Getting ready for VMworld with AI deep dive with 14-year employee

Artificial intelligence is changing literally every part of human life. I mean, today I came from my house in Campbell, California, to VMware's headquarters here in Palo Alto, where I'm taking this video. And my lunch was prepared by my oven, June oven that has an Nvidia card in it and a camera. And you put a piece of toast in it, says, oh, that's toast, and cooks it to the right temperature. On the way here, Spotify used an AI to pick the music on Discovery Weekly's playlist. And my Tesla drove me here. 99 of the driving on the way here, I didn't touch the steering wheel or stop at stop lights. And AI is starting to come along with the autonomous cars, which, by the way, are done right across the street here, Tesla's R D factory. But VMware saw my latest book on spatial computing that I co wrote with Irina Cronin, and spatial computing's augmented reality glasses, robots, and autonomous cars, all working to change our lives deeply. And we wrote about seven industries that are going to radically change over the next decade because of the. And AI runs all that. So VMware invited me to talk a little bit about that today and interview one of its AI experts. And let's dive into that interview right now. The reason we're doing it next week, bmworld, 100,000 people are gonna be online learning about the latest in technology that runs the world's business from this great company, VMware. So it's an honor being on the call with you, Eric. I was thinking just about how deeply our lives have already changed because of AI. I met my lunch today was made by a June oven, which has an Nvidia card in the oven and the camera. And you put a piece of toast in the oven and it says, oh, that's toast. And it just starts cooking it. Right?

Image recognition, that work area.

Yeah, it's awesome, right? And on the way here, I was listening to Spotify's Discovery Discover Weekly playlist, and I met the guy who built the AI for that. So I know that's AI that's picking my music. And of course, I'm sitting in a Tesla. And Tesla is a block. There are an Elab here whose block is right across the street from. I'm sitting in front of VMware in Palo Alto, California, by the way. I came and trying to get some pictures and stuff for the posts I'm going to do with you, with you all over the next few days to get ready for VM World, which is going to be Virtualized for the first time because of COVID this year. Last year you had tens of thousands of people. And I don't know if we should reveal the number, but I hear it's pretty close to 100,000 attendees right now signed up for.

I know for a fact we can, we can talk about it. I mentioned in my podcast that we, we exceeded the hundred thousand mark right now and we're, you know, still a week away, so really good numbers coming in.

And I'm sitting, you know, to continue the theme of how AI has changed your life. I'm seeing this Tesla and itself drove me 99% from Campbell to here. It was self driving. It stopped at stoplights, it change lanes automatically, all because AI. Right, right. And when I, you know, it's funny walking down the road here. I remember visiting here 13 years ago or something like that when I was a journalist and blogger and I met with the founders of VMware and, and got a look at some of their early products and where they were going. And now, of course, it's turned into a major, major company that runs so many of the world's businesses. Right. And does so many different things from security to management of computer systems and data centers to virtualized machines. So you can run all sorts of different kinds of operating systems either in the data center and on your local machine. Right, right. And you know, I just finished this book about spatial computing, which is how autonomous cars and robots and virtual, virtual beings, we're soon going to meet virtual beings, which is a whole separate conversation and all of this new stuff coming in augmented reality glasses. I mean, yesterday Facebook was showing off prototypes that they're wearing around with seven cameras on there and AI chips to recognize things in the world. Right. And be able to talk to the computing system. So our computers are about to really radically change and our businesses are radically going to change over the next few years. And that's what my click was about. And that's, that's what led me to talking to you.

It's like, that's great.

VMware wanted me to talk a little bit about some of the things that is going to be discussed at VM World in a week.

Right.

Everywhere I turn, I see AI everywhere.

You can't escape it now. Yes, yes, taking over.

And soon consumers are really going to understand that. I mean, I'm an early adopter of Dune, Evans and Tesla cars and, you know, Spotify and I've been around and I understand how these things are built. But soon you're going to be wearing a pair of glasses and everything is going to be AI, right?

