforms and facilities from which data sets could be acquired. And the second one is even if you have data sets, it is lacking documentation, it is lacking ways to interpret it meaningfully. So there's clearly a journey that we need to traverse here. And to get there in, say, 5-plus years, we need tools. We need incentives to generate good data sets and to basically even understand what is good data. So the panelists here are really going to talk about data sets and the challenges and the availability and what do we need to do to bridge that gap. And finally, I want to bring this up that we are taking concrete steps towards addressing some of these problems at the grassroots level. So personally, I'm the PI of a recently awarded community infrastructure project from NSF on data set generation sharing and maintenance tools. So our goal here is really to design tools from which users who are not, say, greatly experienced in the nitty-gritties of a test bed operation should still be able to go fetch relevant data. And there should be preprocessing tools that are able to clean and filter out data so that they can make it adaptive for their own research. And this is only possible if you have the support of, of course, the federal government, and we're very grateful to the National Science Foundation, but also with the support of other companies. And we are, in fact, working with NVIDIA, National Instruments, et cetera, to work together to create this data set generation sharing tools. And I think this could be a way to democratize data sets across the board. So enough of me talking. You have not come here to hear me talk. You have come here for the panelists, and I just want to stop my share here. And while [ Alex ] brings up the slides, I would love to introduce to you our first speaker for today. We are very honored to have Nageen Himayat, who's Principal Engineer from Intel. And she really conducts research on next-generation 5G and beyond systems for mobile broadband architectures. She has a number of different publications and greatly involved in the design and contributing in standardization efforts, et cetera. So Nageen, over to you.

Thank you. Can people hear me well? Just checking.

Yes.

Okay. So I guess the discussion point will be, will 6G be ready for native AI and machine learning? But that's the discussion we'll have. But just to kick off the discussion, I would go to the third slide, Slide #3. So we do believe that even though maybe it's not native, but since AI and machine learning will be integral to 6G networks, and there are 3 different aspects that I wanted to highlight, which people are familiar with because Kaushik touched on them. So one is AI/ML for designing of 6G network all the way from physical layer to end-to-end communications. The second is that 6G networks, the devices that comprise these networks, would be able to support AI computations more pervasively as we go further. And then third, as more and more 6G-critical applications such as autonomous services get deployed on 6G networks, then AI would have an important role to -- towards codesigning some of these networks with these autonomous services to sort of better meet the quality of service -- of these services. So next slide. So I'll start with the first little bit of background on. So as we see that the scope of AI and machine learning applications and design of 6G systems continue to be getting broader and broader. I mean, there are more application areas that we get -- that are identified. And the publications continue to show promising results in here. So the first one is basically, this is work done by my colleague at Intel, where they've applied AI/machine learning for channel estimation by just kind of looking at one symbol and using the pilots to develop channel estimates. And the performance is pretty good and robust with low-complexity neural networks. The second, I think our next panelist would talk about it, which is basically they are proposing AI to be sort of native to the air interface and -- the air interface, which is based on AI. So I think I'll let Jakob talk about that. Also, radio resource management has been a big problem in -- it continues to be more complex as we go along because of the variety of resources that you need to manage, including massive MIMO resources, interference and so forth, control and so forth. And here also, some of the works done by my colleagues show that multi-agent reinforcement learning techniques actually are very promising. And they show results which says that we can have a better trade-off in terms of summary throughput versus cell edge throughput even when using decentralized approaches and come close to very centralized approaches. And then finally, I think most communication systems currently are defined by reliability transmitting bids. And now the discussion is, can we reliably convey the semantics associated with these bids? And there is some recent work which actually show promise in that regard. So it's just like a growing tree. Next slide, please. And also, we note that there is a lot of momentum for supporting AI compute over wireless networks. And people are familiar with predictive learning, which allows end devices which collect the data to collaboratively learn AI models while keeping the data private. So there's a lot of appeal in that particular approach. But to scale this type of method, we really need to address a lot of challenges, which have also been identified. And some of the work that we are doing are trying to address some of these challenges of how do we efficiently use AI resources, which not only include compute and communication resources but also data resources through better sampling techniques. We are also looking at coding techniques to improve redundant computations to improve the reliability of computations because the wireless edge environment tends to be very dynamic and unreliable. Other work by our partners are sort of focusing on what are the computational architectures that are well suited for the resource constraints of the devices that occur on wireless networks. So you want to learn smaller modules on end devices and larger modules on more capable devices. And then you need to find a way to compose them together such that you have a more powerful module overall. Finally, there is room to optimize wireless networks specifically for AI workloads. And going further, can you use wireless medium itself for doing computations and over-the-air combining and so forth? And there are some interesting results reported in this paper. So the next slide. Finally, I will not go into too much detail. But as we see that when we introduce very stringent safety critical services over unreliable wireless networks, the requirements of reliability, latency and so forth are just go through -- become even more stringent. So you need time reliability and so forth. And some of the work that we've done with University of Pennsylvania sort of challenges that. So the idea is if you can codesign -- you can use AI to codesign your network resource allocation, which is more aware of the control states of what you're trying to control, then you can really relax the reliability requirement. And I think that's a very sort of important area to focus on because it's very difficult to meet some of these requirements otherwise. So finally, next slide. So there are lots of -- I think there's one more slide that you can back up. One more slide after this. Yes, thank you. So I mean, all this work basically show a lot of promise for using AI and 6G system. There are lots of opportunities. Of course, work needs to be done, as is pointed out by Kaushik as well. There needs to be standards which allow some of the information to flow, and that work has already been started in O-RAN. I mentioned -- data was mentioned as one of the main challenges to improve your AI models, and some of the work on predictive learning or privacy-preserving predictive learning can help in that regard. AI can be used for generating synthetic data sets. So there is an important role for AI to play there. And then the codesign thing that I've just talked about to relax the requirements on 6G networks. Of course, there are lots of challenges, which I think the panel will talk about. Robustness and security continues to be a challenge, especially for mission-critical networks; the role of self-learning approaches and self-validating approaches so that you can minimize the human-in-the-loop in the system. And then I think people realized that just analyzing the performance of a single AI loop is hard enough. So if you have multiple of these working together, we really don't have a handle on that. So that would be a big challenge to address. And we are partnering with NSF, as Kaushik also mentioned, on some of these programs such as Resilient and Intelligent Networking -- Next-Generation Systems, the ML GANs program. And we expect to address some of these challenges going forward. So with that, I hand it over back to Kaushik.

Thank you, Nageen. And while this next speaker gets ready, I just want to pose -- throw one question at you. It comes up in the chat here. So the question is, "You mentioned codesign. I'd love if you can reinforce that concept a little bit because, A, either you do a real codesign or you just use AI/ML to enhance the operation of a given network, so I give you the network, versus a codesign network. So can you shed a little bit of light on that?"

