← Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs
Discover Dauntless XR’s Vision for Smarter Workflows
00:44:11
Discover how Dauntless XR is using AI and XR to transform training, data visualization, and technical workflows across industries.
Loading summary
Transcript39 lines
- [00:05]
A
Welcome to Reshaping Workflows with dell Pro Max PCs and Nvidia, where innovation meets real world impact in high performance computing.
- [00:20]
B
Welcome back to another exciting episode of Reshaping Workflows with Dell Pro Max and Nvidia RTX GPUs. And today we get to do two things that we normally don't get to do on the show is one, we get to talk to one of my favorite partners, Daunless xr. And two, we actually get to talk about xr, which we don't do a whole lot on this show. We talk about M and E engineering workflows. But today we get to talk about XR and how it's impacting kind of the world and their customers. Kind of all enabled through Dauntless xr. So with that we have three people with us. We have Laura Lee, Sophia and James. So let's start with Lori Lee. Give us a quick introduction on yourself and then we'll go to Sophia, then we'll go to James.
- [00:59]
C
Yeah, thanks for having us. So my name's Laura Lee Elliott and I'm the CEO and co founder of Dauntless xr.
- [01:06]
A
My name is Sophia Lazzaro and I am the co founder and chief Product Officer here at Dauntless xr.
- [01:12]
D
My name is James Ire and I'm the Chief Technology Officer here at Dauntless xr. Thanks for having us.
- [01:18]
B
Of course. Now, absolutely thrilled to have you all on. And we've done some work just to set some context and background. So we've been working together, think it's been maybe not a year, maybe nine months now where, you know, when we kind of first had initial conversation, you were using some Dell equipment that wasn't really designed for your workflow. We'll just leave it at that. And was able to kind of upscale, you know, kind of the, the training timelines because a lot of what you do ultimately does involve kind of AI and a bit of machine learning. But we'll get all into that. But before we do, Laur Lee, maybe give us a little bit for those that don't know. I would assume that maybe some people listening are familiar with Dallas xr, but I would assume the vast majority aren't. Maybe give a little bit of background on kind of the company where you started and just a quick overview of that and then we'll dive right in.
- [02:01]
C
Yeah, of course. I co founded Dauntless with Sophia back in 2018, we'll call it. And Dauntless came about after I had been working in engineering and construction for a number of years and was really frustrated with the lack of technology. And at the time I thought, you know, the obvious thing to do was to start a tech company to solve this problem and provide more digitally native tools to the frontline workers in these kind of hazardous and out there industries and jobs. So we got together and came up with the initial concept for our first product which is called Katana. And Katana delivers XR guided workflows along with any data that you need to complete a particular task. And we'll, we'll come back and talk about Katana a bit more later when we talk about what we're doing with AI. But yeah, we started off with a focus on kind of construction and engineering, like you know, these very hands on frontline jobs. And then during the pandemic, we had the Air Force reach out to us and ask if we could maybe use Katana on Air Force bases to help with training and getting aircraft ready for flight. And we'd never thought of that before, but we were like, there's no technical reason why that shouldn't work. So we started working with them. And we've since gone on to work with NASA and a few of the other agencies and kind of out of our work with, with them, we came up with a second product which is called Aura. And Aura takes machine generated Data and creates a 40 beautiful holographic display of your data so that you can go through and review what you have in a way that is a bit more intuitive and makes, makes more sense to a human than to a computer. And the kind of first iteration of Aura was to display flight data. So we would take information off of the information and data off of an aircraft and we would turn it into this great 4D display over a 3D topographical map. And you can go through step by step of the flight and see what was going on, where you were in 3D space, what the aircraft was doing at any point in time. And you know, after that we were like, well, this is just the first kind of step for Aura. We realized that there was a lot of data out there that is underutilized because it is really difficult to parse through. Since then, we got to work with NASA to do some visualization for space weather data. So looking at, we created a kind of digital twin of our inner solar system and pulled in the live or liveish satellite data for all of the fleet that's monitoring the sun. So when you're in an aura session for space weather, you can see what all of the different sensors out there are doing and it makes it a lot more, I think, approachable, especially if you're not a space weather expert or if you don't have a degree in astrophysics. You can still, you know, jump in there and understand what's going on. Going on.
