Today’s episode features an interview between Matt Trifiro and Mahadev Satyanarayanan, also known as Satya. Satya is the Carnegie Group Professor of Computer Science at Carnegie Mellon University, and one of the true Godfathers of Edge Computing. In this interview, Satya shares how he started thinking about distributed computing infrastructure for mobile devices back in 1993, how much of his vision has come true in the decades since, and his views on the future of cloud, edge, IoT, and much more.
Today’s episode features an interview between Matt Trifiro and Mahadev Satyanarayanan, also known as Satya. Satya is the Carnegie Group Professor of Computer Science at Carnegie Mellon University, and one of the true Godfathers of Edge Computing.
Over the course of his multi-decade research career, Satya has pioneered many advances in distributed systems, mobile computing, and IoT, and in 2009 he co-authored “The Case for VM-based Cloudlets in Mobile Computing,” the groundbreaking research paper that led to the emergence of Edge.
In this interview, Satya shares how he started thinking about distributed computing infrastructure for mobile devices back in 1993, how much of his vision has come true in the decades since, and his views on the future of cloud, edge, IoT, and much more.
Key Quotes
“In 1997 we said, given that [the compute capability of] mobile devices is always going to be a challenge, how do we get substantial applications that require compute-intensive processing to run on a mobile device? The answer is to offload computation to the infrastructure. We were the first to demonstrate that capability in a published paper."
“When it comes to user experience, people have learned that it's not the mean that matters, it is the tail. Human user experience is greatly overweighted by a few negative experiences. You may have one hour on a zoom call, and 58 minutes of it may be excellent, but you will remember the two minutes that were miserable. This is generally true for augmented reality, and for all of these other [latency-sensitive] use cases.”
“It's useless to deploy 5G without edge computing. The truth is 5G is only going to improve your last mile quality. That's it.”
“How much end-to-end latency is acceptable is very much a function of the application. But it's a two-way street. The applications that get written depend on what today's technology can offer…If the application's demand gets too far ahead of what the technology can offer, then the application will die because it's not viable.”
“The ultimate beneficiaries of edge computing should make strategic investments that incentivize the creation of edge-native applications, and apply a different success criteria from what is traditionally applied by venture capitalists...That investment is valuable even without a hockey stick growth curve, because you are creating long term demand for the core product that you're creating, which is edge computing itself.”
Sponsors
Over the Edge is brought to you by the generous sponsorship of Catchpoint, NetFoundry, Ori Industries, Packet, Seagate, Vapor IO, and Zenlayer.
The featured sponsor of this episode of Over the Edge is Catchpoint. Catchpoint gives critical knowledge to help optimize the digital experience of your customers and employees. Learn more at catchpoint.com and sign up for a free trial.
Links
Open Edge Computing Initiative
OTE011 Satya transcript
[00:00:00] Matt: [00:00:00] this is Matt Trifiro, CMO of edge infrastructure company, vapor IO, and co-chair of the Linux foundations state of the edge project.
[00:00:06] Today. I'm here with Mahadev Satyanarayanan, also known as Satya. He's the Carnegie group, professor of computer science at Carnegie Mellon university, and is considered one of the fathers of modern edge computing. Personally, I'm very excited to have this conversation because some of his early papers on edge circuit, 2008, 2009 helped me personally wrap my head around the true revolution at hand and why it's inevitable.
[00:00:28] We're going to talk about Satya's career in technology, the evolution of edge since Satya first introduced the concept of cloudlets back in 2009 and where edge is headed in the future. Hi, Satya, how are you doing today? Right?
[00:00:39] Satya: [00:00:39] I'm doing great. Thank you very much, Matt.
[00:00:42] Matt: [00:00:42] Yeah. Terrific. You know, before we, we dive deep into edge computing, I'm just really curious.
[00:00:47] How did you even get your start in technology?
[00:00:50] Satya: [00:00:50] that's a great question, too. so I was always interested in science from a very young age, probably from the age of five. and I think by the [00:01:00] time it can time to go to college. I picked electrical engineering. This was in India. the computer science didn't exist at that time in India. And, it was only in my junior year in college that my university received its first computer.
[00:01:18] It was a large mainframe IBM mainframe. This was 1973 and there was a open, lecture series to teach students programming. And I took it and I was hooked. And so for the rest of my life, I've been interested in computer science and that's how I got my start
[00:01:41] Matt: [00:01:41] It's interesting, you know, having started in electrical engineering on the hardware side of it, and then moving to software, but in a way, a lot of what, a lot of the problems that you're studying and solving for and John will edit out this pause. They [00:02:00] get down to the physics,
[00:02:02] Satya: [00:02:02] that is a very insightful observation. And I think the point you're making, which I do agree with is the fact that my background came from doubly means that I really truly understand at a very visceral level, what it takes to do computing. So most of computer science is all about abstraction. It doesn't matter what Greg
[00:02:25] Matt: [00:02:25] in computer science are solved by another way of faction.
[00:02:28] Satya: [00:02:28] precisely.
[00:02:29]but the point is, given my roots, I am conscious of how all of this high-level stuff gets finally, how execution actually happens. And so you are correct in observing that some of the insights that led to edge computing. Burrows because of that awareness.
[00:02:50] Matt: [00:02:50] Yeah. At some level we're pushing electrons and photons and around, and there are constraints that those does introduce. so, so tell me [00:03:00] what you do at Carnegie Mellon. Like what is your role there?
[00:03:03] Satya: [00:03:03] I'm the, I'm a professor of computer science. So I do the things that professors do, which is we teach the undergraduates and graduate students. I do research with my PhD students with post-docs and with faculty collaborators, and also with industry collaborators. So I have a very rich network of people that I work with at all levels.
[00:03:28] And equally importantly, I worked very closely with industry to both learn and sense the new challenges that they face. And also to help transfer the knowledge that we have gained through our research. And so it's a two way street. And, vapor for example, is one of those companies, Seagate is another, we have a, group of companies called the open edge computing initiative that was created by my team.
[00:03:58] And, so there are [00:04:00] many companies which paper and Seagate uphold a very important parts. And there are many other companies that I'm sure you've heard. familiar with that are also
[00:04:08] Matt: [00:04:08] Well, yeah, in fact, I think we came, became familiar with your program through crown castle, one of our co-owners and, and obviously a big partner in your lab.
[00:04:16] Satya: [00:04:16] Crown castle has been a strong supporter and they are great partners to work with.
[00:04:22] Matt: [00:04:22] So when I, participated in the first state of the edge, the first day of the edge report, which was 2018, I did some research to try to figure out what the origins of edge or, and the first reference that I could find, was the founders of Akamai. When they're at MIT, wrote a paper, really describing what was to become a CDN, in fact, the largest CDN in the world Akamai and they described, a lot of what we think of today as in pushing the stuff and I would call it, you know, compute and data.
[00:04:54]but at that point it was mostly data. It was caching, farther out to the edge to solve for a lot of the [00:05:00] problems then, which was just, you know, consumers wanted access to video and large images and. But the internet at the time, was it good at supporting that at the kind of quality of service they wanted?
[00:05:10] And so that solved for problem. when did you become aware of edge computing as something that you thought was important? And what did you see about it? That was important.
[00:05:23] Satya: [00:05:23] That is a long that's a, the question requires a long answer. So I hope you will be the bear with me as I give you the answer.
[00:05:30] Matt: [00:05:30] would like the long answer. So yes, I'll be very patient.