So yes, I think the work that was shown here, which we didn't get a chance to actually go in detail, is while I'm actually allocating the resources, I don't need to allocate resources to control a loop which is already very stable, a control system which is already very stable. I can relax those requirements and steer away my resources to some other. And likewise, if my control system is also aware of the network conditions, then I can probably take some mitigation actions to do this. And if you do that, then you can show that I don't need like 99% faster delivery ratio. I can probably get away with 70% fast delivery ratio. The role of AI there is interesting because you can learn the control behavior of the system you're trying to control, and you don't need to know it a priori. So that's how you can actually generalize this approach to many, many systems. I think that's the idea there.

Got it. Got it. Thank you very much, Nageen. Thank you for your insightful comment. We'll move on to our next panelist for today, Jakob Hoydis. So Jakob is the Principal Research Scientist at NVIDIA, and he is working on the intersection of machine learning and wireless communications. And prior to that, he was the head of a research department at Nokia Bell Labs, France. And so Jakob, over to you, and feel free to share your slides.

Hello, can you hear me?

Yes. Jakob, yes.

Okay. Fantastic. And you can also see my slides, I hope?

Yes, very well.

Okay. Perfect. Yes. Hi, everybody. Yes, I'm really pleased to be on this panel. It's a topic that perfectly fits my research activities and interests. And yes, to get started, I would present here a very simple 6G vision that tries to make an analogy between what's currently happening in the world of processor development and what might happen when we transition from 5G to 6G. So thanks to Moore's Law, we have seen increasing processor performance and efficiency over the last decades. Now Moore's Law has essentially come to an end, and domain-specific accelerators that accelerate a specific type of computation are one way to continue scaling performance and efficiency even further. Now wireless communication systems are traditionally designed to work robustly in quite a wide range of scenarios. And a bit like a general-purpose processor, you can see them as a general-purpose wireless bit pipe. And so now we believe that like a domain-specific hardware accelerator specializes to a certain task, 6G systems should be able to specialize to the radio environment in which they are deployed and to the application they serve. And in order to enable this specialization at scale, this must be fully automated, which can only be achieved through software-defined radio access networks and machine learning. And also, digital twins or siblings are another enabling technology to achieve this. Now let's focus just on the physical layer and think about from a research point of view, is it actually possible to have a communication systems that can actually adapt to any type of environment? So I would claim that based on the work that many groups in the world, including myself and my colleagues, have conducted over the last year, that it's possible to learn the entire physical layer. So it means we can learn new codes, especially for black lines, that achieve state-of-the-art performance. We can learn new modulation schemes that allow us to get a bit of demodulation reference sequels. We can learn new waveforms that embrace hardware imperfections while actually still being very spectrum efficient. And the key to all of these are fully neural network-based receivers that largely outperform traditional methods in most realistic scenarios. But -- by the way, a few references here to each of these points. And then the next question many people would ask themselves is, "Why should we actually learn a new physical layer?" And well, one could argue that OFDM with QAM is good enough. And if you go to terahertz where you could see -- you could just them as a simple single-carrier scheme, and I would say that this is probably a valid point for the traditional cellular broadband type of networks. But the more we go towards highly specialized networks for distributed sensing or intelligent applications, robot control, joint communication and sensing, our power transfer also becomes less clear. Now beyond that, I think there was a couple of good reasons. So first of all, a physical layer that is learned can achieve unprecedented performance, especially once you take hardware constraints into account. And this will become a dominant factor in the terahertz spend. And I've also witnessed over the course of the last few years that it's actually much faster to train a neural network than to implement the baseline algorithms that are used for benchmarking. And because you can go more or less directly from data to a hardware implementation, this traditional algorithm development and deployment cycle can be dramatically dropped. And lastly, I think that if we expect that 6G will support an even more diverse set of use cases than 5G, it will become very difficult to standardize for all of these cases. And so ideally, you could instead standardize a learning-based method that would allow the system itself to automatically adapt to a specific environment. Now what do we need actually to succeed and make this happen? Quite a few things. First of all, I would say that having a learned transceiver is something nobody has ever implemented in a product in the world. And to make this happen, it requires, I think, a paradigm change in the industry to enable it. So if you look, for example, how long it has taken for MIMO to make it from research into a protocol standard, it seems a bit challenging to be ready within the 6G time line. Now another big question is what we actually standardize. So would it be just the learning outcome like a trained model? Or would it be the signaling and procedures that enables it? And lastly, I think that trust and reliability are very important issues. Because from an economic perspective, it is actually impossible to develop simultaneously a machine learning base and a traditional solution to the same problem. And before any company goes ML first, they need to trust their solutions and must be able to test and troubleshoot them efficiently. And yes, that's it for me, and I hope that we can discuss some of these points in the panel.

Wonderful. Thanks, Jakob. So one point stands out, as you were talking, I was making notes here. You said there is a possibility now to take -- to jump directly from data to a hardware deployment stage. Now does that mean that, that data that you are using for this sort of a jump has to be comprehensive enough that it must be representative of every practical use case that can exist? Because once you go to the hardware stage, you are somewhat less flexible, let's say, in your modification. So how do you make sure that jump is sound?

Yes. I think it's a very, very, very good question. So I -- so we have done quite a few tests and trials over the last few years. And it turns out that in -- I would say that in many cases, it might be good enough to have training on a very rich data set in an offline mode so that, as you said, it's representative of many cases you would expect. Similar as you test your algorithms on a wide variety of 3D models, you could just test it on a larger data set and then be confident that it will most likely work on the large cases. Now -- but I think if you would really like to get the last, I don't know, 5% or 10% in performance, I think there is no way around online training in the field. So this gives me -- that's what I meant actually with a network that -- a physical layer that can specialize fully to a scenario. So if you think about a receiver in some factory setting, you only want it to work in this particular setting. So it doesn't need to work in any other things. And if you now have to learn a new solution that can essentially overfit to this environment it feeds, that brings you, I think, the real performance gain. And both are required. So I think there's a need for online, I would say, deployment side-specific training.

I think that what you just said also answers one of the questions that came up that can a layer one adapt to rapidly changing channel conditions. So I think this basically answers that you need to have an online component.

Yes. Now the thing is always, if you think about this learning the physical layer, most likely, you don't want to change what is transmitted, so essentially the waveform. That's not something you should possibly learn all the time because the overhead base is too large. But having this trend in an offline mode on a very realistic model and then just have the receiver adapt to whatever is the mismatch between the model and the real world, I think this compensates quite sufficiently for it. But training end-to-end all the time, I don't think that the scale makes sense from a feedback perspective.

So you can hold that thought. We're going to come back. We're going to come back soon on this. But -- so we're going to go on to our next panelist. And our next analyst is Dr. Nada Golmie, who received her PhD in Computer Science from Maryland at College Park. And now she's at NIST, where she's the Chief of the Wireless Networks Division in the Communications Technology Laboratory. So she has many, many papers and actively leads a lot of standardization efforts involving simulation modeling and so on. So Nada, over to you.