So potentially, potentially, that car that you're in, I believe has 17 different neural networks operating in different chips, reading the sensor data and deciding is this a stop sign? Is this car getting too close to me? You know, all kinds of predictions going on in the car itself. And it's really fascinating how it's pervaded the world so quickly, you know.

Yeah. So who are you? Just introduce you to the audience. Because my audience, first of all, my audience probably doesn't know VMware very well. They probably heard the term on CNBC or something, you know, the company. But they don't know what you guys do because it, unless you, unless you're working in IT or data centers, you probably don't have a good idea what VMware is. And then my audience probably doesn't know who you guys are. So who are you?

Well, I'm Justin Murray. I've been at VMware 14 years now. I joined when VSphere VMware. VSphere was young and I've moved through a series of roles, mainly in technical marketing which have to do with emerging technologies. A couple of years ago it was big data, but now very much in the last three or four years it's been machine learning and GPUs and particular. And there's a very tactical answer I could give which is customers want to know how to use GPUs the best way on vSphere. But the field is much, much bigger than that, I think. I'm interested in helping customers understand how to do machine learning with VMware as the operating system kind of thing. And it is becoming the operating system for all kinds of clouds. AWS has it, Google has it, IBM has it. It's not just an operating system for on premises data centers anymore.

But.

What fascinates me like you, is the amount of innovation and uses that are going on. Like I've been dealing with medical images for the past few months here and detecting Covid or non Covid in those images using a machine learning model, which is, yes, you know, that's not in the doctor's office, but it is in the research institutes in the nih.

When I went to Israel, I met a company, Zebra Imaging, that can see cancers in x rays and MRIs. And it's more accurate than the doctors that see seeing certain things. Right.

More than the human. Exactly.

And AI is a pattern recognizer, right?

Exactly, exactly.

And so if you teach it to see a certain kind of cancer in Just see, as soon as it sees that pattern, that's that cancer. Right.

And the reasonable thing about it is it will give you a probability that it is or a probability that it's not. It's not giving you a 1 or 0, it's giving you a, you know, 90% probability or 0.9 or 0.1 probability. So it's, it's being a little cagey about its decision, you know, and so there is a need for human decision makers in the loop is what we've learned that this thing can't take over and take over our jobs completely. It's not quite there, but it's getting closer and closer. And voice recognition, image recognition, and the way I explain it to my customers and my colleagues is, you know, business intelligence that you've been doing for years. You've been generating reports and SQL statements against your massive amounts of databases. Well, that's kind of reporting on today's situation or last week's situation or last month situation. ML is different. ML has the in common with that, or machine learning has in common with that. There's a big chunk of data that you can throw at it, but it's not about reporting on what's happening right now. It's about predicting what's going to happen in the next minute or second or hour or day. And that's. Yeah, that's really the interesting thing is it's, it's future looking. It's trying to tell you ahead of time what the patterns of data and the past data have told it. And the bugbear of all that is, is the past able to establish the patterns actually that are going to occur in the future? Could we have predicted in January that we'd all be working outside the office six months for March onwards?

You know that.

And so in reality, patterns in data are changing all the time. And so past data is only to be trusted for a certain period of time as giving you the indicators of patterns in it. And so you need to constantly refresh and redo your training. Training being that part of machine learning that applies the data to a training algorithm.

And we should talk about training because I've already seen a new trend that the cost of training, a new system, it is coming down. It's almost like a Moore's Law, it's coming down exponentially and it's becoming faster to train things. Right, right. And here's where I was going with this. I was talking to the guy who runs augmented reality at Trimble, which is a construction company and they're they're, they're buying these robot dogs from Boston Scientific that you might have seen on YouTube. Right. Little dogs. And the dogs walk through the construction site and image it with a lidar. So it's imaging it in 3D. And then the AI looks for safety problems, things that might hurt a human tomorrow. Like you said, Prediction, prediction. This car is predicting a kid is going to walk in front of it. So it slams on the brakes. Right? So. Right. And it's always trying to predict what's the next action of things on the street to threaten the car. But this can mean a lot of things to all sorts of businesses from retail stores to construction companies.