[00:05:33] Satya: [00:05:33] Okay. so what's your describe was Akamai coming at this problem, starting at the cloud and moving towards the edge. I came in the other direction. I started at the edge.
[00:05:47] And moved towards the cloud.
[00:05:49] Matt: [00:05:49] Okay. And when you say the edge, what are you're talking to the device.
[00:05:52]Satya: [00:05:52] yes. Devices. So let me walk you through the logic. very early in the evolution of [00:06:00] mobile computing and I'm talking 1993. Okay. At that time, the only mobile devices in existence were laptops. They were pretty hefty things. The way, the tongue, the first smartphone like device, we thought the Cilla component came out about 1997, 98.
[00:06:23] These were things like the Palm pilot. that was a compact device called the iPad. They will call it personal digital assistance. I PDX.
[00:06:32] Matt: [00:06:32] actually in that world. I built operating systems for personal digital assistance.
[00:06:37] Satya: [00:06:37] Fantastic. So you appreciate the point that making I'm talking about five years before these appeared. Okay. So the editor of IEEErdfdr computer. In 1993, sent me an email saying, you know what, you've been working on this brand new area called mobile computing. And many people ask me what the heck is new about this?
[00:06:59] You know, so [00:07:00] this laptop is like a shrunk down desktop. the bits arrived by wireless rather than on the wire. So what, why is there anything really fundamental about mobile computing? And so that forced me. I took on the challenge to write a short, just a one-page dthought piece on mobile computing thing.
[00:07:21] Why does different invites, would it be fundamentally different? So one of the points that I realized as I wrote that was the deep insight that has been true ever since. And it's true to this day, which is the very fact that devices are mobile. Requires them to be small, to be lightweight, to not get too hot, not burn the user to have I have long battery life.
[00:07:48] Right? All of the things that we expect of things that we carry well, where on her body. And I realized this is 1993, and this is a published [00:08:00] document that you can find. that this was always going to be with us. This is not going to change human beings. At least in the timescale of 10 years or 20 years are always going to be faced with this challenge and that what you were going to have to pay as the price of portability mobility was the compute capability in terms of cycles, memory.
[00:08:25] Storage, et cetera. In other words, you can think of this as a mobility premium that you have to pay to make the computing device small enough, light enough, et cetera, to be mobile. So that realization was the foundation that led eventually. Nearly, 20 years later, almost two, 15 years later to edge computing.
[00:08:49] And so he has to logic, sometime a few years later, 1997. we said, okay, given that mobile devices are always going to be [00:09:00] challenged in this way, how do we get substantial applications that require compute intensive processing? To run on a mobile device. How do you ever get that capability? The answer is to offload computation to the infrastructure.
[00:09:19] And so that capability, we were the first to demonstrate in 1997, again, in a published paper. And it demonstrated how speech recognition could be done on an AIPAC compact device using compute offload.
[00:09:36] Matt: [00:09:36] Did you call that edge computing?
[00:09:37]Satya: [00:09:37] no. At that time we did not demand that term. This was just mobile
[00:09:41] Matt: [00:09:41] compute? Offload. Was that phrase around?
[00:09:43] Satya: [00:09:43] In fact, the term I came up with in that paper, it doesn't have a name in a subsequent paper.
[00:09:49] I generalize this idea and called it cyber foraging. Okay. The vision was mobile devices of the future would have to find [00:10:00] compute cycles around them and leverage them. And of course, there are many papers would that name and the term offloading is also used for the same concept. So compute offloading is the term that is more commonly used today.
[00:10:15] But if you see the term cyber forging, it's exactly the same. So fast forward now to 2008, 2009, the timeframe of the paper that you mentioned by then cloud computing had become big. The iPhone had been introduced, applications such as Siri, the speech recognition application of Apple was in fact doing exactly what we had shown.
[00:10:41] Right. So, the thing that I realized immediately was that people's mental model was that you would have the mobile device, you would have the cloud and that's all they needed. And I realized that this was not going to [00:11:00] take us down a good path for the future. And the reason is because the underlying economic model of the cloud is based on consolidation giant exascale data centers, because it is the economies of scale that make it possible to deliver such low cost.
[00:11:24] Per unit of compute, that is the economics of cloud computing. And so I realized that if you only have a few large data centers in the world, on average, most of them are going to be far away, which means that the speed of light is going to take a while for your bits to get there and back. And from there, it's obvious why latency end to end latency.
[00:11:52] Becomes a showstopper for applications in which the offloading has to happen, in a, [00:12:00] very tight time budget things like augmented reality.
[00:12:06] Matt: [00:12:06] talk about that. And let's try to make this, simple for. The folks in my audience that maybe don't live in this world all day long, like you and I do. speed of light sounds really fast to me,
[00:12:18] Satya: [00:12:18] So,
[00:12:19] Matt: [00:12:19] 86,000 miles per second. How can that be slow?
[00:12:23] Satya: [00:12:23] so first of all, that's the speed of light and free space. It only travels two thirds that fast and fiber
[00:12:30] Matt: [00:12:30] That's right. It's
[00:12:31] Satya: [00:12:31] I, so that's already one third of the speed gone.
[00:12:35] Matt: [00:12:35] It still sounds fast to me.
[00:12:37] Satya: [00:12:37] It's still is very fast. No question
[00:12:39] Matt: [00:12:39] Two-thirds of almost infinite is hurting
[00:12:41] Satya: [00:12:41] It's still in fit. Yes. The problem to think about, the way to think about this is 186,000 miles per second, actually translates to roughly 300 kilometers per millisecond per second.
[00:12:56] Okay. 300, 300, 300 kilometers per millisecond. [00:13:00] So on fiber it's about 200 kilometers. So. Say roughly 150 miles. So thereabouts, anywhere within that space, you could get there in one millisecond. The problem is actual end to end networking involves many network hops, and it is never as fast as the speed of light.
[00:13:22] Typically, if we come within one order of magnitude of the speed of light, people are very happy. So something like, 10 millisecond or less latency would be considered extremely good. So if you did an end-to-end test, say from where I am right now to a typical cloud location, like, Amazon West. Yes West, which is located in Portland, Oregon at the speed of light, the wrong trip should only be 30 milliseconds.
[00:13:57] Okay. That's the width of the country in speed up light [00:14:00] terms. But if you actually measure it, the numbers you will get back the mean will typically be close to between 70 and a hundred milliseconds, depending on the day, you know, all kinds of factors, how heavily loaded it is. But that's only the mean, okay.
[00:14:19] If you plotted individual things, you would have observed that they worked out to have a distribution of which the mean is about between 70 and a hundred, but the tail is way up. It's around hundreds of milliseconds. Now, when it comes to user experience, People have learned that it's not the mean that matters.
[00:14:46] It is the tail. Human user experience is greatly overweighted by the few negative experience you may have one hour [00:15:00] on a zoom call, 55, six 58 minutes off. It may be excellent. But you will remember the two minutes that were miserable, and this is generally true for augmented reality. For all of these use cases, it seems to be an innate character.
[00:15:21] And in fact, the guys, Daniel Kahneman and Amos Tversky in the work that one Danny Kahneman, the Nobel prize actually commented on with this, that loss aversion is the term they use for it. Seems to be a deep characteristic of human users, both cognitively and reasoning. And also it has turned out to be an user experience.
[00:15:44] So for all these reasons, you have to keep both the mean and the tail of the distribution as short and small as possible. And as you have more hops, keeping the tail short becomes very [00:16:00] difficult.