Thank you, Kaushik. I am delighted to be here to participate in this panel and a few thoughts to sort of set the stage for the conversation to come. Next slide, please. So basically, the vision, people are talking about 6G, next G, whatever G basically, where we're going to continue to see -- to handle huge amounts of data at very high speed and very low latencies. As a result, a lot of things are happening. They're going to be transformative, and we're witnessing that from many different sectors, from smart cities to telehealth to autonomous driving, to agriculture, robotics and the like, right? So we're going to continue to see that. And really, this transformation is being fueled by the 5G-AI combo, I would say. And in fact, some people are predicting, they've already put some numbers, dollar amounts on this combination, despite that this is coming from the government that we do care about the economy. So there is basically an expectation that by 2035, the revenue -- direct revenue from 5G and AI combination is in trillions of dollars, right? So that's great. And that's, I think -- so that's happening already. If you look at the impact of AI in current networks deployed, that is already -- we're seeing that AI has been used for some time already for the -- with the service providers and the telecom operators. And this will continue. Next slide, please. So the one thing that people that have been using AI in the deployed networks today, whether it's 4G, 5G, whatever G you want, is basically you get real benefits if the data quality is there, right? So data quality is important. For people who don't have data, I mean, that's mostly been the case in the research community, I think there's more good news and more data that is coming your way. And we'll get to that in a bit. So it's important to keep in mind that, as also Kaushik introduced, it's important to -- not all data is created equal and that data quality is important. That's what people that are using AI today in networks are reporting. Next slide, please. So continuing with this AI -- using AI in 5G, in next G, and this is going to continue, right? So as Kaushik introduced, AI for 5G or 5G for AI, this is transformational. And as the network of the future, it's going to support so many different scenarios and applications. There has to be a native support for AI, right? Because we're using AI in so many different sectors. We're using it in communication and so many different applications, whether it's for connectivity, adaptability, high-capacity spectrum. We're going to see more of this native. And Jakob provided a very nice case for the learned ran -- learned file, right? So -- and we're going to see more of that. Next slide, please. So what happens next is that we are going to have these enhanced communication intelligence, these smart communications, fueled by AI with next G, 6G. And then the components there, obviously, are measurements, models because we need these. And we also need new metrology, new methods for evaluating this combined communication stack, right? We're no longer dealing with something that is static or devaluing its performance. And then we're dealing with something that is -- can be learned. So how do we assess it's performing as it is, as Jakob listed a number of challenges. So we need new metrology, new ways to measure in a SaaS, new models to utilize also machine learning and also to develop communications data sets. At NIST, we have a number of projects that tackle different pieces of this puzzle. Obviously, we have very long history with propagation models and measurements. We have data sets that we provide to the community for this type of application in machine learning. Next slide, please. So the story here is sort of -- is one of partnership and collaboration. I think many of us are already involved with the RINGS project that was mentioned. We have also consortiums that basically provide data in terms of measurements and models with the next G channel model alliance. I'm delighted to see Kaushik's announcement about the RF data factory, and I'll be very happy to work with that effort in order to really look into more dissemination of data sets for these types of efforts. And I think I will summarize. Basically, just the key takeaways is that AI and ML is being used today in various forms of network design, configuration and operation. Good data quality is absolutely needed. Next G is going to do more of that, if not less. And also, we're looking at a more native approach for supporting AI for the different scenarios and applications. And then measurement methods and data are key. And we have as -- what it takes as a community to come together to do this. So thank you very much and looking forward to the conversation.

Thank you, Nada. So one question comes up here. You mentioned that there is metrology. There is the science of data sets. You mentioned technology themselves, massive MIMO. You mentioned telehealth and so on. Each of these are domains in themselves. Now when you start to bridge across these different domains, a question that comes to my mind is, do we have a trained workforce to really tackle them together jointly at a go? And I only bring this up because, I mean, from this, you well appreciate that the science of measurements needs very careful design of experiments, very careful understanding of even the act of how the measurement should be done. So how do we train then this next workforce to be proficient in these wide varieties that you mentioned and -- so that they are -- they become productive members of this wireless community and really push this collective moving forward?

Kaushik, this is an excellent point because, again, we need to train the workforce into the use of machine learning. We need to sort of train the wireless engineers into data science. I'm told that part is easier than training data science into wireless engineers. I don't know why. I think everything is possible if one really wants to learn. But yes -- and this has already been happening, right? So obviously, in every organization that is serious about solving these problems, this has already -- the train has left the station, so to speak, in terms of looking at this. And that should start perhaps with the academic institutions. So what are you doing in your -- in the universities to start that so that we're not really looking at this joint competencies? And -- but 10 years ago, this was very difficult, right, because people were talking about apples and oranges and not really understanding each other. Today, things are -- have moved quite a bit, but we still have a long way to go.

Thank you, Nada. You are a tough panelist. You've thrown the question back at me as the academic hat-wearing member. But yes, we will work with you on this. But let's move on to our next panelist for today. And we are very happy and honored to have Bill Wright, who is the Head of AI/ML and Intelligent Edge, Global Verticals at Red Hat. And really, he is designing technical strategy across vertical industries and global accounts at Red Hat. He's also a member of the Red Hat AI ILT and Co-Founder of the Enterprise Neurosystem open-source AI research community. There are many more accolades to Bill, but I'm going to let Bill talk about his work himself. So Bill, over to you. Thank you.