Correct? Correct. The simple example is you bought. I'm being old fashioned here, but you bought album A, so you might be interested.

I remember those days.

You remember albums.

Big things that are plastic, you know, aren't they wonderful though? You put them on a record player and put a needle in them.

You know, little works of art. Right. But you streamed program A, so you'll be interested in streaming program B. Recommendation engines and recommendations. Engines are a huge industry in themselves. You know, whole teams of software are building those things. And that's the simple example. But the slightly different example is, okay, you're at a checkout in the store. Whenever you go to a store these days.

Wait, wait, time out. You already using old terminology, checkout, Right. At Amazon, there's not a checkout. You just grab a bag of chips and walk out of the store because the camera is watching you, picking you up. And a 3D sensor and maybe some other things watching you shop. My best friend is doing R and D in the retail industry and they're forbidding him to use the term checkout. There's not going to be a checkout.

If not a checkout. I've been through one of those Amazon stores, I think in Seattle you like.

To talk about old things. I already know said vinyl and checkout counters.

Yes, well, drawing examples purely. But the one I was going to give was if somebody faked the barcode and took a cheap barcode and put it on an expensive item. The only way around that is for a camera to tell it, hey, that's a $4,000 bike that he's selling for $20 here. You know, that's actually a real use case that we've come across. So there are, there was a company.

I met who can take a picture of a driver's license and know whether the driver's license is real and has been modified.

Wow.

Right by the Pattern in the, in the plastic on the driver's license, Right? Yeah, you're right. It's endless. There's not going to be a part of human life that's not touched by AI soon if already.

Yeah. And I'm fascinated by that tension between the business world is full of tabular data, data that comes in tables, and the exciting spreadsheet. Yeah, the exciting ML world is full of images and voice recognition and stuff that has nothing to do with tables. Now I used to be confident in my view that you need to accelerate the latter workload. You need to throw as much GPU and hardware at it as you can because it's doing neural networks, they're heavily compute intensive and you don't need that for tabular data. But I'm not so confident in that view anymore because why not? I see all those companies that make their living out of ML for tabular data, you know, Salesforce, lots of other companies, smaller companies like H2O and DataRobot, the vendors in this world, they're now adopting GPUs as fast as they can. And now you're seeing traditional tabular models like XGBoost, that's been famous for a long time in the tabular world. You're seeing that getting accelerated by GPUs and Spark and I just came off a project where we applied spark with GPUs to XGBoost and got a huge boost in performance. Last year or the year before, I would have said to you, you can do that with regular processors, you don't need any special hardware for this. The specialized hardware that's popping up all over in this ML world is really fascinating.

Yes.

The going to be.

In fact, my Tesla had an Nvidia card when I bought it and they took that card out a few months ago and put a Tesla built AI chip that has 18 billion transistors in it. Right. So there's, there's an Apple just, just announced a new iPad with AI capabilities built into a piece of the GPU. Right.

Why?

For people who, who don't know this new world and are coming into it. Why is GPU so used? I keep hearing this all over. This is why Nvidia stock price keeps going crazy, right?

Oh yeah. And they're now purchasing yet another company. Yeah. So the GN GPU as you know, stands for graphics. So these devices were intended to accelerate rendering and shadowing and ray tracing on sophisticated workstations for AutoCAD and that sort of thing 10 years ago or more. And they've been hugely successful in that. But what was going on on the covers there is that matrices were being multiplied by other matrices at high speed. And those matrices indicated coordinates in the. In the graphics, somebody who is very smart said, what's happening inside a neural network? A collection of modules that are strung together in a network that's exactly the same thing. It's multiplying a matrix of weights by a matrix of inputs, matrix multiplication. Where does that operate the best? It operates on parallel processors, parallel cores, because you can do all of the layers in the neural network, all of the nodes in the layers at one time in parallel. So a very bright person said, let's apply this to executing neural networks, and got a 10x or 50x performance increase. And that's why GPUs are being talked about so much. But they're not alone. You know, GPUs can be tied into fast networking such that a GPU can talk to another GPU over the network without going through the cpu, without going.