- [04:52]
B
So the only thing I know about space weather is one time. This is completely weird story, but I was in Nicaragua, you know, on a vacation visiting a friend down there that had a business down there. And there was like a big coronal mass ejection I think is what you call it. And he was like, oh, this is happening. It means there's going to be an earthquake. And I thought he was totally full of crap. Like it was like pseudoscience. But I'm not lying. Like within a day there was a minor one, right? Not a huge one, but like still it was an earthquake. So that's all I know about some space weather. But I, I think you make an interesting point, right Is that, I mean someone on this show recently, and I, and I agree with them is that, you know, a company's data or the data that you have the rights to is kind of the new oil, right? It is the new gold rush. But so many times there's really valuable data, but it's setting in a form in an Excel spreadsheet or a sharepoint or wherever it may be where it's not really usable, readable or really consumable by anyone in any way because it just can't be consumed. So so if you tell us and we'll get more into Katana but like tell us a little bit about, about Aura and how the product you design really makes it easy for someone, whether it's, you know, small company, up to, you know, some of the federal agencies to take that data in and be able to visualize it and some of the benefits they've seen, yeah, a lot.
- [06:07]
A
Of it comes down to the format that is in and I think the, the way you put it is, is, is a great way that the data companies have is their oil. Unfortunately, accessing it can be not useful. So we'll use the aircraft example. So if you are a pilot flying a training mission, learning a new maneuver, learning a new landing or takeoff strategy for a new airframe, the feedback that you have is what you remember from when you were in flight, which is we all know can be fallible. What your trainer remembers or your co pilot remembers. Also quite valuable. Or this giant CVS file with more headers and columns than you want to know because some of the sensors on these flights, their data sampling rate is as high as what James, I think the max one that we saw was like was 200.
- [06:55]
D
Every quarter of a second. Yeah, every quarter of a second maximum, yeah. That we were experiencing in that file.
- [07:01]
A
Yeah. So if this is, you know, a flight from Houston to Atlanta, imagine how many, how many files are in there and you're trying to look at a specific area of the flight where maybe you had to maneuver around some winds or things like that to see what the readings were telling you and how you react, acted. You're parsing through rows and rows. I mean, to the point that I opened one of these and it crashed my non Dell computer when I tried to just open the file. So what Aura does is it takes all of that data and then parses it into something you can step inside of. So in the case of an aircraft, right, we're going to lay a 3D topical graphical map of the bounds of the geographical data that we have in that data set. Then we're going to draw to scale what the flight was and then we're going to extract the most crucial data. So the things that the pilots are looking at and then we're going to represent that data in the fastest way for their brain to consume it. So in the case of flights, it's going to be a primary flight display, which is the way that they're used to looking at a lot of their primary controls. And then what was so fun about this platform was because we were building it to not just solve for flights, we had to really use a lot of abstract thinking to be say, like, okay, we have to categorize these sets of datas. So these are minor controls that you always want to see. These are data sets that drive the baseline scene, which would be the map. These are data sets that drive the bounds of what the user can and can't do. And that was what enabled us to then take the platform from a product perspective and adapt it so quickly to space weather. Because then again, we're rebounding, right? The bounding data is now what is the farthest satellite that we are taking data from at to Earth, right? So then, then our new bounding becomes the inner solar system. And so that was from a product line perspective. What was so interesting about that one is because we could take this really inaccessible data and kind of categorize it into ways users would experience it in 3D. And that is what enabled us to take that platform and apply it to different use cases.