Well, thank you very much. I'm honored and honestly humbled to be here. I think the other speakers have actually taught me quite a bit today, and it's been fascinating to listen. It's funny to see Red Hat in this conversation for many because they look at us and say, infrastructure provider. And they don't realize the kind of depth of work in AI/ML and also in 5G and ultimately 6G that we're up to at the moment. But one of the interesting things that happened for me, I think, about 5 or 6 years ago was I went to visit a friend at América Móvil, who was responsible for global infrastructure for a good portion of their network there. And we were having a conversation over lunch at this beautiful restaurant called Loma Linda in Mexico City. And we were going back and forth. And he said, "It's really interesting. Red Hat is really well positioned to kind of unify and bring together all the disparate elements that are taking place right now in AI and ML. And I don't know if you've really thought about that through the lens of mobile communications." And I said, "Frankly, no, I haven't thought about that yet." And we began to have a dialogue. And I've always found that the customer dialogues like that are the most productive in terms of understanding and getting a gauge on, I think, where the industry is headed. And so most of my career, I tried to base our activities on what customers truly need and desire and not to lead from within, so to speak. And in this effort, what we've done is we've come up with something that's really an open-source community called the Enterprise Neurosystem. You can go to the next slide, please. You get a sense of where Red Hat's business in AI and ML has gone, and I don't mean to make this Red Hat-specific, but it just kind of puts it in context. Again, many people haven't thought of us working in these areas. But again, in almost every element of our business, and I won't go into each one, we have different aspects of AI and ML that have basically infused our entire product line. And next slide. What we arrived on was really kind of a meta-level approach in view of where this was all headed. And we stood back and said to ourselves, "Well, look, from an infrastructure vantage point, where do we see the core problems arising?" It wasn't in the radio, 5G, 6G, basically the transmission rates, et cetera. What we saw was a preponderance of AI and ML applications both at the edge and at the core, and the large majority of them were completely lacking integration. Many of them were point solutions. Many of them were really at the application layer communicating with one another. But there was no core, I guess you can say, meaning or context being derived from the totality of AI applications in the enterprise. And we took a look around and we said, "Well, look, you've got AI applications coming from vendors that are DIY and built internally. You've got them out at the edge. You've got them at the core. How do we basically provide meaning and context from that?" So we decided to go ahead and start a community, open-source community with a variety of partners called the Enterprise Neurosystem. And the concept is to really create a singularity, to come up with a core AI/ML federation of intelligence that basically connects all the different AI and ML instances in an enterprise, whether it be a mobile network, whether it be a CPG, consumer packaged goods company, a bank or financial institution and basically take all the data and cross-correlate it in a meaningful way and present it up to management. That's really the idea. And so really, what we're looking to do is, especially with 6G and the advances that are taking place there, we can now begin to see the emerging of the edge and the core, the merging of devices basically, and create that single intelligence, utilizing all resources in the network to basically provide that real-time view of operations in the enterprise at any given time. So we're starting with the telco and scientific communities. But again, this can expand to almost any enterprise scenario in a variety of different Fortune 500 companies. And next slide, please. So this is where we're starting. This is the Cookie Box. This is a very interesting device that was basically used by Stanford and their linear accelerator for X-ray -- basically X-ray laser research. And it's in their LCLS lab. And the next slide, please. One of the interesting things about the Cookie Box detector is that it has a frame rate or, I guess, a data capture rate of 120 frames a second at the moment. But that's about to expand to 1 million frames a second over time and over a terabyte of data coming in a second through this device to basically capture the different events at a molecular level and understand what is taking place on -- in terms of like the new physics discoveries that need to be advanced. And so the conundrum here, even though it seems to be only a science project, this really -- the same paradigm extends to mobile networks, that extends to the enterprise. And if you think about the incredible data rates at the edge being taken back to the core and then processed there, there's an interesting load balance effect you have to look at in terms of the kinds of algorithms that will be posted at the edge and what will be basically taking place at a neural net level at the core and what programs are really optimal for that. And so we're looking at all these different areas. Next slide, please. And so really, what we're looking to do from a new code perspective to create within this community is a central cross-correlation engine, whether that will be a federated group of AI instances or a central AI model that will basically take all this and cross-correlate it. We're looking at all of that right now. We're actually looking at the middleware layer, and middleware is probably the wrong term for it. But we're looking at Layer 7, I mean, all these different areas where we can basically integrate all the different applications in an effective and secure manner. We're going to be building additional AI models for missing functions and also, of course, synthetic data environments for model training. Next slide, please. And so really, the benefits here, I think, are pretty clear to everybody on the phone, I would imagine, or on the conference call. Wide frame real-time business insights will be really valuable. We're actually looking at creating a hologram advisory that will then, in turn, turn around and advise the C-suite and management as to what's happening in real time around the enterprise to streamline operations and enable those cost savings, pretty much everything you would imagine with both a combination of AI ops and business intelligence. That's the idea with this system. But if you think about it, all these different elements have really been sitting on the periphery. And these key leads have been drifting together for some time now. So we're just looking to create an intelligence in the middle and then basically integrate everything at the edge. That's the idea. And then last slide. So the reason for all this development and the reason we started with telco from this perspective is if you think about the pervasiveness of a telecommunications network today and the size and scope of these kinds of networks, we've got companies like Equinix and América Móvil and we've had Yahoo!, et cetera, these are some of the biggest wireless networks and CDMs you can imagine. And the idea was to really start looking at the meta-scale question of how would we manage climate change. But also, if you look at the terrestrial and extraterrestrial networking that's going to take place in 6G, I mean, you're going to have amazing uptime at almost every location on the planet and maybe a little bit beyond. Could we do extraterrestrial object detection as well like meteors and impacts like that? How would we create a system that basically would manage and operate all this? So again, this is a research and development consortium. We don't plan to answer these questions next week with a new product. The idea is to incorporate academia, government and industry to basically sit down and talk about these issues and come up with a system that could eventually have a really positive impact on how we interoperate with our planet. So that's really the idea and the concept. We're working on it actively now. We have a kickoff actually tomorrow to support the Cookie Box PoC with Harvard and Harvard Analytics. We have UC Berkeley's Data-X program now involved. We've got -- gosh, I could go down the list. We have Intel and other partners that are helping us, and IBM Research is also contributing one of their scientists as well. It's been a fantastic effort. So we are going to continue this on over the next few years and see where it leads us.

Thank you, Bill. So when you mentioned central AI, I got a little nervous because I thought, now this is where the next reference is going to be hiding or something like that. But your vision of central AI is not -- it's saying that you are not at the edge. You are -- you can work with any edge that might come up. So if you can just demystify a little bit the world centrally doing complex AI for very large distributor systems like wireless architectures. So if you could bridge the 2 concepts, it would make it a lot more clearer, if you can just comment on that a little bit.

Certainly. And I actually had a slide about this that I'd culled from the stack because I didn't want to go on too long past 5 minutes. But in essence, we'll be taking a look at all the lower-level functions like FFTs and different forms of AI that will be deployed in the field. We'll then basically do a tiered architecture, and that's at least the first pass of this architecture we're looking at. And so it will actually -- every segment of the business will basically be funneled into yet another segment and yet another segment into a tiered architecture that basically gets into GANs and transformers and takes it up into a reporting instance that then cross-correlates all the larger data sets and the metadata to basically create an output in real time that can be interpreted at the top level. So it's really a tiered approach for the time being. There are folks within the community that are advocating a central model that will manage all of it from an efficiency perspective and an accuracy perspective. We're going back and forth on all that right now, so I think it's an interesting dialogue we're having. And we'll be doing things in the lab to basically try to prove that out and understand what that looks like, but we'll be basing it on the research we'll be doing.

[ Well, very ] interesting. I'm sure that there'll be lots of active debate on both sides of the aisle on this, so looking forward to that. Great. So well, all the panelists have provided their vision talks, and there are certainly words of wisdom that at least I have extracted. I have a bunch of questions myself. And I do see a couple in the chat, so what I'm going to do is I am going to pick out a few and I'll leave it open to the panelists. If anyone wants to jump in, please feel free to do so. And then you can also voice other concerns or opinions that may not be specifically addressing the question. We welcome that as well. And especially if you don't agree with your -- what your colleague has said on the panel for -- go ahead, please, and say it loud.

Okay, so with that said. So one of the questions is do you envision data collection platform being integral part of all network deployments towards next G network architectures? So in other words, will the data collection [ must ] necessarily exist in network deployments? Or do you think that, that is perhaps not the role of the network operator who is sort of trying to optimize the operator's own performance? Any thoughts on that?

I just want to clarify when you mean data collection. Of course, operators collect a lot of data on the network performance and so forth, and they will continue to do that. Is -- does it mean like whether they collect user data? And they also do that, but there will be increasing concerns of privacy. And that's why some of these new technologies such as distributed, privacy-preserving learning have -- coming up [ and see whether ] they will have a role on that. So I don't know whether that is the clarification...