Through the main memory.

And that's an experiment that we've been doing and using it in clusters so that the type of data scientists that needs more than one GPU to do their work can get it from the other side of the network without having.

To somebody doing simulations, really, in depth simulations of chemistry or physics or military or. Right. Or Uber. That's some crazy stuff. Yeah, true.

Uber and Twitter, who went to the trouble of inventing distributed frameworks to get the best out of, you know, they've got three or four hundred GPUs spread across their data center, and they're utilizing them at 50% or less. And so.

And that brings us to VMware. Exactly right. VMware for people who don't know the history of VMware. You know, back in the old days, businesses had to buy a rack of their own computers and set them up in their own data center. Right. And was really expensive. VMware and a few other companies comes along and goes, hey, we can split that machine up into virtual pieces. Right, right, right. So that we can rent piece one to this company and piece two to this company, you know. And that's cloud computing today, right? Exactly. And VMware EN was the software that did that. Right. So you guys are doing the same thing now for GPUs?

Yes, we are. And you know, I'm kind of surprised and delighted in the way that Nvidia is helping us with this. They invented the concept of virtual GPU, which allows you to do. Exactly. For GPUs which we. What we've been doing for CPUs for a long time.

Yeah.

And that's. That's kind of method and I think.

I have a thesis and it'll be interesting to see. To talk to you about this thesis. GPU training something. Right. If I'm going to train this car to see a stop sign, I have to take a lot of pictures of stop sign or video, I have to tag it and then I have to throw it into a training system that runs a lot of GPU for a short period of time to build the. The training set that it runs the car. That gets uploaded the car and runs the car. Right.

Or train it for a long period of time too. Yeah, but.

But it's bursty. You don't need it all. But it's not like Twitter, where the stream of information going through Twitter every day is pretty much the same all the time, 24 hours a day.

Right.

When I'm working at Tesla, I one day I'll probably have a really big training problem and the next day a little training problem. Right. And. And I'm not using the same amount of GPU every day. Well, why would I want to buy my own GPUs in such a world? Rent them from somebody like Amazon or Google or, or Nvidia or. Right. And. And I bet that's why they're talking to VMware, because VMware enables that whole process.

Correct? That's absolutely right. And yeah, you're right that there is lots of bursty use some of this. Some of the customers we've looked at are 50% or less usage. When you need it, you need it all, but you don't need it, you need more all of the time.

If you're gonna like. Tesla's data set is getting pretty sizable now. Right. So if you're gonna retrain on new images that, you know, worldwide with a million cars, that training set can get pretty crazy at. You want to throw as many GPUs at that problem as possible to build that training set and then upload that to the cars, right?

Correct, correct. And you have 17 or more of these neural networks being trained at once. That's a lot of power you need there. So you need to be able to commandeer power when you need it and then relinquish it when you don't. And that's a VMware problem.

I think it's funny that Tesla's R and D lab is right across the street from VMware. I'm sitting in VMware in Palo Alto and literally across the street is Tesla and Pivotal, which is another cloud computing company.

I know that building. I know that building. Yes, they're right there in Deer Creek. Yes. Yeah, yeah, yeah. They're, they're an advanced user of all this for sure. And they're some of the leading lights in the ML world, work for them. So we've had some of them to speak to us at VMware, in fact. But yeah, I met somebody.

They still from your cafeteria here before, Before COVID they'd come across the street and eat lunch.

We have the best cafeteria for sure. Well, we're into sharing. We're into sharing. We see people from Uber and Tesla.

I hinted at it before that the cost of training is coming down and partly because you can rent servers and not have to buy a lot of servers. Right. Not have to buy a lot of GPU and you can scale it up to the workload that you have. And if the workload gets bigger, you can do that. What's happening in your view, what are the trends that maybe a little bit. What are you guys going to talk about next week at VMworld and what's happening that you're seeing that maybe people don't know about in terms of the cost of these things? Yeah, the capabilities.