[00:16:02] Matt: [00:16:02] Yes.
[00:16:02] Satya: [00:16:02] Okay. So it is a combination of both do exactly. So cloudlets edge computing saying, look, put the compute close by.
[00:16:12] Don't try to fool mother nature. Okay. Means I can get there typically in one hop, one wireless hop from a mobile device, plus maybe one more fiber link. That's it. So it's one hop a one plus one, and the total end to end latency is typically the airtime of the wireless communication, which for 4g LTE today is about 12 to 15 milliseconds.
[00:16:44] And then if it is a 5g, of course, it would be even less. If it's wifi would be less and then you have a millisecond or less of fiber on top of that's it? That is the networking portion. What this means in [00:17:00] practice is that F enact very tight time budget, total end to end latency. I may have a budget of 30 milliseconds in that time budget.
[00:17:11] I've only eaten away at one or two milliseconds, or in this example, 15 or 16 milliseconds. I still have quite a bit of that budget left to do real computing, to make a disc access, to do a face recognition. All the kinds of compute that you need to do to create the kind of applications of the future that we would love to have.
[00:17:33] Matt: [00:17:33] Yeah. And so that, that one or two hops from the edge of the wireless network, the base band unit, or the, you know, like. Cell tower, so to speak, in metaphorically, that one or two hops, that you want to have in say sub five milliseconds means that data center that's running that cloud.
[00:17:50] Let, whether it's a street side, cabinet or a building, needs to be within 30 to 50 kilometers.
[00:17:58] Satya: [00:17:58] it varies. It [00:18:00] depends on the very much. It depends very much on the details of the network interfaces, et cetera. But in general, if you have fiber connectivity, Orland connectivity. A cloudlet or two for a small city is reasonable. more will work fine. You don't need it. Exactly.
[00:18:18] Exactly. At the cell tower. The answer to the question, how, where is the edge is ideally you would like the edge to be far away. If you could have, if he could actually afford to do it. The reason is
[00:18:32] Matt: [00:18:32] get the economies that you mentioned earlier, the more
[00:18:34] Satya: [00:18:34] The economies of scale. And, the reason for the economy is also management complexity. if I do not have to go from cell tower to manage these distributed compute nodes, it's so much less expensive.
[00:18:51] Matt: [00:18:51] whether you're orchestrating workloads or sending humans out to replace parts, it's more complex. Yeah.
[00:18:56] Satya: [00:18:56] So centralization consolidation, which is the [00:19:00] name of the game for cloud computing. Has many advantages of economies of scale centralization, ease of administration dispersion, which is the essence of edge computing has two advantages, one which we talked about low-latency and to enter it and see the author, which is also important is improved resilience.
[00:19:23] The longer the connection to the end point, the more, the chances that it can break.
[00:19:31] Matt: [00:19:31] Right. Well, and you mentioned having two cloudlets in a municipal area. That means you potentially have fail over for your low latency workloads. And if you lose connectivity back to the main cloud or the large data center, you potentially have some fallback compute
[00:19:44] Satya: [00:19:44] Excellent. That's indeed. That is precisely correct. You'd have to engineer your applications to be able to leverage this capability. but indeed the presence of cloudlets and edge computing allows you [00:20:00] to take advantage of it.
[00:20:01] Matt: [00:20:01] Well, yeah, and software high availability with fail over, you know, look at Macy's and Kubernetes and technologies like that is becoming much more mainstream. And so you can see how those worlds may merge at some point where the same way we get. High availability inside a massive Exoscale data center, could be translated to multiple data centers, you know, instead of multiple racks in a single data center to multiple data centers in a single city.
[00:20:27] Satya: [00:20:27] exactly. And also keep in mind the division of cloudlets and edge computing also extends to end points that are mobile. So a vehicle, a car that you're driving. Or an autonomous vehicle of the future has onboard a club. Okay.
[00:20:46] Matt: [00:20:46] That's just like, so, so we started talking about cloud this, but we haven't really defined them. And you have this paper that you published. I think it was 2009. and. as I understand it, that part of the origin story of that was a meeting at Microsoft, that Victor Bahl [00:21:00] and some of his cohorts such as yourself, put together.
[00:21:02]can you tell me the origin of cloudlets and what cloud is?
[00:21:08] Satya: [00:21:08] sure. so, so sometime around 2008, the concerns that I just shared with you about latency, et cetera, were becoming very apparent to me. So, at the 2008 Moby SIS conference in Breckenridge, Colorado, I had lunch with Victor Bahl. And I said, Hey Victor, and many other things that we talked about, I said, Victor, you have something I've been thinking about.
[00:21:32] And I shared my concerns with them and he thought about it for a moment. And he said, you know what? You're a hundred percent right. let me think more about this. Completely separately at the same conference. Roy I want of Intel was also there and I shared these thoughts with him and he also was quite deeply, engrossed and thinking about the consequences of this.
[00:21:54]another person was, another coauthor on that paper was, from at and T research, [00:22:00] Ramon cus Russ. And so he was another one. So in some sense it was, and I forget when exactly this conference took place, but you can look up the dates. It was sometime like the summer or fall of 2008 a year later.
[00:22:14]and I forget, I think it was probably 2009 or so. No, it was still in 2008, probably 2008 or so. Victor happened to have a meeting in Seattle at Microsoft research. And I forget what the reason for the meeting was. It may have been a program committee on which a few of us were already scheduled to be there and was just a gratuitous opportunity to get together and maybe build on this earlier conversation.
[00:22:41] And we planned an hour long conversation, and I recall we used it the whole afternoon. And then spend part of the next day also talking about this and came away, convinced that this was going to be important. And at that meeting was where we said, look, cloud computing, the idea of [00:23:00] using, technology such as virtual machines, that was the technology of 2009 containers were just starting to be viable then.
[00:23:09]but all the machinery. That goes along with system administration, et cetera, all those are perfectly useful and is the only thing you need to change is to shrink it and put it closer. And so the notion of a cloud in a box or a data center in a box was really what we thought of as clouds, how big the box was, how heavy it might be, how much cheat it might generate.
[00:23:38] We deliberately left that. Open because we thought we don't quite know. And in fact, it might take many forms and indeed that has proved a cloudlet in an automobile or a cloudlet in an aircraft things of a hundred percent aircraft Boeing seven 37 will apply again. given that I haven't actually gotten on a plane in six months, the, you [00:24:00] know, the connectivity inside an aircraft is excellent.
[00:24:04] It's wifi connectivity. But between the aircraft and the brand, the connectivity is very bad. you know, you know this, if you this, right? I know that people are doing heroic efforts to try and improve it. But once again, I'm going back, you know, don't try to fool mother nature, an object moving at 500 or 600 miles per hour, 30,000 feet above the earth.
[00:24:32] Trying to get low-latency connectivity and high bandwidth connectivity to a ground-based station is extremely difficult. And if you can achieve it briefly hats off, but to sustain that reliably, economically, if a price point that people can actually afford. So architecturally a model that says there's a cloud on both the aircraft.
[00:24:58] It serves most of your [00:25:00] needs once in a rare while asynchronously that cloud left on your own behalf may communicate with the cloud on the ground, but you don't need any kind of synchronous communication. That vision is, I think, far more achievable from an engineering point
[00:25:15] Matt: [00:25:15] Well, and to some extent on an airplane that exists when it comes to video streaming.