So you're right, Nageen. Maybe I should call I -- the way I interpret this is that should this data collection platform be more accessible to perhaps outside the operator's own domain. I think I interpret the question as in that manner, but yes, I think you did touch upon the point and privacy-preserving learning, et cetera. I also see Nada with her hands up. Nada, go ahead.

Yes, sure. So as Nageen said, there's a lot of data that's being collected and will continue to be collected, but maybe one thought perhaps for the future networks is that maybe no data needs to be collected, right, because things are being learned or sensed or being -- and dealt with right there. Right now all we do is we take the data. We shuffle it back and then we process it later, and then we find the insight and then we act on it. Perhaps maybe that's not needed, right, if things are going to be learned more quickly. Just a thought.

Indeed, indeed. No, yes. Nageen, do you want to add a follow-up...

Yes. I guess, I mean, of course, there needs to be some data for validation and so forth, but yes, you could really reduce the amount of data that needs to flow by using local learning and analytics. And that's really the way to go. I mean there is no need to [ ship ] so much data around.

True. And just a few seconds of IQ sample data itself is gigabytes, and at this rate, there'll be nothing left. Or perhaps I shouldn't be saying these comments. And probably there are other more -- better people have said things about storage and have been proved wrong, but I do agree that storage could become a concern pretty quickly. Okay, another question is that there were -- so there were mentions of technologies like terahertz and whether 6G will involve frequencies that high with all the stream bandwidths. And so at those bandwidths, computing time probably is the dominant factor in the end-to-end latencies. So clearly if computing time is the dominant factor, then running AI/ML at [ line ] speed or in the loop is going to be even more difficult. So just a thought here that, if 6G were to have terahertz, how would you do some of these machine learning methods as part of your online system. Any thoughts on way to -- ways to handle that? Jakob, you can go.

Okay, sure. I can start [ with the ] question. So I'll -- I see 2 possibilities. So one is, of course, in the waveform design itself, but that doesn't mean that you need any form of online learning. Essentially you would actually come up with a new waveform that's entirely learned based on models and validated with some data, but once you have it, essentially there's no learning anymore. And then [ I think ] that's easy. And then the second part is, now for the detection, it becomes actually extremely challenging at terahertz to deal with [ hardware and protections ]. And as an example would be excessive phase noise. And there has been quite some work already, but I expect much more to come up, that actually shows that there's a benefit of using, for example, neural networks for detection. And now whether you can actually operate these terahertz kind of throughputs that's required, that then boils down to do you have an efficient hardware implementation for it or not. And I think we still don't have it. We don't have these, I would say, communication-specific [ hardware ] accelerators, but I think it's just a matter of time that these will come. And then I -- it could actually even become possible that this will be a lower-complexity implementation compared to a traditional algorithm. So yes. And I don't expect any online learning. So I don't see that -- most likely you won't do online learning on top, so [ I'd rather see it ] you train it from time to time and then you deploy it and use it. So...

Just to follow up on that. I think the audience member who asked that question also offers a solution, says that should we move to quantum devices or -- and that opens up a whole new kind of world, but if -- but that's one thing that we haven't touched upon in this whole discussion of AI/ML with regards to 6G. But where does the quantum story fit in? Or are we too early yet for that? Any thoughts on this from the panel?

Do you mean too late? Are we too late?

[indiscernible].

[indiscernible]. That's a good perspective indeed, but is there something that we should start to incorporate already? Knowing that there are economic challenges in creating these sorts of computing engines, but should we start [ to build clearly already for this ]?

I have some thought. I think it's we are not ready for this, and it's too costly and won't work at scale. And then one also need to think about these terahertz links. They -- we won't use them for coverage, right? It will be mainly -- I don't have many good use cases in mind, but I think some that -- kind of idea, these data towers. You're in some shopping mall, and for some reason, you want to download terabytes of data then go to this data shower. Boom. You get everything you need in a couple of seconds. I mean that's really a [ neat ] use case. And the question is do you want [ as in ] quantum devices there that costs a lot of money, or can you do something else? So I think it's a bit too early. I mean the quantum [ disruption ] is coming, but I personally don't expect it to be there for 6G. So...

That's certainly point of view, Jakob, well taken. Nageen, do you have a counter to that? Or are you...

No. I was going to throw it to academics -- sorry. I was going to throw it to an academic who are doing quantum research because I'm not that familiar with this yet. I don't -- I also sort of concur with Jakob here that it's probably...

Indeed, okay.

If you want to assign a G to it, maybe 6G is too soon, but if we say next G, some G, then maybe it is. And the research has already started. Question is, is it commercial ready? Probably not, right, but it doesn't mean that it shouldn't be worked on. And as for the terahertz -- and as long as we don't assign a spectrum or frequency to a G, I think everything is on the table. Everything is fair game, right? So we should use whatever makes sense. So for terahertz, things that are probably very short distance, sensing, things that where is -- terahertz is best, but we're not going to try to deploying it maybe in an urban area, at least not in 5 years.

It makes sense. No, that's insightful. So here is a question from the audience, and this one is actually pretty close to my heart. We are talking about using machine learning to -- purely in the RF domain, but if you look around, you find massive amounts of cameras and acoustic sensors, lidar, et cetera, right? We are surrounded with these. When will these multimodal sensors start to come inside a regular wireless optimization loop and become part of a multimodal decision process? Is 6G going to do it, or should we wait longer? And why is it that we don't see that much of, let's say, active -- or perhaps there is an active development, but an active incorporation into the standard of these multimodal sensors for RF-based applications. So there's a lot of questions, so I'll be happy if each of you take a crack at it. Jakob, do you want to go first on this? And then we can go Bill, Nada, Nageen.

Sure. [ I mean looking at ] multiple parts to the question. So in terms of standardization, I think there are even already discussions about this; and how could you, yes, incorporate additional sources of sensing information such as cameras into the network. I think it just hasn't materialized yet, but definitely already 2 years ago, I was aware of a couple of -- involved in a couple of [ discussions ]. And it makes sense. Now the question is -- I think, is this fast enough? So in many cases, having a camera or something like this, is it really slow compared to how fast wireless works? So if you get 30 frames per second, that's too little. From -- in wireless, it happens on [ the microsecond space ]. And in many cases, that's possibly a bit too slow. So I don't see right now a lot of use for optimizing wireless using additional sensing information. I think possibly -- I mean there are many cool applications arising. It's just I haven't really seen it yet, so -- but I think that it might actually be other -- the other way around. So saying most likely 6G will be an additional sensor that can complement cameras, lidar. So for example, there are some also some nice work or with people that actually use these for virtual -- kind of the headset for augmented reality and then use RGB cameras to get these depths. You use depth cameras to actually be able to visualize something, like an virtual object [indiscernible] you need to estimate the depths where you are, but when it gets dark kind of, these cameras don't work so well. And then [ we're going to use ] millimeter-wave or terahertz frequencies to do -- to replace or to take over the role of these cameras. So I see them, I think, in one -- in the other way around. And yes. And I mean, otherwise, going to 6G, I mean, that's one of the big use cases. In 6G, we go to higher frequencies. The base station will essentially become a sensor. I mean that's key to it. So I think that's one of the things I expect -- [ will ] expect to happen.