There are a whole bunch of things happening. A whole bunch of Nvidia, for example, acquired Mellanox. Now, why does a GPU company acquire a networking company? Because.

Well, you give the hint.

Exactly.

To hook them up together fast.

Hook them up fast, exactly. So that it looks like one big computer, but in fact it's distributed around. And so distribution is going to be a key theme in this machine learning world. And it hasn't really hit there yet because distributed training is a hard problem. Distributed training involves swapping all those weights that we're adjusting in the neural network between all the participants at intervals during the training process. And so Haravod was invented by Uber to do just that. It was a refinement of something Google had already done. But it's. That's, that's really in its early days that that whole distributed training business, the majors have figured it out, but it hasn't pervaded the rest of the industry yet. So that sort of distributed stuff is coming along with that. What you were saying, which is dedicated hardware that is designed for training or design for inference that's coming to companies like Graphcore are building specific hardware and ASICs for executing training better than anybody else. And you can see Google thinking about this with tpusc and Nvidia has a range of products now, one of which, T4 is destined for inference. Whereas the A100 which just appeared on the market, the new high end machine, which has sharing in it by the way, built into it, that's for training. The high end is for training. So people are customizing. When people start customizing the hardware, the specific software needs, you know you're onto something important here, you know. So those two themes I think are going to be very big. Distributed distributed operations and custom hardware. What, what is fascinating to me is the pruning of the neural network tree, if you like to optimize it. So neural networks can have 150 layers in them, with a thousand nodes in each layer or 500 nodes in each layer, they can be huge. They might not fit on one GPU in the memory of a gpu. So a whole bunch of academic research is going into do we really need all those layers and do we need the layers in a tree shape or in a curved shape or what shape is best? And there is an interesting paper from a whole bunch of Google folks, including Jeff Dean, the guy behind a lot of innovations there, on where is the best location to place layer one as opposed to layer two and layer three and layer four as far as a GPU layout is concerned. So this is not there yet as far as mainstream is concerned, but it turns out that you can find patterns using machine learning itself as to where the best placement of the individual layers within a neural net network should be.

Wow. So the AI is making the AI smarter.

There you go, more efficient.

Which efficiency is going to matter? Right.

We've turned, we've turned it on itself now. So if we've got a problem to solve, which is where is the best combination of GPUs to execute these 150 layers or prune version of that? Why can't I give us the answer? Having viewed a whole bunch of neural networks histories in the past, I would.

Expect VMware is using AI all sorts of places, particularly with its multi cloud systems. Right? Because when I built websites, I'd buy a machine and put it in a rack somewhere. Right? That's what I used to work at Rackspace. That's where they got their name from, Right. But now, yeah. Why doesn't an AI decide where to put that in a virtualized machine and maybe even move it around depending on price or speed or closeness to the customer? I mean, if I'm playing augmented reality games, I want those packets to be really close to me. So the virtual machine should move closer and closer and closer to me to have less latency and stuff like that.

You're exactly right. Traditionally at VMware we call that distributed resource scheduling, which is suiting the workload to the layout of the hardware that we've got under our control. We've been looking at algorithms for that for years. But you raise a broader point, which is can VMware improve its products using AI or machine learning? And we're full of really smart engineers at VMware, so they're all looking at that very closely. And the ideas are. We have internal innovation conferences where the ideas just flow like crazy about that. So there's a whole bunch of things in the pipe for exploiting machine learning for our own products as well as helping customers do their machine learning as well and make their businesses better. So there's kind of a three legged stool here, three angles to ML. One of which is improving our own products. Another is improving our own business. Like setting pricing based on ML for our products. That's done today using ML.

That's right.

And then applying it to business situations that customers have, like fraud detection and banking or in healthcare, like we mentioned, or a whole bunch of different businesses. I read recently that something like 40 to 50% of the new Kubernetes applications will have some element of machine learning in them. Yeah, it's a huge portion. Yeah.