[00:25:21] Satya: [00:25:21] indeed. And you know what? They don't serve food anymore on these planes. So all that space in the galley. It's available for servers,
[00:25:29] Matt: [00:25:29] That's great. That's great. So now we're going to be playing video games and, you know,
[00:25:32] Satya: [00:25:32] augmented reality. Exactly.
[00:25:36] Matt: [00:25:36] yeah. That's really interesting. So, so there's a couple of threads that we put pins in and I want to come back to, so, so I want to talk about latency again just a little bit, because I think one of the other.
[00:25:47] Elements of Lacy. So, so you used a lot of examples that involve humans. You have the, you know, how fast my phone responds and how much I'm willing to carry and things like that. But there's, another phenomenon, which [00:26:00] we broadly could call the internet of things, but it's these electronic devices, these machines that are out there.
[00:26:05] And, you know, as you very well much know, but again, some of our audience may not be familiar that the latency that. Machines are sensitive to and tolerant of is orders of magnitude different than humans. You know, humans operate in ones of seconds. I mean, fractions of seconds for like video and stuff, but essentially like it's not near the nanoseconds and microseconds that machines operate.
[00:26:26]how is this world of machines of ever-present sensors of all this effecting. The world of edge computing, from a latency, but from any perspective, how is that driving edge computing?
[00:26:43] Satya: [00:26:43] a very good point there, Matt. Let's break the problem up into two parts. Think of something very simple, a very simple IOT device, like a video camera. Okay. If you wanted [00:27:00] real time video analytics. Okay. You want to, have this video camera do a human recognition face recognition in real time. Okay. And let's say it sends you a text message when John.
[00:27:14] Is John in this, in this corner of the building, let's just make sample, right? you might want that. you can see how this requires continuous video from the video camera to some compute node that can do face recognition. And then the rest of the pipeline is pretty straightforward, but that continuous video feed.
[00:27:39] If it is HD video, typically the bandwidth required is 10 megabits per second. If it is part camera, indeed. And if it is a 4k video, you multiplied that by the right number eight K video. It's, you know, obviously we can even higher. So as resolution increases, the spend [00:28:00] with demand increases and that's per camera as you punch it up.
[00:28:03] So if I have. You know, a hundred video cameras distributed through some enterprise, and all you say is, Hey, find me where John is so that I can quickly say something to him because I really want to have at least a video face-to-face with them or something. you really need much higher bandwidth scalability.
[00:28:23] Sending all this data to the cloud would require a total inverse bandwidth. That is huge. So even if there's
[00:28:32] Matt: [00:28:32] just isn't even a fiber in the ground for it,
[00:28:35] Satya: [00:28:35] right, I mean, in special cases, you can always lay enough fiber. You can always solve this problem in a dedicated point solution, but the cost now becomes higher. So if you go this to do this economically at scale, you, this is not going to be scalable.
[00:28:52] So bandwidth, scalability, even without any latency. Is already a case for edge computing, you push the compute [00:29:00] close to the data, right? This is the aspect of edge computing that is closer to what Akamai originally proposed, but the differences in the
[00:29:11] Matt: [00:29:11] the network, you're isolating the bandwidth to that branch.
[00:29:14] Satya: [00:29:14] right, but in the case, the source of the data was the cloud. And you were trying to send it to it's the edge, that's the opposite direction. It's in the uplink direction.
[00:29:25] Matt: [00:29:25] the reverse CDN.
[00:29:26] Satya: [00:29:26] Isn't that reverse CDN. So cloudlets can be viewed as reverse CDN. That is the right way to think about that. So that's half the story.
[00:29:34] Matt: [00:29:34] difference, although Akamai certainly has the ability to run workloads and all the CDNs seem to be going that direction. And increasingly sophisticated, you know, like StackPath is allowing VMs and containers to be run on their, content delivery network. but so, so it seems there's a fundamentally new dimension, of edge computing of cloudlets and that is, it's not just.
[00:29:57] Capturing caching storing [00:30:00] the data. It is processing
[00:30:02] Satya: [00:30:02] yes, that is precisely the point. Once you have a CDN that actually allows general purpose processing by definition, you have embraced cloudlets. You may still call it a CDN, but you know, it is a cloud. So you, in 2009, when we described cloudlets CDNs, essentially data staging nodes, where content that had been produced in the cloud at a website could be pre-positioned close to consuming sites and something else.
[00:30:37] Also remember it aloud. Real independent customization for each part of the edge. So advertising content, for example, for Pittsburgh, with the New York times, it could be different from what it is for city of Cincinnati. So they, there was a little bit of customization happening at the edge through the edge mark [00:31:00] of language that Akamai might pioneer.
[00:31:02] For us all very constraint, and the notion of taking an Amazon like data center and moving it to the edge is closer to our vision of what cloud would start. And in fact, of course, now that Amazon has partnered with Verizon to do exactly this, you know, we are exactly there. but I, yeah, sorry,
[00:31:24] Matt: [00:31:24] Oh, I was just saying that I've been saying that for years that, you know, most people don't realize most people aren't in the industry. Don't realize that if you want to precision any provision in ECE, two instance in the United States, you have exactly two choices.
[00:31:36] U S Western us', but we're quickly approaching a world where you have San Francisco, East and Chicago, West Pittsburgh, North, where you can deploy these cloud instances.
[00:31:47] Satya: [00:31:47] and those are the cloudlets that we have in mind.
[00:31:49] Matt: [00:31:49] Yeah. that's neat how this world is converging,
[00:31:51] Satya: [00:31:51] Let me continue though, because I've told you the story of IOT that we started out with something like a video camera, but it could also be [00:32:00] a vehicle engine. It could be any of the, any sources, rich data fits this description. But now suppose the IOT device is a drone and on board, the drone is a video camera. And in real time, the groan is observing what's below it. Now today, most drones capture the data. They store it on storage device, on board, the drone. And when it comes back, you upload it into the cloud and
[00:32:34] Matt: [00:32:34] Yeah. a human takes the USB card out of the drone and walks into Starbucks and uploads the video. That's how it happens.
[00:32:43] Satya: [00:32:43] So, so you can improve this very easily by saying, let me transmit continuously even while it's being captured. Okay. If I had a 4g LTE connection and my bandwidth was sufficient, you could imagine doing this in real time, but it's [00:33:00] expensive. And if I had a swarm of drones, this would not be scalable.
[00:33:05] Right now, come see. Interesting question. Suppose after you process the data, you discover that halfway through the flight, there was something really important that seemed to be present. And you would like to take a closer look. It takes you half an hour or more to discover this because the wait for the drone to come back.
[00:33:31] And then you have to send the drone out again to take a closer look instead of matching. You could do this analytics in real time during the flight.
[00:33:45] Matt: [00:33:45] Especially if the life safety issue like someone's life is at risk, you don't want to wait a half hour
[00:33:50] Satya: [00:33:50] Yeah, I mean, if this is after a hurricane or earthquake,
[00:33:53] Matt: [00:33:53] fires
[00:33:54] Satya: [00:33:54] is a rescue right prize in California, this is the rescue mission. You have a, [00:34:00] possibly a person, a survivor, you know, that half hour, the guy might be dead, right? So, so it really can be a matter of life and death. if it is a military target that half hour, may make a big difference.
[00:34:15] So imagine in real time being able to do this analytics and before the drone has moved very far, maybe no more than a few feet. You're able, based on what was seen recently to say drop down to low altitude, to take a closer look, right? That is actuation based on real time sensor data.