Great. So you did touch upon the sensing question. We are going to come back to it. That deserves a whole set of responses from the panel in itself, the role of sensing, because we spoke a lot about communications. So we will definitely touch on it, but let's come back to this question on multimodal sensing for -- or multimodal sensors to shape wireless optimization. So Bill, would you like to go next? And then we go to...

You saw me smiling here. I wrote a white paper internally at a company I was at about 10 years ago about this topic and something kind of similar. What I was proposing at the time was to basically take mobile devices and use them as resources as part of a cloud pool, you could say. And it's really funny because, when you think about it, there are, I mean, millions upon millions of chipsets sitting there. And as I'm talking with you on this conference right now, this is sitting here totally latent, not being utilized whatsoever. It could be part of a pool of resources that could be used from both a sensory perspective but also from a compute perspective as well. I think 6G will open the door to that, but then there are also the issues of do I want to opt my device in for something like that. How do I protect my onboard data? How do we take a look at that? There are some issues with privacy that come to the fore, but again, we give away our privacy all the time. We opt in to all the usual programs that we do on the social media side. So it's fascinating to me. I think, in some respects, we have to be very guarded and careful about it. And I think, in others, we've already given that freedom away for some of the things that we use on a free basis every day. How do we personally want to navigate that? But the resource pool, I think, for devices is massive. I'm just -- I think that's something that's an untapped resource we need to take a closer look at. And 6G will help unleash that, I think, over time.

That sounds great, Bill. Nada, would you like to go next? And then we'll go to Nageen.

Sure. 2 comments. One is that we are already using AI-based video and cameras to do propagation modeling, for example. That's we need to -- when we look at lidar data and RF data and in order to come up with good models. So that's already happening. And then you said you wanted to take the sensing, but there's basically lots of work on joint wireless radar-type sensing and applications, comms and sensing, that's already happening, and standards too. So we'll talk about it perhaps later.

Sure, sure, but before we move from you, Nada: You're saying that this is going to be an integral part of 6G. Or should we wait further on?

See, I'm not so sure. When people talk about 6G, maybe they have something in mind. They already have some requirements. To me, the next G is the next G. At some point, the marketing folks will come in and will say, hey, this is 6G. So I don't know. From my point of view, it is the next G, meaning it's not there yet. Or it will be there maybe in 5, 10 years, but I'll leave it up to the marketing folks to determine if that's 6G or not.

[indiscernible], yes...

[ Yes. It is impressive ] marketing, [ yes. I'm sorry ]. No, I was laughing and agreeing completely with the marketing aspect, but yes, it -- just seeing that funnel increase has been nice, though. I think it has opened up a lot of freedom, but my apologies. Please continue.

No, no. This is great. [ Please continue ] to chime in. So Nageen, over to you, please.

Yes. I guess I don't have too much to add beyond what people have already said. So having mobile devices as sensors and base stations offering a platform for additional sensors is definitely very promising, on the other side which, I was going to make the point which Jakob made that wireless as a sensor is actually very interesting because it's also [ a part of life ]. A, first of all, it's pervasive. And B, it also has privacy properties, which makes it very interesting. So just to add that point.

Great. So since we are on the topic of sensing and wireless as sensing, right, let's just drive home a little bit further down. So how do you see it? Do you see that, the data or the information that you will be able to collect from using a wireless as a sensor, you will learn on them and then you will optimize your actual wireless network operations using this? And would you be able to do this in a -- in time scales that are meaningful enough? And if so, does it need to be at the radio front end itself? Or can it be -- or can you offload some of this into some sort of a cloud in the back? Is -- are there some examples that you can share, anyone from the panel, that...

So you mentioned something about sort of using the RF to do sensing but -- that's fine but also for many different applications, right? So we're using the RF, the wireless to do detection, presence, to do counting, to do surveillance, to do -- without the camera. So we're using this, the RF, the communication to -- for different sensing types applications. So that's an area that's very promising, it seems. And it's already started because we also don't want to pay the price twice, right, the comms price and the sensing price. So we want to just say, okay, since you're communicating, we're going to use that signal, that RF to -- in order to figure out other things about the environment. And then we can come back and optimize the wireless, right? So optimized wireless could be one application, but it could be also for other types of application.

I think -- I appreciate that point. So it could even be it may have nothing to do with the wireless optimization. It would just be understanding the environment for a variety of different other [ purposes ]...

Yes. For telehealth patients, are they breathing? Are they not breathing? How many people are here? This is COVID. We're reaching the limit of...

Right, many, many examples. And these help to -- for me to visualize some of the use cases. Any other thoughts on that from the other panelists?

Yes. I mean I fully agree with what Nada said. Possibly I see more interesting cases arising there, but I think there is already a lot of, I will say, use of RF sensing for optimization of the network itself going on. So for example, you can detect -- [ and probably ] we can estimate [ how fast to uses move ], right? That's essentially radar. And depending on that, you would choose at -- the frequency with which you send [ channel point ] feedback, channel estimation, et cetera. Then you can use RF to localize [ users ]. That's what's happening. And this type of information, you know -- [ you can know what to make that scheduling ] decision which [ users ] should be jointly [ multiplexed ] to ensure some, I would say, [indiscernible] between those channels. So in some sense, that's already happening, but it's not like -- the marketing hasn't gone well. You're doing this to some extent already, but it hasn't -- you don't need 6G for it. Or you don't have to do specific radar there. So I think it's going on, and most likely -- I would say that this is just going to continue. So people will discover. Okay, now we can actually not only detect the speed, but now we can really pinpoint what's the use? I can really figure out if there is a blockage coming and I need to hand over to the [indiscernible] kind of things, but most likely they -- this won't be a marketing argument. It's just something that, I mean, naturally happens. So I think, for 6G then, to brand it, it will more like the what can you do on top of communications. [ That's all ].

Sounds good. Sounds good, Jakob. So let me switch gears to a different topic. And it came up when we were talking about -- and I broadly call this as the role of the human [indiscernible]. Now when you talk about AI and machine learning, the thought process is naturally that here is where the algorithms and the devices, the computing infrastructure takes over, removes complex decision-making from the individual network operator, is able to reconfigure change, adapt faster than what perhaps a human perhaps could anticipate in both near-term and long-term scales. Or perhaps it will handle vast amounts of data that is really impossible for a human. Now what is the role of the human operator in this future world where AI/ML is firmly entrenched into a 6G standard? Is the human only an end consumer and that's it? Or is there some guiding role that a human can play to shape what the AI/ML does? So this could be a question for the entire panel where -- and if anyone would like to go first, please unmute, and then we'll do it that way.

So just -- sorry. Maybe I'll just state the obvious. Look I -- just there are still some questions about how bounded and how reliable AI solutions are. And until those things are better understood, you may still need AI to provide some recommendations and get some human in the loop to address. So I think you do see in -- for a foreseeable future that human should stay involved, but as things get better understood, I mean, that's obvious then. [ People will get less ], but yes, yes.