Agileness. I mean, some of this was to build new kinds of products. I mean, and there's some commonalities. My iPhone, my Tesla car and my June oven all have improved after I bought them. They got better. Right. In a big part because of AI. AI lets you build exponential learning systems that get better over time and learn as they learn more about the tasks that they're trying to do, like cook toast.

Right.

They get better. And they get better for everybody who has a June oven because some engineer at June trains it better to see different kinds of toast. And all of a sudden everybody's oven does more right after it improves. But my Tesla card changes lanes today. It didn't do that when I bought it.

Right, right.

You know, and so it is amazing. Right? And it totally changes. Auto industry is having to react to, you know, these ideas. Right. Because right now it's the expectation. When I bought a BMW or Toyota, I never expected it to improve after I bought it. Right now I do the next car better have that already built in. I mean my. It's so deep that my brakes have improved because they downloaded to the car. New algorithms to work the brakes better.

Adjust the brakes. Adjust the brakes.

Yeah. And even like I'm using a little gimbal, right, which has three little motors in it. It's a $128 DJI. It's using AI to figure out how to turn the motors to keep my iPhone steady.

Well, you raise, you raise an interesting subject here, which is how small can we make this? How small a device can we fit this into? And now you're beginning to see why Nvidia is interested in arm. Because it could be, it could be the case that the actual big money here is in the inference phase, not the training phase. Now, we've made lots of money on training so far because that consumes gigantic machines with lots of GPUs in them. But how many millions or billions of devices like your gimbal or your control panel in your car?

Security camera at the door here. Right. On and on and on.

Right.

So factory floors are being digitized in a whole new way with a process called digital twinning. Volkswagen told me about this and they're using AI on the factory floor to design the factory floor. So it's, it's crazy, it's endless.

Super interesting. But the challenge for us techies is how much memory can we put in that low power device to hold the trained network and how fast can we make it? So how small can we make it.

And how battery efficient, Right?

Yeah. And Google came out with an experiment not so long ago which said all of our cell phones that we're carrying around with us, why couldn't they be participants in a federated distributed training exercise where the phone is in charge of its own data? The data doesn't go anywhere, but the data is being used to train a model. That trained model is then sent back up to Google and amalgamated with all the others to make one super trained model. Now that's a fascinating idea that you federate the job out to every phone user that you know, and the results come back in miniature form from the device. They've used local data like what site says Justin visited on his phone recently. And we can train a central model based on all that data. So this is federated and distributed at the same time. You know, yeah, there. This is the new renaissance in computer science, this one. This is the business, you know, it's endless.

When, and you know, you. Yesterday Facebook was showing off these prototypes that they're going to have a few hundred workers walking around with, with seven cameras and a AI chip inside that recognize analytar. Right. So that it recognizes the world and recognizes everything you touch and everything you look at. And they're trying to figure out the privacy implications of that. And, and how, how, you know, what they have to do to keep from freaking people out. Because we can go into black Maryland real fast, right? Yes, but what you just said with Google, if I'm touching a Coke can in a, in a, in a grocery store and I teach it something new about Coke by talking to it, that's going to want to be shared by everybody else, right?

Yeah, yeah.

And so that it recognizes, oh, there's a new Coke can maybe for Christmas or something just showed up.

Yeah, yeah. There's a lot of possibilities here. On a more mundane level, customers have come to us like airports and have said the cameras are feeding video data back to a central data center somewhere. We want to be able to tell immediately if somebody's in a place they shouldn't be. We want to be able to do surveillance using machine learning. And there are companies now that specialize in that. And, you know, there's a whole security angle to this and there's totally. There's a bias angle to it as well.

There's a what?

There is a bias angle to it in the sense that have I trained my model on a representative set of images?

Or.

Is this a fair model to women and minorities? Or is it biased in the favor of us guys who do, who do tech, you know.