[00:34:45] Matt: [00:34:45] Yeah. You know, and there's a, there's an interesting conversation in the edge world today, which is. you know, a lot of the Silicon manufacturers are getting very clever at building in like image recognition into dedicated Silicon that can be [00:35:00] processed on a drone to some extent, but you raise a couple of good points.
[00:35:02] So one is the fact that, well, if you're trying to interpret from a swarm, like. Those have to be communicating to each other or to a central locations of you can aggregate and scan all that data simultaneously. But there's another really interesting part to this, which it comes back to your original, you know, 1990s paper.
[00:35:19] So a drone is a mobile device in the sense that you're describing it. And it has. The same car. So an autonomous car you can imagine, well, okay, it's generating a bunch of power. You could stick a small data center in the car, which is actually a fact that what they do today, but you can't on a drone because it doesn't have enough flight power.
[00:35:35] It doesn't have enough battery. you know, rather probably invest that energy into the RF signal instead of local processing. So you're back to those constraints that you identified back in the nineties, and it really does speak to the need for cloudlet.
[00:35:49] Satya: [00:35:49] a flight is very interesting because it is like human, wearable devices. Right. Light lightweight does matter. And keep in mind. One other [00:36:00] thing today, drones are. quite easy to fly in rural areas or in test ranges. There's a huge amount of interest in making these drones work and deliver value in urban areas.
[00:36:18]but in fact, just last month, there was the national Academy of science report on drones, which exactly makes this point that there's a huge pent up demand to have drone services in urban areas. But there are fundamental issues of safety, right? A drone falling on your head. If it is large and heavy, is going to cause a lot of damage.
[00:36:42] So from a safety point of view and regulatory point of view, small lightweight, drones are fundamentally intrinsically safer. Right. I mean, even if it goes bad and falls on you, the damage is less [00:37:00] likely than a bigger The smaller and the lighter weight. The drone, all of the problems that we've been talking about are exacerbated.
[00:37:11] Right? You can still carry an eight K or 16 K video camera. No problem. You can still carry a 128 gigabytes of flash, no problem and it's small and lightweight, but now it cannot carry anything much heavier than maybe a smartphone. That's the amount of compute I can carry it. So in our work on drones and edge computing, that is the guideline that we have been using, which is pretend that the drone that you can use cannot carry
[00:37:42] compute larger than the smartphone. Okay. And now ask the question. What can I do in real time with this? And it's very interesting. What you can do is you don't have to sort of send all the data to the ground. Even the [00:38:00] compute onboard a smartphone, especially if it has chips like the new SOC on the Apple chips.
[00:38:06] Right. they, that they can do is non-trivial amount of processing. What you can do is use the offload concept, offloading to the edge, but use the onboard processing to do early discard. In other words, you're able to scan incoming data and use the local compute to be pretty confident that there's nothing interesting here.
[00:38:33] This is not worth transmitting.
[00:38:35] Matt: [00:38:35] Yeah. And you could even store the raw data locally for
[00:38:39] Satya: [00:38:39] Later processing.
[00:38:40] Matt: [00:38:40] were afraid that maybe you might lose something.
[00:38:42] Satya: [00:38:42] Right? So you can, you don't need to lose any data. Now it becomes, I use the local processing to prioritize the incoming data and decide which subset of it is valuable enough to transmit right now, rather than defer to the [00:39:00] later time. And then on ground at an edge computing node on a cloudlet I can have a GPU, a whole bunch, of course, you know, power.
[00:39:11] That's not an issue. I can revalidate the compute that you did on the crumb to verify a high accuracy, whether indeed you are correct. And what we have also looked at is dynamically constructing, doing machine learning in real time on the cloudlet. And transmit up to the grown and improved machine learning model so that as it is flying, it's processing capability is being tuned to the terrain below it. Right? So you're the, this is true edge computing sort of on steroids.
[00:39:53] Matt: [00:39:53] Yeah, that's really interesting. So, so we're talking about mobile device devices and wireless. and I don't think we [00:40:00] can touch that topic without also talking about 5g. And, I recently watched, one of your, one of your keynotes, the OpenStack keynote from a few years ago. And, one of the things you mentioned there, and I probably won't get it exactly right, but I'm gonna paraphrase you basically say it's useless to deploy 5g without edge computing.
[00:40:15] Can you explain what you meant by that notion?
[00:40:19] Satya: [00:40:19] I've meant exactly what I said, you know, I don't know exactly what words
[00:40:22] Matt: [00:40:22] to that. I mean,
[00:40:24] Satya: [00:40:24] it was. It was pretty blunt, you know, I'm sure, you know, I have many very good colleagues in the telcos, right. They've been working closely with me, Vodafone and Deutsche Telekom show all of them, French when I said this. but it's the truth.
[00:40:38] I mean, the truth is 5g is only going to improve your last mile quality. That's it
[00:40:47] Matt: [00:40:47] Yeah. from the tower of the small cell to the drone or the phone or the car?
[00:40:51] Satya: [00:40:51] No question. That's incredibly valuable. If you can reduce 15 milliseconds of latency down to one millisecond. Fantastic. That is [00:41:00] great. But you know what, if it takes me a hundred milliseconds to get to Amazon West and you've reduced 15 to one, I still have 84, 85 milliseconds. We'll have to deal with.
[00:41:11] Matt: [00:41:11] Yeah. I had a friend of mine saying it's like painting the inside of your car, trunk gold.
[00:41:17] Satya: [00:41:17] That's a very good metaphor.
[00:41:18]Matt: [00:41:18] it's impressive, but it's not doing anybody any good.
[00:41:22] Satya: [00:41:22] Now, on the other hand, if you push the compute to the cell tower or close to it, as we have been discussing, now, that improvement makes a big difference. You know, this is old, very old, these things, this line of thinking, in fact, it's so old that there's actually a name for it in computing called Amdahl's law.
[00:41:42] Matt: [00:41:42] Okay. I don't know this.
[00:41:43] Satya: [00:41:43] Oh, it basically says in a computing system, if you have an end to end pipeline and, you know, as you can know, from the name Amdahl, it arose in the day. So the mainframe, because I am doll was the designer of the IBM three 60 architecture. Right. And it had to do with [00:42:00] the fact that how much improvement that faster computer will give you depends on the balance between processing and storage
[00:42:10] access. If most of the time it goes to storage access, making the computer faster. Doesn't help. It's the same idea except instead of storage, replace it with the network now, right? If your network, the computers written, I know the 5g part, the storage is now replaced by the rest of the network. Did the distant cloud.
[00:42:37] And it's sustained argument that says if you only improve one part of it and the rest of it dominates, you have not won significantly.
[00:42:47] Matt: [00:42:47] Yeah. and I think that the, one of the things that makes us fundamentally different from some of the other sort of arguments is that you really can't escape the physics. I mean, I mean, maybe quantum [00:43:00] computing at some point will create this, you know, magical entanglement version of cloud computing, but, you know, certainly for the foreseeable future, The only way to make things faster is to reduce the distance and the number of intermediaries.
[00:43:16] Satya: [00:43:16] exactly. Yeah. I'm not betting my future on a quantum computer. you know, I'm sure it is phenomenally important for the future, but not the near, not the short term.
[00:43:28] Matt: [00:43:28] Yeah, it'd be a long time before we're handing out, SDKs for quantum computing to the general it developer.