- [08:58]
B
I love that. And I think it's so smart. And I don't want to like be an episode where I'm just telling you how great you are, but, like, I'm going to give you a compliment is that I've seen a lot of startups because I talk to a lot of AI startups, right, and their core competency might be X. And someone then asks like, well, can you do Y? And then they pivot very quickly into something or they build something so specific for one customer that ultimately when a new customer comes along, it takes a complete redesign, which that's what I think I love about Aura, is that it can take really anything, right? I mean, space weather data, chronomas detection versus flight data, kind of very different and that. And it still works. And James, this kind of leads me to a question for you is that, you know, when you were building kind of that framework and you know, what are some of the things like data wise that Aura can like ingest? Right. Obviously there's, you know, timescale data, there's, you know, all types of stuff kind of describe one kind of all the data that all the hypothetical data could take in. But two, you know, for companies that are out there that are like, wow, this would be a really interesting use case for me. What do they need to think about with their data to be able to utilize it for, you know, Aura or XR in general?
- [10:12]
D
Yeah, that's a great question. You know, I think that a lot of times what we come to is, you know, like, what kind of data do your users need to play with in order to make insights into what it is that you're trying to analyze. So for the Air Force use case in particular, we had to make a separation kind of right off the bat in terms of what is session data versus what is like an application layer data versus what is a user interface layer data. And so there were some separations that we had to make. You know, like the raw data that we were ingesting from the CSV files were from the flight data recorders. And so thankfully, because they were an understood data schema, it didn't take a whole lot of work in order to make them all compliant with each other. It was an understood ontology in terms of like what the relationships were between the data points and what they were trying to represent. You know, like latitude and longitude can only mean so many things. The roll, pitch and yaw of the aircraft can again only mean so many different things. And we also had to create that again, that ontology or like the relationship between those representations in the XR space when it came to the user interface layer and the application layer. So for the user interfaces, you know, as Sophia kind of touched on, we recreated a lot of the displays for the primary flight displays and instruments that the pilots were experiencing in the cockpit itself, because it really helped from what we were understanding, speaking to the pilots and the instructors that were doing these, that were conducting these debriefings, is that there was a lack of accountability in terms of you're going through these files. And so like the. Sorry, not to like side sidetrack a little bit, but the way that they were doing these debriefings before is they were printing out reams and reams of these sheets from the flight. I'm not kidding. Yeah, it was like straight out of the 1950s, where they're just going line by line and they're just like reading timestamps and they're saying what it is that they're doing. And at the end of the day, there was still a lot of, like he said, she said, in terms of, you know, what actually happened during a flight, because, you know, you can both see the data point. And so the instructor says, you were climbing too fast or you're going too slowly at this point in time, whereas the pilot would just answer, no, I wasn't. And then so, you know, it would just go back and forth because you can't actually see what it is that the instrument was reporting. And so there was no context to the data itself. And so what we were able to do is create a user interface layer that was able to take the session data and say, okay, well, this is what it looked like in the cockpit to you. And so now we can all sit here and make an agreement or a more informed decision about what it is that you should be looking at next time in order to avoid this mistake. And then the last point that I wanted to talk about here is that the application layer also had its own data stream or its own data domain that you had to think about in terms of how this session data was being presented to the user in the context of everyone else. You know, especially in mixed reality, where you are projecting the hologram or the data into a three dimensional object that's being rendered for everybody. You have to recon. There's a reconciliation that has to happen between the digital space and the physical space that you're all standing in to make sure that when you're pointing at a specific point in space to reference it to someone else, everyone else is seeing the Same thing that you're pointing at. And so there's that application layer that has to understand the space that you're in and project things in a way that everyone else is getting the same experience that you are. So when we were first working on this project, you know, like Aura's main advantage is that it takes that Venn diagram of these different domains and layers and is able to do a lot of the heavy lifting in terms of what is relevant from the data that you're uploading from a single fil. And how does that get parsed out to the different layers that are, that need to receive the data source? So in terms of, you know, like in a roundabout way of answering your question originally, you know, it's able to ingest a lot of different data types. You know, we created it in a way so that it is able to read from multiple different file sources at once. Because again, with the Air Force use case, they needed to be able to see how a single flight was coordinating with multiple flights that were happening in the same area at the same time. And so there was a consideration that we kind of had to take right off the bat in terms of being able to ingest multiple data streams and multiple sources to create a single session.