And because of the training, you can sometimes pick up wrong thing or you can pick up patterns that shouldn't be right. Racism is a pattern that shouldn't be. We want to get rid of racism in the world, right, and make things more fair. But if you look at home buying behavior, right, You'll. Your AI might pick up a pattern, right, that shouldn't be right. And then all of a sudden it reinforces that pattern in the future on people. That's a bad thing. I assume you're going to have some ethics discussion, many ethics discussions, and there.

Is a whole field of discussion going on around that. How do we introduce fairness and prevent bias in the machine learning training?

Here's what I. I was at the data Science conference one time and the founder of Truro, which Apple bought, was giving a talk and he said, you know, we had a problem one day we were training AI to see Wolf and Husky, right? And we went through thousands of photos teaching it how to see Wolf and Husky. And he said it. After the training was done and they started running the system, it made an occasional weird error. And this thing should Never give an error. Right?

Yeah.

And they, they had to unpack the training data that it builds. Right. The neural network.

Right.

And they said what was funny is it started learning on is there snow or grass in the photo? It wasn't even looking at the dog or the, or the husky or the wolf. It was looking at is there snow in the picture? Which was usually accurate because a husky is usually on snow and a wolf is usually on grass. But once in a while, you know, they would switch and it would throw an error and it learned on the wrong thing. He said we had to remove the snow and the grass from the pictures to retrain on the dog's face, you know, and on the dog, on the wolf. And that's a very simple idea now, but, but this stuff still happens at a big scale. So if you're business, you really have to have experts who really think through what kind of data you're going to teach it. Right. Where are you going to teach it?

Precisely. And you put your finger on something in that discussion that's extremely important, which is if I'm a person working in a bank and I'm in charge of recommending somebody gets a loan or doesn't get a loan, I'd better be able to explain my decision making to a regulator. And explaining what a neural network is doing is tough. It's really tough. Unraveling a neural network is, you know, an almost insurmountable problem. And so we've heard of situations where unless it can be explained as to what it's doing, unless it can explain what the most important weightings and the most important columns in the data are, then it's not going into production because we can't explain ourselves to the regulator for that reason. And so there's a whole field of interpretability being researched. And actually companies like H2O and others are quite mature in that world by building a simpler model that you can explain that produces a very close result to the sophisticated model. You can say by correlation, I can explain the simpler model to you. And so the more intense and sophisticated model is doing the right thing because they came up with the same answer. You know, that sort of interpretability idea needs to be solid before this can go mainstream in many businesses. You can take.

We planned on 20 minutes of discussion discussing AI. We could go for hours and hours.

Sorry about that, I didn't get into my.

No, it's such a big topic. We could go for hours and hours in a VMworld. I'm sure you'll have several Sessions on different topics we covered today. There's hundreds of sessions. Right. For people to watch on video.

If you were looking for the theme, I think in the VMware sessions, there's lots of stuff about innovation. But if you're looking for the theme, it's VMware providing the infrastructure machine learning rather than providing the application logic itself. We have that and we've built our own and done it for ourselves in our business. But I think our core competency is how do we optimize the infrastructure, partly to do with sharing, but partly to do with dedicating as well, so that the data scientist is happy without even knowing they're on a virtual machine.

Thank you.

That's the core competency here. That's what we need to get right. Yeah.

Well, thanks for spending some time. I could talk to you all day long.

Likewise. Likewise.

You're in the middle of it.

I'm in a very lucky place. I'm having the best time of my life here.

Yeah. Thank you.

Thank you. Yeah. Have a good afternoon.

Yeah. All right. We'll end there.

Call again if you need some more on this. I'd be happy.

No, this is. That's cool. I just talk to people like you to understand how. Where. Where are shifts that I don't see coming, you know, and where. Like, you gave me some insights on inferencing as a splitting apart and visit. Visibility of these neural networks is also another thing that's coming. I had heard that.

But yeah.