[00:43:35] Satya: [00:43:35] Okay, indeed. And even then let's be honest here, there are going to be specific applications such as factorization. You know, the things like breaking in production keys. It is for those tasks, but for the kind of tasks that we are talking about, augmented reality drone control, the precision and accuracy of conventional computing.
[00:44:00] [00:43:59] Rather than the probabilistic model of quantum computing is actually crucial.
[00:44:05] Matt: [00:44:05] Yeah, there's an interesting, and I don't know if you've heard this, but an interesting discussion. That's emerged very recently in my edge conversations. and it goes something like, like this it's. So, so let's say three to five years ago, the most discussion of edge computing was around latency.
[00:44:23] Right. We need to reduce the latency and this is why we need to move closer to the edge. And there's other reasons, you know, data sovereignty. And, but essentially it was around latency. Now there's this, this meme I'll call it, which is, well, we were wrong about latency, that the most important thing isn't latency, at least not sub 30 milliseconds I see some of this phenomenon in my business. So, so I see that there is more demand for sub 100 milliseconds, maybe, you know, 30 ish, but I don't see a lot [00:45:00] of demand for cloud, like low latency over a last mile network, in the. Sub five milliseconds, sub 10 millisecond. But I also believe that is just a function of where we are today.
[00:45:12] And I would be interested in your opinion on that, which I'm pretty sure is going to be. There's lots of applications that need that super low latency. And I'd like to know your opinion and also like what kind of use cases are going to require that are interesting and going to be important to people in society.
[00:45:29] Satya: [00:45:29] First, I agree with you completely. The basic premise that you've sketched is exactly correct. How much end to end latency is acceptable is very much a function of the application. But it's a two way street. The applications that get written depend on what today's technology can offer
[00:45:51] Matt: [00:45:51] is such a great way to describe it. Yes, they are related to each other.
[00:45:55] Satya: [00:45:55] Okay. If the application's demand gets [00:46:00] too far ahead of what the technology can offer, then the application will die. Because it's not viable, you know, if it's a startup company, it doesn't have a big enough market.
[00:46:12] Matt: [00:46:12] If you need to play a billion dollars for the cloud infrastructure in order to run your application, you're probably should do something else.
[00:46:18] Satya: [00:46:18] right. So this is a recurring theme in technology, right? So it's crucial that technology and the applications that leverage the technology kind of have this type dance. Where, you know, one gets a little ahead of the other briefly and the other catches up. It pushes the other further in that direction.
[00:46:45] The other catches up and so on. And so there's almost like a virtuous cycle where each influences the other.
[00:46:54]Matt: [00:46:54] I like that very optimistic way of looking at, but also very realistic of description of why we are where we are [00:47:00] today.
[00:47:00] Satya: [00:47:00] Right. So I have described in papers that I've written the notion of edge native applications,
[00:47:08] Matt: [00:47:08] I'm
[00:47:08] Satya: [00:47:08] which are application, right, which are applications that are so dependent on the edge that they simply do not work without edge computing.
[00:47:17] Matt: [00:47:17] So let's talk through some of those.
[00:47:19] Satya: [00:47:19] Certainly. an example of that would be any kind of augmented reality application that.
[00:47:26] Involves offloading to a cloud lab, but it's also, sufficiently visually immersive that in fact, more than a 20 or 25 millisecond latency is going to cause extreme discomfort to use it.
[00:47:45] Matt: [00:47:45] So outside of gaming, what's an example of a legitimate use case
[00:47:51] Satya: [00:47:51] okay. Let me give you one. It's actually a class of applications that we've been working on that we call wearable cognitive assistance. Okay. [00:48:00] So it is a kind of AR kind of augmented reality. The way you should think about it is combining augmented reality and AI.
[00:48:11] Matt: [00:48:11] Okay.
[00:48:11] Satya: [00:48:11] Okay. So let me give you the simplest example.
[00:48:15] I can think of, imagine you are making a cooking, you know, making some dish. one of the earliest applications for Google glass was the recipe app. So you could say next, and it would go to the next screen and show you the instructions to do next. You could say next time do that. And, you know, this is Google glass came out in 2013.
[00:48:42] So this is a seven year old technology. Compare that with what a good friend would do if he or she was standing next to you. If the instruction said heat, whatever, saute the onions until they're sizzling or something, [00:49:00] your friend would say, wait, they're not sizzling yet. You haven't cooked them long enough. Right? Sensing the real world, testing, and determining whether it meets the criteria to go on to the next
[00:49:17] Matt: [00:49:17] step
[00:49:17]
[00:49:17] that's a great example because I have a nanny and she cooks a lot of our meals. My kids are obviously school from home now. And so I need a lot of help around the house. I'm a single dad and, she's a very competent cook, but she doesn't have as much experience with the, as you say, there's like.
[00:49:33] Experience and observation elements of cooking, like when are the onions caramelized? Right. And yes, I can sit there and I can say, okay, this is what you're looking for. But even for me, I have to think hard to verbalize, you know, the heuristics that I've learned over, you know, 20 years of food experience.
[00:49:51]but even more important than like the cooking example, which I think is a funny and accurate example. Like imagine somebody fixing, You know, a jet engine [00:50:00] or a nuclear reactor, or I, you know, I can, or, you know, the trope in movies where the, you know, both the pilots go unconscious and somebody has to land the plane.
[00:50:10] You know, there's all these like really potentially dangerous or super technical tasks that could be done much more efficiently, accurately, safely, if you had some assistance. And you're right. Having an AI AR assistance that could. Interpret. So that's really interesting, like actually processing the image data.
[00:50:30] So this is your 4k cameras uploading and maybe stereoscopic. So you've got a ton of data. You got to process that in real time because it doesn't do you any good if you're not updating the screen fast enough? Cause you might make the person sick or you might not give them the information fast enough.
[00:50:44] Yeah, that's a really, so why
[00:50:45] Satya: [00:50:45] and one of the things the software needs to know that you're trying to make a cheese souflee.
[00:50:51] Matt: [00:50:51] Yes.
[00:50:51] Satya: [00:50:51] is not general AI. This is very targeted. Now I suggest to you. That [00:51:00] you and the audience listening to this podcast every day of their life, use an example of such an application, except you don't think of it this way.
[00:51:08] Matt: [00:51:08] Okay, what is it?
[00:51:09] Satya: [00:51:09] It's GPS navigation. Yeah.
[00:51:12] Matt: [00:51:12] Okay. How so? That's interesting. Yeah.
[00:51:14] Satya: [00:51:14] To what the world was like before you had GPS navigation, you had paper maps.
[00:51:19] Matt: [00:51:19] sure. It was life
[00:51:20]Satya: [00:51:20] the first question was, where am I on this map? Next question is, how do I get to where I want. Right. I mean, it was, you needed skills in map reading, and if your marriage survived a
[00:51:33] Matt: [00:51:33] needed a good map.
[00:51:34] Satya: [00:51:34] right. and if you know, your marriage survived a trip to an unknown city, using paper maps, it was a sign of a good marriage, right.
[00:51:43]so compare that with what we have today. You're given step by step directions. If you don't follow, like it says, take this exit and you don't it. If it detects the fact and gives you an, a correction, how does it [00:52:00] work? First? It knows the task very deeply. It has the roots, et cetera. Second, it uses sensing, in this case, GPS based location.
[00:52:11] And it's a system that depends on. A smart human being. It only has to say, take the next exit, unlike a robot, right. Where you have to plan, you have the movement of the autonomous car and all the details. It's have to tell you, I assume you are competent driver and the rest of it follows. So generalize that metaphor where I'm now doing, not just GPS sensing, but vision-based sensing.