- [14:47]
B
I love that. So quick follow up question when I love your answers. By the way, don't ever think you're not talking enough because it's, it's all just gold. It's data gold. But from the use case of printing reams and reams of paper on a 1983 Okadata printer with the side things that used to spool it through to what experience that Aura provided them. Like just quickly talk about the time savings it went from used to be feedback debrief took seven hours and now it's down to minutes. Just give us kind of. And that's what I think XR is so interesting is that it can take that and just shorten the time. So now you're on to the next, you're on to the next, you're on to the next.
- [15:28]
D
You know, I think that the most extreme example that we can think of is there's a named event that happens called Red Flag where it's multiple, up to hundreds of flights happening at a single time during a training exercise. And so you have a room full of pilots going through and debriefing again with sheets of paper and they're going line by line. And so it can take, no kidding, it can take days or even weeks from what they are describing. And so Instead, you're able to debrief the flight in real time. And, you know, you have timeline controls that. We were pretty adamant about developing this into the application from the start, where you were able to pretty much like, have video controls where you can seek to a point in the timeline, you can fast forward, rewind. Pretty intuitive stuff. And so it would cut that debriefing time down to easily half an hour, under an hour, you know, depending on the situation. Because you could just fast forward to the point in time that you needed to talk about. You don't always need to go minute by minute.
- [16:28]
B
I love that. I mean, I, I dread, no offense, I dread corporate meetings, like in the office for multiple days. I couldn't imagine, like. So. Yeah. So, Laura Lee, do you have anything to add, Add to that?
- [16:41]
C
Yeah, I just wanted to jump in, I think, with, with Red Flag, you know, like James said, there's, you know, dozens of flights happening at a time to coordinate on a single mission. And if they were to debrief minute by minute, it probably would take days or weeks. Realistically though, I think when they're out there, they. A debrief can take eight hours. And that is a, an. An entire conference room full of people sitting there and going through each action and maneuver for eight hours. And it remind a lot of what, what we had to do in EPC when we do like 3D model reviews and you are all sitting in a room going through a 3D model, kind of, you know, piece by piece. And yeah, with, with something like aura, you could debrief a very complex mission with, you know, multiple aircraft in the air in an hour. And the, the time savings is definitely, you know, like a huge benefit. But it's also so much more enjoyable because you're not there kind of, you know, defending your memory, being like, no, I'm pretty sure, like I fired here at this point point. And they're like, this guy said you didn't. Or you, you know, you turned too early or you turned too late. It takes that kind of element out of it. So, yeah, there's definitely the, the time savings, but then there's a, you know, like cognitive ergonomic impact as well.
- [17:54]
B
I mean, if my wife had this, then she would be able to call me on all my stuff because at the end of the day, I remember nothing. So it's like, it's real time memory, which I love. But now we've kind of talked about aura then. You know, with xr, I think what is so interesting is the AI aspect of It Right. So now you're able to visualize. But how can you use that experience to one train an AI to teach you how to do something or how do you flip it on its head to use the AI for xr? So you kind of briefly mentioned Katana. So Laura Lee, if you don't mind, give a, give an overview of Katana, kind of what it is and then we'll move on to Sophia for a follow up question.
- [18:30]
C
Yeah, of course. So yeah, Katana was our first baby. It was our, the initial app that we built out and it's a hands free guided workflow application. So on the web app side you go and upload all of your procedures, checklists, training, all of that kind of content in very human generated content and it turns it into an XR native workflow for you. So you can see the instructions for what you need to do, you can see any associated data. So if you have like spec sheets, Data sheets, drawings, 3D models, things like that, you can associate that with a step and it will all show up beautifully in your field of view and then you're able to progress through those steps to kind of assist you in a task. And one of the things that we had, people had been asking us since Cortana was just an idea in a pitch deck was the ability for image recognition. And you know, in, in 2018, we didn't have the kind of concept of computer vision or AI or most people didn't have that kind of concept of those technologies like we do today. So when the opportunity came to start layering in computer vision, object recognition, image segmentation, all of, all of that good stuff into Katana, we jumped on it because people have literally been asking us for it since day zero. Not even day one, since day zero.