But we still have to deal with the human in terms of the delivery, right, of services, right, because we're dealing with the human [ as in ]. And like you said, but -- so the human is there, but also a lot is happening between the machines themselves, right? And for the comms to be smart or intelligent, that's we want to remove the human, as much of the human that is needed for doing very simple tasks, repetitive, things like that. We want to remove the human, right, but we want to make sure that the human is getting what it -- he or she needs. But then these things are supposed to be autonomous of sorts, right, the comms...

Yes.

Yes, yes. So Bill, go ahead.

I was just going to like once again step back and look at the meta level again. I guess that's all I'm good at basically, I think. What's funny about it, though, is if you think about it, agriculture in the United States: I think 80% of the population was engaged in agriculture and related activities in 1800. Today, I think it's 10.9%. And if you look at that evolution, over time, it really came down to automation in many respects. And so what happened to the rest of that population? Well, it doesn't mean they're necessarily sitting around at home. They've found -- they've been retrained. They're doing other activities. They have other professions, et cetera, but that's taken place over many, many years. I do think, if you take a look at, for example, image recognition in the medical domain and X-ray detection of different, let's say, anomalies or different health conditions -- I think the error rate for the machine algorithms currently stands at around 6% or 7%. For humans, it's closer to 3% for X-ray experts at a PhD level, but if you combine the two, the error rate drops below 1%. And so the marriage of human and machine, I think, which is a fascinating proposition, is that it becomes an adviser. It becomes a synergistic relationship between the two because of the strengths inherent on both sides. So I think over time we'll see a melding of the two, as opposed to one displacing the other. Just a thought.

Absolutely. Jakob, any concluding thoughts on that?

Nothing really smart to say. Sorry.

No. It's okay.

No. I just think it's not like -- I was thinking about what Bill said about farming, so I struggled to make an analogy between operating a network. And it's just a few hundreds of people doing this. And when you see you have automation, you're automating thing, it's like -- most of network operators, they have maybe a couple of hundred people that changed one of the settings in the network, get their KPIs, look at this and then make informed guesses about what's really happening and try new things [indiscernible] again and again. And I think -- and it seems natural that most likely machines could do this job in the future, but we are not speaking about percentages of the population. So I think it -- sorry about that, yes. So I...

Right. No, it's a fair point. I think -- so my takeaway here is that I think the panel uniformly has agreed that the human still has to be involved and sort of guide the AI/ML into, say, maybe faster convergence or perhaps identifying what sort of data to consume given various options. So there is certainly a role to play, but I think it's an open research question now, that where should -- interruption and the interjection of the human is most effective and when it is actually a collective action rather than sort of not allowing a child to really learn from his mistakes. So I think I do feel that that's an open question as well. And perhaps then -- as a community we could look into it a little bit more from the context of a network and says, if I don't interject here, what will happen in the network? If I do, then what would happen? And really understand these use cases. But yes, a fascinating question about what does the human do. So here is another thought from the audience. And the audience asks that, well, how do you handle the uncertainty in machine learning in a way that it'll help 6G networks to be more responsive? And in parenthesis, it says "no latency in dynamic environments" and especially for safety, mission-critical applications. So bottom line, uncertainty in machine learning exists. We know that. There is -- and how will that be incorporated or somehow managed or even utilized by 6G networks in extreme environments like ultra-low-latency or extremely dynamic environments, mission-critical environments, et cetera. So handling [indiscernible] question. Is there -- I mean go ahead. Jakob, I think you're saying something.

I'm not fully certain how we can leverage uncertainty in a positive way. [ It's this is a ] question, I think, if -- I rephrase [indiscernible]...

You measure it. You measure it, first.

[ I would have said ] how can we use machine learning to actually make 6G systems to be more responsive and robust in mission-critical applications but not how can you use the uncertainty of machine [indiscernible]. Maybe this is something in the question I don't get. So this is one use case I really dream about, what I think might actually happen. I mentioned this in my introductory slides. I think there's a huge potential for digital twins. So we think about this fully automated factory: You have a digital twin of it, meaning everything that's happening in there, robots moving around, et cetera, but you can fully essentially simulate it. And now you can -- now the question is how can you design a networks that ensures communication with -- as if it was scaled, so no errors. And now because you can learn everything, you can use machine learning to decide when would you place your access points that you can actually fully anticipate what's happening and actually honestly resource-allocation problems you need to do, when to hand over, [ what to do ]. This is not a -- you actually take the randomness out of the entire problem. And the randomness in communications is uncertainty, right? [indiscernible] and then you can essentially use machine learning to predict everything. And so there is no randomness and uncertainty anymore. So I will say using digital twins with machine learning allows us to get rid of the uncertainty we have in classical communication schemes. So that's kind of one use case.

Well, it's interesting, yes, yes.

I can -- but -- yes. Obviously it wasn't a concrete use because I [ can't ] kind of imagine what could happen. So that's it. [indiscernible].

That's great...

Yes, I was going to add to what Jakob is saying. So there are basically ways to reduce uncertainty. And Nada actually just alluded to one, which is having more diversity in the system, like multiple digital twins, human. And you mentioned, Bill, that human combined with AI, plus combined with digital twins, will reduce the uncertainty; and that's basically using diversity of options. And then second, I will say that there are, of course, now efforts to have safe AI or bounded -- kind of bounding the output of AI solutions. So mathematically, there is a lot of work that is being done to sort of at least see how robust a particular solution is. And those techniques will be important.

Jakob -- I'm sorry. Go ahead.

Go ahead, Bill. I was about to ask other inputs from the panel. Please go ahead, Bill.

No. I thought Jakob's line of reasoning was fascinating, and I couldn't agree more. I -- just sharing a personal story: I was thinking to myself a little while ago. I was thinking, what if you could create a digital twin of a human body and you could do it for each individual based on the kind of wear and tear they've had over the years? Because I recently took up playing rugby again in my 50s, and it was funny. I look back that -- when I was 18 to 25 and all the crazy things that I'd done to myself, and I was like, what if a digital twin could actually tell me what my injury rate is going to be at this age going back to the rugby field? And playing touch rugby, not tackle rugby, but just for the fun of it. But what was interesting is the digital twin model could apply to all these different areas of biology and different systems. And I mean across the board it could become a massive data problem but a massive data solution as well. So I think he's set me on the right track there.

Great. I mean I'm just thinking aloud. Maybe I can send my digital twin to my office hours while I [indiscernible].

No, no. That's called cloning.

[ Yes ]...

No. We'll just have to wait for Jakob's technology to catch up.

It's probably not 6G but maybe some other G.