[00:52:41] I am. I have knowledge of the task, not just for navigation, but for things like cooking or repairing of complicated piece of machinery that is broken down or a medical procedure being done by somebody out of a med in an emergency, [00:53:00] because there's no way to transport the patient. Okay. If you use whatever medical skill is available.
[00:53:06] And so you can imagine something like this being useful. In a developing country in rural areas in the military. Right? All of these, the ability to deliver just in time instruction is I think very powerful. And that is another example, which is not just entertainment. It is. It is everyday life.
[00:53:34] Yes.
[00:53:36] Matt: [00:53:36] So
[00:53:37] Satya: [00:53:37] Yeah,
[00:53:39] Matt: [00:53:39] you've you've been talking about this for 30 years, at least it seems, it's not here yet. A lot of these things aren't here yet. Is that a long time? Why isn't here yet? When's it going to be here? What do you say to the skeptic?
[00:53:54] Satya: [00:53:54] well, here's what I would say. Nature took a billion years to evolve us. [00:54:00] We've come pretty far in 30
[00:54:01] Matt: [00:54:01] You're a patient man.
[00:54:04] Satya: [00:54:04] Okay.
[00:54:05] Matt: [00:54:05] of a billion years, it's meaningless, but in all seriousness and all serious. I mean, I clearly these things take time and I don't mean to make light of all the advances that have happened, but it feels like we've been on the cusp of augmented reality of like, ubiquitous, augmented reality.
[00:54:22] We've been on the cusp of that. For the last five years, people have been talking about it and heavy part of discussion. Now it's kind of falling out of favor. Now. I think it's an amazing technology. I think it's going to revolutionize the way we interact with our world, but it's, it has fallen out of favor.
[00:54:35] I mean, when, so when do you foresee that? in a, an advanced nation, like the United States, we will have, Relatively common. and by that definition, I'd say maybe not a lot of consumer applications, but like there are people who go to work and on a regular basis, Don pair of AR something or other, when do it, that's connected to a network that's doing cloudlets.
[00:54:58] When do you see, when do you foresee that [00:55:00] happening?
[00:55:00]Satya: [00:55:00] so the first use cases are always going to be the high value use cases. The high value use cases, either in industrial troubleshooting. And here's a number. I mean, I realized, as I mentioned earlier in our conversation, I haven't been flying for four months is the longest period in 40 years that have not been flying, but when I was flying a lot
[00:55:26] Matt: [00:55:26] I imagine you have some serious airline points.
[00:55:28] Satya: [00:55:28] yeah.
[00:55:28] See this. but, and I'm sure you have too, but the point is that all of us take flights, you know, running an airline is a very sophisticated. Real time operation, right? I mean, hats off, we are annoyed when the flights are delayed, right? We're very annoyed when the flights are delayed, but you know what kudos to those guys to get those flights out on time.
[00:55:50] Matt: [00:55:50] Right.
[00:55:51] Satya: [00:55:51] Here's a number that is stunning. when a fully loaded aircraft of Boeing seven 37 is not enabled to push back from the gate [00:56:00] because of some trivial mechanical problem. Right. The estimated cost per minute was $600, five years ago, a 20 minute delay for the mechanic for the mechanic to come from the other end of the airfield to fix whatever problem it is, do the math, you know, $12,000.
[00:56:28] If I through the technology we are talking about. Could get an airline employee who's already here, right. He just is not an expert in this particular piece of equipment.
[00:56:40] Matt: [00:56:40] the, as much about the breaks as the specialists who can dial in so to
[00:56:44] Satya: [00:56:44] Right. So to speak, if I can help somebody who's already here to unwedge whatever's currently stuck. I have saved the savings arent trivial to telecommute directly quantifiable number to the bottom line. [00:57:00] So my belief is it is going to be an industrial troubleshooting factory automation.
[00:57:06] And these settings that this technology is going to find the lowest hanging fruit and be the most valuable and like everything else that's happened repeatedly in computing. Once you pay the cost of recouping of repute grouping the cost of the initial investment. Then the marginal cost of applying it to other areas of the applications becomes cheaper and cheaper.
[00:57:32] Matt: [00:57:32] Yeah. the $12,000, every 20 minutes for an airplane pays for the tractor version.
[00:57:38] Satya: [00:57:38] right. I mean, so, so that is the path that I see. entertainment is fine, but I don't believe it is entertainment. That is going to be the driver of this kind of technology. There's going to be situations of the kind we just described
[00:57:52] Matt: [00:57:52] So what, so where, so again, so the timeframe,
[00:57:54] Satya: [00:57:54] I am confident that in five years in the five-year timeframe, there will be use [00:58:00] cases of this technology in industrial troubleshooting.
[00:58:04] Matt: [00:58:04] 2025. Yeah, I totally can see that. I could see it happening faster
[00:58:08]Satya: [00:58:08] and, you know, the technology doesn't have to be perfect what we're working on. So we have extensive experience. We will proofs of concept. We've demonstrated this. we are also demonstrating how, If the software gets out of its depth , you can, the software itself can escalate the matter to a distant expert, the human expert, right.
[00:58:33] That mechanic. So that mechanic was at the other. Yes, well, but the point is that even if you don't, your AI is of limited capability, right? effectively what I'm doing is creating a zoom call. With the expert, but the difference is this, one expert now can help maybe 20 workers, because most of the time he doesn't have to [00:59:00] directly work with them.
[00:59:01] He only has to handle escalations, right? Just as today, the way we are coming at this is from the other direction. When you call in for help on anything, you first boat through this jungle of AI that tries to do triage tries to try to get you off just with an answer so that you never actually talk to a human.
[00:59:26] And then finally, if you have passed through all the hoops, you finally get to talk to the human here. You're doing this idea, but in the inverse direction, at least the software can help you. You never need to bother the human expert. It's in the rare case of exception. that you need to contact them.
[00:59:46] Matt: [00:59:46] And as as they, the automated assistance becomes better, you need fewer of those experts and yeah,
[00:59:52] the
[00:59:52] Satya: [00:59:52] so that one expert is now being scaled out better. Right? That's the way to think of, think about this [01:00:00] technology as a way to scale out expertise by definition experts in any field are rare. Otherwise they wouldn't be an expert, right?
[01:00:10]Matt: [01:00:10] or jet engine repair.
[01:00:11] Satya: [01:00:11] Right. And so that is always going to be scarce using this technology to scale out that expertise is I think a fundamentally powerful metaphor.
[01:00:22] Matt: [01:00:22] Yeah, I agree. That's really interesting. I want to shift gears a little, cause there's, some research that you've worked on that I think is extremely important. I know you're passionate about, I don't think it gets enough air time. and it's your concept of a privacy firewall? Can you.
[01:00:37] Tell the audience, what a privacy firewall is, why it's important and how you envision us using it.
[01:00:44] Satya: [01:00:44] that's an excellent point. And thank you for bringing that to the attention of your audience. one of the big concerns with IOT, especially in consumer deployments, as opposed to industrial deployments is concern about privacy. So let me give you a simple example. [01:01:00] You can go to nest and buy devices.
[01:01:03] You can buy from other companies, things like water meters, which report, right? the waterflow on a fine time green basis that you can have a video camera that looks outside your door. It looks at the back of your, VR where, you know, people deploy these. The question is today, all the data raw data goes to the owner of that off that software chip goes to whoever has provided you that device in the cloud.