- [19:52]
B
That's awesome. I mean, and I'll, I'll take a brief point is that if you were at Dell World this year, we had Donald, you know, doing a lot of things, but one was in the booth to kind of show the power of this, which is it was the example, I'm going to tell you, was designed to be fun for a trade show. But we're going to get into some real world use cases. Here is the coworker of mine built this, you know, had this LEGO set and it was build an AI factory with Nvidia. Right. Not a hard. It had the book and everything. Even though I had to track down the book and my daughter was real mad that I ripped apart the set. Anyways, doesn't matter. But what. And we'll come to James on this question here in a second after Sophia, but basically was able to use the instructions to be able to use computer vision to figure out which piece came next, kind of where it would put, you know, where you would put it, and then kind of the next step. Right? So that is a prime example of kind of the XR experience and then layering in AI to kind of enhance it. Right. So, Sophia, a question for you is that, you know, within Katana, you know, some examples that you can share. Do you have a couple of use cases from customers or even, you know, hypotheticals that you've thought about where you've injected AI via Katana into kind of the pipeline and ultimately what the results were.
- [21:08]
A
The goal for Katana from a product perspective is if we think about, you know, the premise behind the Internet for a lot of people was we're democratizing access to data, right? It's now so much faster and easier to access it. So why isn't everyone experts in everything? Well, because you need the right data at the right time in the right context, right? And this is where XR and AI, I think, work together so well. And really what we're trying to do with Katana to advance towards that goal of democratizing so that you do have access to the right data at the right time. And we'll use a comparison here. The best way often to learn something, as many of us know, is to learn alongside someone who already knows how to do it. So let's use the example of like changing your oil in your car. Very useful thing to do. A lot easier to learn if you're next to someone doing it, rather than trying to go up and down from looking at a screen or a printed PDF with pictures over here and trying to translate the 2D images you're seeing into the contextual 3D environment that you're in. But up to this point, that is kind of like the gold standard of how to learn is to be standing next to the mechanic doing it together, or the person who knows how to do it. This has been one of the fun use cases for xr and I think we're going to release it for a couple of models on our public version of Katana. If people want to try it is rather than having to have those reference documents or have that mechanic, because maybe you don't have access to them, is being able to just put on a pair of glasses and then we can see. Katana can see what you can see. And it's providing the instructions from the Source. Right. So right data, right time, but it's also providing that real time corrective feedback. And that's the AI component that the version of Katana that we launched years ago couldn't do, but we always wanted it to do and it brings it so much closer to that step. So that's one example is really hands on things. Like, I mean, Laura Lee can talk about this as well, but her experience in commissioning in on energy projects, these are real experiences where you really do need to be hands free. You're climbing things, you're climbing over things, around things, carrying a binder with you or finding the right person to go with you at the right time. Not always as accessible as AI enabled smart glasses. Right. So those are some of the scenarios that we see. We're also, I think interested in less industrial ones as well. I think those are going to come later. But one of the ones that James has talked a lot about in his experience is in like a lab setting where you have humans and robots and you have a lot of super expensive equipment all in the same space, providing people with right data, right time and corrective feedback where they're looking to prevent them from having to run to the other side of lab and potentially eventually knock something over. So those are some of the use cases that we're really interested in right now. Another use case that we've gotten into with the corrective feedback is the earlier stage of pilot training. So we talked about debriefing with Aura, but another stage that we've talked about approaching with Katana is when pilots are learning early with switchology and cockpit layout. There's a lot of muscle memory that they have to build and I kid you not, they build it today with a poster. With a poster that folds out like a diagram. Like, you know, I know a picture one of those, like if you did like an 8th grade science fair, those like trifolds, but it folds out like a cockpit and then you're pretending to push these buttons. Right? And they do it. It's called chair flying. And they do it to build muscle memory over and over and over again. Because in times where they have to do a lot higher order thinking and decision making, the last thing you want to be doing is remembering how far back over your head you have to reach for that for that specific switch. Right. But the, that is, it's expensive to do in a simulator. But doing it with a paper tiger or a cockpit poster can be really inefficient because you could accidentally do it incorrectly and no one is there to tell you. But if you're doing it in a 3D model in a headset with AI error detection, which adds some like. Interesting because now the computer vision has to segment not just the physical objects in the view, which is like your hand and what it's doing, but then also the digital objects. Like did you reach for the right switch and did you pull the knob out the right 20% or whatever. There's some really interesting computer vision challenges that we're, we're having to face there. That's yet another layer where we can remove a lot of friction and add a lot of like real time expertise that frankly we couldn't have done before and we couldn't have done without the new Dell equipment. So the Dell Pro Max can, and James can get into this more specifically, is able for us to train a lot more data a lot faster, which enables us to real time test all of these iterations and then make them available to customers.