Justin is if we look at the explosion in AI, Right. And I look at the web apps that came along, and then I look at IoT and my thermostat has a computer in it now, but now everything is exploding into. Not only do you have to have a computer in almost everything, but now your computer has to think. And to our point where you don't have the horsepower in the device itself, it means you connect it to a very fast net, back to somewhere. And then everybody that's building apps now not only have to worry about their app and their web app and the interface to the altitude device and network speed, you also have to worry about back end bgpuc capacity to provide the necessary horsepower to do this work.

Right.

Yeah, you're absolutely right, Eric. And when you're building the app now for the device, the question you're asking yourself is how can this app learn? How can it learn from the patterns it's seeing and how can it do better as a result of that learning? Which is what Robert was saying about this.

Yeah. Well, how, how can you collect data that then you learn with? Right? And that's a big one. That's a big one. A lot of businesses have to really rethink themselves. Like what, what do we need to start collecting to be able to build a training set, to build a system that learns?

And Tesla must be thinking, why can't Robert's car, why can't Robert's Tesla learn by itself as opposed to sending data back to us to do it all centrally? Why doesn't learning take place and learning is an active verb in machine learning? Why doesn't it take place in the car itself? Send us the results.

For me, pinning it back to VMware, you've already got a ton of infrastructure that does the web computing, that does the connecting the networks and talking to the IoT device. And so VMware coming in going, hey, and by the way, we can just extend our infrastructure to give you VGPUs and allow you to extend your application into this space. Makes a lot of sense because instead of needing to rack more hardware, we can give you that layer that can connect to all the clouds. That, to me is like, the ability to scale out, to meet this demand is going to be one of the big IT challenges. We have the infrastructure to do this, and, hey, guess what? We're investing in the ability to manage your VGPUs and allow you to manage all this and scale it out.

Well, my friend right now is up in a Microsoft data center putting in GPUs, and he says these data centers are so massive compared to what I was looking at when I worked at Rackspace. He said some of them are almost a mile long and just shoved with machines and they're putting a lot of Nvidia cards into them. He says. So you can see, I mean, that's just one company buying this stuff, right? And I'm sure Google and Amazon are doing the same thing, just buying a crapload of these Nvidia GPUs and hooking them up.

And if that's going to be a cloud data center, you can bet that Microsoft will be carving up the GPUs and giving you half and me half or a third each. Between the three of us, it's not the case that they're going to dedicate the whole device to one of us unless we pay the premium. But yeah. So I look back at the history just Yesterday, and since 2013, we've been doing VGPUs at VMware for graphical reasons, to share the GPU among graphical applications. It's only been a year or two, maybe two years that we've been using those same VGPU concepts for compute, for heavy duty neural networks, so that, you know, a lot of our customers have GPUs for reasons that are not machine learning, but by nighttime, when they're not using them for their desktops or workstations, they could switch those over into machine learning quite easily. Yeah, that's another.

That's interesting. That's actually interesting. You know, Nvidia just came out with the 3080 card this morning. And that's like a supercomputer. It is a supercomputer, really. You know, and if. If you're buying this for a VR company, right, you have a lot of GPU just sitting in people's desktop, unused at night. You guys could come in with a layer and say, hey, we can use all that. Right.

We call it desktop by day, machine learning by night. We have a whole theme of guys looking at that.

That's interesting.

Yeah. All you have to do is switch the VGPU type from graphics intensive to compute intensive, and off you go. The same device can be used for the new purpose.

That's cool.

Yeah, that's cool.

Can mine some bitcoin too, right?

We used to do that, but now we've turned to machine learning instead.

Actually, if I remember right, somebody read an AI to sense whether somebody is cheating and cheating the clouds and using the GPUs for Bitcoin mining because they don't want that workload.

I think it influenced revenue for a little while, but the shine has gone off there, I think.

Yeah. Well, Cole, it's been an honor talking to you.

Likewise.

Hopefully this comes up. Send me the recording and I'll chop it up. And I'm gonna go do another intro in front of the sign and put that in.

Sounds like you're on a plan. Love it and can't wait to get it in market. So, Justin, thanks a lot again for taking time out. You're just amazing as always and curious. We will stop recording now.