Some next G. I fully agree with you. So one question that a lot of our students have -- and I ask this because we have some of the top companies here in this panel who -- or many of our students who, when they graduate, is planning to work for, right? And as such, they see the activities and the approach that large companies take -- that are really the movers and shakers in this whole area. They would like to -- they see these and they would love to emulate it and so on. My question is this: When it comes to data sharing and transparency, there's always going to be a barrier that is going to exist because companies exist for a reason. Now not all data can be shared and not all data -- perhaps there are certain economic [ costs and ties ] that limit these. I do understand that, but then -- but the general thought process is that it's going to be very hard to get data from enterprise-scale networks. For example, I can maybe do some front end on my lab and get something out of a channel, but it's hard for me to get backhaul data from some enterprise network. And this is just for example. There could be many others. What is it that industry can do to ease this accessibility? Is there a real intent as well to do this, in the first place? That's number one. And if so, what could be done so that more and more students and researchers don't have to wait to join the company and then start to contribute but can do so at a grassroot level and get trained on some of these complex problems that exist perhaps in these companies. So I'll stop here, but I'll really open the floor to your inputs.

Yes. I can go first. I think it's a fantastic question. I mean I think that's really at the heart, I think, what's hopefully going to happen and change over the next few years when -- because I mean it's not only students. I see this massive kind of separation between academic research and what's relevant to industry. And that's not only a problem of [ getting access to data ]. It's actually in academic research you have limited compute, but in many cases you want to do [ something mathematically, sort of ] resort to simpler models where you can prove something and show something. But that's a very long way from actually putting it into product. And that's why it's not immediately relevant to industry, but there is really -- nobody bridges this gap. And so I think what needs to happen is that the academic research uses data set, realistic simulations, possible test beds to make research more relevant, but someone needs to develop essentially these platforms that allow you to do actually these realistic simulations and building test beds without too much effort, I think. So this is where industry comes in. And I really hope that this is going to happen, but I mean there are increasingly -- with Open RAN, there are more and more opportunities actually for even universities to build 5G test beds, hopefully 6G test beds, to use it, try the algorithms on real system and not just a toy model. So it makes a research more credible. This will also allow to create data sets that are relevant. I think you are having one of these initiatives. I think the [ MIT ] versus RF challenge, which I found quite fascinating. So I hope that this is now really gaining momentum, but I think that's the change that's needed in our community to bridge this gap. But I think all the elements are there: increased openness, sharing, reproducibility of research. That all goes hand in hand, so -- and I think we don't need data of millions of private users for our research. We speak about RF data, so it's most likely price is not so much an issue for this type of research.

And I want to -- if I may, I want to say that this is already happening. The community is coming together. Things, partnerships like the RINGS that we're -- some of us are part of, we're going to see more of that, the NSF power test beds. There are test beds throughout the country. That is happening. I think what we can also do is perhaps push the test bed, the research to things that are not just already last G but sort of in the prototype development phase, right, so that we can actually work with the community, the research community, for things that are not already there but will be there so that we can test things and prototype things and -- but we're coming there. I mean this is -- in the last 20, 30 years that I've been in this community, I don't think we've been any closer to this than now. And I think this is going to continue.

Kaushik, you are muted.

Yes. Sorry. Yes, well, when guest speakers talk, I sort of mute myself and then -- but so, Nada, I appreciate that. Nageen, do you want to conclude something there?

No. I just want to echo Nada that some of the RINGS program and also our MLWiNS program is just trying to do exactly that in partnership with NSF. And NIST is, of course, involved in some encouraging data sets and so forth. It's going to be continuing to be challenging for industry to share data, but through these challenges, I mean, that's one method; and methodology of AI methodology itself, like Jakob mentioned digital twins. I mentioned predictive learning and so forth. So those are tools that you can apply to solve some of these problems as well.

So we are -- as we're about to close the panel, I would still like to charge the panel with one additional request, okay? And that request is -- so let's say we have a 10-year span from now, okay? What would you like to see that the community has enabled and now is part of the 6G standard? It's -- [ so about like ] 10 years from now, 6G is sort of midway through. It's sort of perhaps picking up steam, so what would -- is it -- what is the one thing that you feel, if it was there or you would like to see it there, and you feel that would be a big game changer? I've put you all in the spot like this. These will be the -- [ all ] questions I'm throwing [ out of the gut feel ], but I would love to get a response from each of you on what is it that you'd love to see in 10 years time.

Okay, I can start again. There are 2 things. One actually is not related to machine learning at all. And as I mentioned that -- at a one point in my introduction, is how long it is taking MIMO to go from theory to practice. And I still think that 5G has not gotten it totally right, so we are still far away from getting true reciprocity-based beamforming where you can scale a large number of users. I think that's not happening. We also see that, at a millimeter wave, MIMO is not [indiscernible]. Then we get these things right for 6G. Now with respect to machine learning, I would really hope that there will be signaling and procedures in place that would at least allow for someone who wanted to deploy a learned waveform and use it. Because right now you can't do it. So I'm not saying that 6G should be based on -- is something that you have learned or it should only be based on this, but [indiscernible] possible to part kind of learning of a physical layer on some parts of the spectrum, if you wanted. I think that would be a big achievement.

Who would like to go next? Yes.

I'll just add, yes. So I guess, since I'm -- this is my biased view. Being a researcher in the area of distributed AI and so forth, I would like to see end devices having capable of AI training and online training and so forth, where it will really make AI solutions much more pervasive than they seem to be and really moving out from the data center all the way through devices. So that's a [ wish, personal wishful thing ].

All right. I'll put it in the box of -- the wish list box that I'm making here -- no, but that's great. Nada, do you want to go next?

Sure. I guess for me it's going to be a very sort of down-to-earth comment. It's that we won't ask where the data is coming from. It will be there. And we will use it and we will be completely seamless. And we won't be asking, "Where is the data? Give me the data. I don't have the data," [ et cetera ].

That's a loaded ask. Well -- but I completely -- it's a small ask that you said. You compressed it in one sentence, but there's a lot going on behind it, so -- but yes, okay, that's certainly in the wish list box. Bill...

I'm just impressed with Nada's sheer optimism now. It's impressive, I think. If I look at where I'd like to see all this go, it's interesting. There is the convergence of cloud and fog. And I look at all these different things coming together; and the convergence of space [indiscernible] Internet, all these different things. What I'd like to see is what was mentioned earlier. It was just to find these kind of large-scale sensory networks that we can create to basically enable things like load-balancing climate change; if something happens in one part of the world, noticing what the impact is in specific ways in other parts of the world; learning to basically contextualize that and make it more intelligent. And then -- because all the pieces exist. We just need to network them together. And I think 6G could be just another step in that direction to enable that technological leap, but I think we just can't -- we can't avoid it anymore. It's something we really need to look at.

Fantastic. I think that's a very positive note, a very positive -- and hope for the future as well. And so let's conclude the panel here. I would -- want to say thank you to our distinguished panelists, Nageen, Nada, Bill and Jakob -- for taking this time out of your very, very busy schedules. I have tried to get on your calendars sometimes at different times. I have been on the receiving end, so I really appreciate that you've been able to carve this time out for the audience. Also, for the audience, just to let you know that, by next week, we will have the recorded sessions made available to you. You will get an e-mail. So you will be able to access all these great programming content at your leisure through -- by logging in into the 6GSymposium website. So in that case, analysts, thank you very much. Appreciate your time. And we'll be in touch on one on one. We'll continue the discussion forward. Clearly this is just a step. Thank you very much.

Thank you.

Thanks. Bye-bye.