- [25:37]
B
I love it. And you teed up my next question perfectly because I'm just kind of going around the screen. So James, so we'll talk about the Dell Pro Max here in a sec, but you know, in the. Because there's a lot of ways to skin a cat like when it comes to AI. Tell us a little bit and you don't have to give away any trade secrets, but tell us a little bit about. Let's use the example of, I mean, God, the paper flying. I mean I just, I, I'm shocked. I'm learning so much. But let's use that for an example. And you're trying to create something or a katana to be able to do that. Like what is kind of the workflow that you use on the Dell Pro Max to ultimately test, iterate, train and deploy for an experience like that.
- [26:15]
D
You know, I think that it's kind of, it's interesting seeing your reaction because I feel like I've been like desensitized to just how much of the world still operates like it's the 1800s. But it really is like, it's kind of shocking sometimes how we have such access to tech and we're just, we still don't know how to interact with it or utilize it in a way that makes sense and is intuitive. And so you know, not to like wax philosophic here a little too much, but I think that at a high level, the thing to remember is that it's really, it boils down to how do we interact with computers. You know, like human computer interaction is a Conversation. I think that needs to happen more often when it comes to AI, because as Sophia is, she put it so succinctly, and I always repeat it now where the last thing we want is another chatbot. You know, like it really doesn't always make sense to have that kind of an interface. And how we are interfacing with computers and technology determines a lot of their use case. And so, you know, when we think about the intersection of AI and xr, when people are using XR technology, the device itself needs to have a holistic understanding of the space around the user. You know, why is it that you are here looking at this object at this point in time and what is it that you are that you were doing previously that is relevant to this object right now? And so being able to understand and make those inferences in real time in a way that doesn't completely destroy your battery life or, you know, completely chew up all your processing power rendering each frame, trying to understand that is only improving. And I think that as we come up with more use cases that cater towards that, the technology is only going to improve. And so when it comes to the chair flying aspect, as Sophia was talking about, there are a lot of challenges that are easier yet harder at the same time in terms of measuring what's going on in the real world space versus what is being rendered in digital space in front of the user. And so a lot of the processes and workflows have very binary decision trees and logic flows in terms of did you do the correct process in the correct order? You, you either did or you didn't. You know, if you did switch the flip, excuse me, if you did flip the switch on at the correct time, then yes, you did or you didn't. And so there's not really a whole lot of AI processing and logic that has to happen at that step. It's more at a high level in terms of getting the application to have the intelligence to understand when it should intervene with the user's flow? What sort of mistakes do they have a tendency to do, and what are the best steps to take that might help mitigate that? And so in terms of having, it's nothing's going to beat having an experienced teacher sitting over your shoulder watching you and being able to correct you in real time, but with that's not available, you can have an AI that the teacher trained that is able to have an understanding of what are some of the methods that are best used to interrupt a student or to, you know, teach some new patterns of behavior. That coincide with how the student has a tendency to learn best. And so the AI can do an analysis in terms of, you know, these are the kinds of processes and these are the kinds of situations that the student has a tendency to make a mistake. And so it can either try to intervene or it can have different kinds of methods to nudge them towards the correct decision. But yeah, it's, it's a lot of just making the interface between the AI and the human feel a little bit more intuitive. Sophia kind of brought up also a little bit. I learned a lot of this with my work. Before I was with Dauntless, when I was working for a company that was doing XR and AI for life sciences labs in particular, we were trying to work with Pfizer Biotech and we were creating what they were calling their lab, the future concept. And it was the