[01:01:35] The problem is that embedded in it. When you provide this data at fine granularity is information that is truly quite concerning to many people. If they only knew. So you may ask, okay, how much privacy is being lost by the fact that I'm my water consumption is being reported every 60 seconds. Oh, come on. Tell me [01:02:00] what does it matter?
[01:02:00] Well, he has hot people. You'll have shown in the published papers, not our work, but other people's those work that you can do, machine learning on that data to discover precisely when the dishwasher turned on, when the washing machine turned on, when somebody in the house flushed at the com mode, et cetera.
[01:02:21] Matt: [01:02:21] Cause they all have different pressure signatures.
[01:02:23]Satya: [01:02:23] you can discover the signatures. Well, it's only a short step from there to be able to infer, Hey, Matt looks like looking at your water consumption. Somebody had diarrhea this morning and you know, family, or somebody looking at the electricity consumption says, Hey, Matt, it looks like you had an overnight guest. Right now, you know, once you start realizing that these are the kinds of inferences that are possible, people start becoming very concerned about [01:03:00] privacy exposure. So the approach that we have been exploring, it's work that we began with Nigel Davis, of Lancaster university, and also Nina Taft at Google.
[01:03:12]Google is very cares, a lot about privacy, and it was great to be working with her on this, is the notion of using cloudlets to be the first point of contact for your raw sensor data and running on the cloud, what has storage? So your raw data is preserved for quite a while, maybe a month, maybe six months.
[01:03:33] And what goes to the cloud? It's a much redacted version of that data, which is less privacy exposure, sensitive. And if the cloud software senses a need for more detail, they can contact you and obtain explicit permission for more fine-grained data. So for example,
[01:03:59] Matt: [01:03:59] in your house. [01:04:00] Could we look at the last week's worth of water usage and we promise to only look at it for that and then discard it.
[01:04:05] Satya: [01:04:05] exactly. And now I'm able to make us mine, my cloud. What has the data? And so you are able in this way, or for example, the data that I normally send has all faces blacked out the video, and maybe I only sent it that one frame every 60 seconds, rather than 30 frames, a second. But if a theft is detected or if the police, for example, have reason to believe that you'll video may have a suspect and they have gone through the appropriate procedure, like get a warrant for, maybe they ask you your permission.
[01:04:43] You may be willing to share, with the face not blacked out. Right? So in other words, giving the owner. Exactly. This is the point that is allowing the owner of the private data. [01:05:00] You are the homeowner, the sensor data is in your home. The video camera is looking at your backyard. You should be able to control who gets to infer details about your life based on that data.
[01:05:16] So this concept we'd refer to as privacy mediators. And the notion is that edge computing is really an ideal way to implement this because you're pushing the processing close to the data. And equally importantly, the cloudlet is not owned by the cloud provider. It's either owned by you in your home or on your behalf, a third-party.
[01:05:46] So for example, it might be Verizon. Or a T-Mobile as somebody who unknown behalf runs a platelet running the privacy mediator, but they're not beholden to the application, which [01:06:00] is the IOT application owner.
[01:06:02] Matt: [01:06:02] Is there anybody in the ecosystem that's trying to commercialize a privacy mediator.
[01:06:08]Satya: [01:06:08] I am not privy to any knowledge of a specific individual. I can only infer that. the number of people who have expressed interest in this work, but that's a very general answer. not a specific
[01:06:23] Matt: [01:06:23] Yeah. And that's a pretty exciting, I think that's really exciting. And I think that is, it's one of these things that the privacy portion of this, I don't think it's talked about enough. And, this seems like a really good solution to at
[01:06:35] Satya: [01:06:35] Yeah. The question is always how to monetize this. Will you pay you the homeowner pay for this because the recipient of the data actually does not want this. Right. So I would say that this is not a technological issue. It's a monetization issue. And the person who figures
[01:06:58] Matt: [01:06:58] issue actually at some point. [01:07:00]
[01:07:00] Satya: [01:07:00] that would be heavyweight. Right.
[01:07:02] I mean, to the extent.
[01:07:04]Matt: [01:07:04] consumer privacy laws, they didn't come around because the commercial interests
[01:07:09]Satya: [01:07:09] indeed. Yes, we agree. But to the extent that these things are driven by value delivered to the end-user, they are likely to be more agile and better solutions than regulatory solutions.
[01:07:22] Matt: [01:07:22] fair enough.
[01:07:23] Satya: [01:07:23] Right? I mean, I would like regulatory solutions to be a last resort.
[01:07:27] Matt: [01:07:27] Well, you know, Apple has positioned themselves, around creating value around privacy. And I think increasingly, at least in my peer group, which, you know, granted it's the star Trek convention, but my peer group, they value that, like, I know friends who have a bias towards Apple products because Apple goes out of their way to at least advocate for privacy, if not actually do it, but certainly feels like they do it.
[01:07:51] Right.
[01:07:52] Satya: [01:07:52] I would agree with you. I think Apple is to be admired for their emphasis on privacy, but it's only one of many companies
[01:08:00] [01:08:00] Matt: [01:08:00] And it's the only one. Yeah. And then a lot of the companies get my information. Okay. Two, two final, quick questions. so, so the first one is more exploratory it's so if you look out over like five years when stuff's going to happen, right, and you kind of look at the dominoes that need to topple to get us there, if you had a super power, which is you could go push one of those dominoes, you could nudge it.
[01:08:19] You could make something that's. Slowing us down go faster. What, which dominant would you push on?
[01:08:24] Satya: [01:08:24] I would push the venture capital domino. Okay. Yes. So I would say that what we need are the ultimate beneficiaries of edge computing to recognize. That the return on investment has to be a long-term payoff, incentivizing the creation of edge native applications. In other words, it, if you have a startup company with an [01:09:00] idea, it's not going to be a hockey stick growth curve because the payout is going to be over a longer
[01:09:06] period of time, today's venture capital is not interested in. But somebody whose future is bet on the emergence of edge computing should take the view that investment in that kind of startup is valuable. Even if it doesn't have a hockey stick growth curve. And the reason is because you are making, creating longterm demand
[01:09:33] for the core product, which you're creating, which is edge computing. So I would say that either collectively, and this is a point I've made publicly in a keynote top class, Jen London, that, the ultimate beneficiaries of edge computing should look at this very carefully to help strategically invest, make strategic investments.
[01:09:57] That incentivize the creation of edge native [01:10:00] applications and applied to them a different criteria, success criteria on what is traditionally applied for by venture capital. I think that is
[01:10:12] Matt: [01:10:12] call to action, and I hope the investors and venture capitalists in our audience, contemplate his suggestion there. so last question is, if we want to find you online, how can we go about doing that?
[01:10:24] Satya: [01:10:24] Oh, wait, you can send me email. That's the easiest way to do it. and my email is public Satya at CS for computer science, that CMU for Carnegie Mellon, university dot edu. And, I is perhaps the easiest way to find me.
[01:10:39] Matt: [01:10:39] Great. Great. Well, thank you. So yeah, this has been such an enjoyable conversation. We, there are so many more parts of this discussion we could have touched on. I really appreciate you being so generous with your time and, maybe we'll do this again.
[01:10:51] Satya: [01:10:51] Matt. Thank you very much for having me, for giving me the opportunity to share my thoughts with you. It's just been a pleasure.
[01:10:57] Matt: [01:10:57] Awesome.
[01:10:58]