intersection of all of these things combined with robotics involved because a lot of their vaccine production during the pandemic was being modernized at a very rapid rate with robotics. And a lot of these scientists, you know, who have like a PhD in virology, they're not necessarily robotics experts and they're not used to working around a three ton, you know, robotic arm that's going to come around and knock them out if they're walking into a danger zone. Which unfortunately was one of the, one of the reasons we were starting to do this. Yeah, it's, you know, because they just weren't used to that. And so a lot of what we were trying to do is under, have an understanding of what is the AI best at, which is analysis and scheduling to a degree that humans, I think, are just not as quick at. And how do we put, you know, how do we separate these domains that we are best at in terms of humans are best at figuring out new creative solutions and AI is best at analysis. And where do we mesh those two technologies to or where we mesh those two things to allow them to work best.
- [31:16]
C
I just wanted to go back to your question, Logan, because you were asking about processes and like where we were using the Dell Pro Max and then we got into talking about Katana processes. But James, when you, when you describe like how your, your kind of workflow for, for putting together the AI application that we debuted at Dell Technologies World, I was wondering if you could also share a bit about the synthetic data and what we learned about that because that was a new thing for us that just came about when we were getting the demo ready for Dell Technologies World and then. But also please talk about the the. Your process and like, where, where everything kind of fit in. I know we went down a little bit of a rabbit hole, so I'm like, we need to go.
- [31:56]
B
I was here for it. I was here for it. I was letting him go because I love it. It's mesmerizing.
- [32:01]
D
I. I can talk. Yeah, thank you for reminding me too. That's a, that's a great point. So for a lot of the AI training using, for the Dell demo that we had, that we were training the AI model on a Pro Max using synthetic data. And so the synthetic data was. It really came out of the fact that, Logan, you were still trying to track down the actual LEGO set that we were going to use and you were still trying to track down the instruction booklet. I really didn't have anything to work with. And so I had to go track down 3D models of these LEGO bricks and export them in a way that, you know, I had to export them as 3D objects. And so I'm kind of sitting there scratching my head about like, well, you know, I can't just sit here and wait for the bricks to show up. And even still, when they would, when they did show up, I wouldn't necessarily have time to generate a bunch of physical training data because that would require taking hundreds of thousands of actual images of the bricks in different lighting scenarios and backgrounds. And. And so we had the idea to just kind of automate the whole process, especially since we had the horsepower of the Pro Max behind us to really fuel this whole process. And so I was able to create a system that would generate a bunch of synthetic data for the LEGO bricks. It would put them in different lighting conditions, different environments. Because we have that Nvidia GPU that's in there, we were able to simulate a lot of very realistic lighting conditions and a lot to add a lot of variance to the. The imagery in terms of obscuring the objects that we were trying to detect or create different levels of challenge for the AI model with the training data. And so we were able to create a system that was able to generate the synthetic data and then turn around and immediately train an AI model on that synthetic data. And so it was pretty close to a one click process where I would kick off the generation of the synthetic data, which would then offload to a training model, which would then train the object detection model and then still be lightweight enough to run on mobile hardware like the Quest 3, which we were running it on at Dell World. And so it really started creating this feedback loop in terms of creating better and more refined training data. And it gave us a really nice framework for being able to capture and use a lot of different types of synthetic data. Like if we were to expand this out further, we could use this to generate real time training footage from, you know, say like a game engine that was able to utilize the Nvidia GPU and the RTX capabilities to really show a lot of variance in terms of the lighting conditions in the environment as well as materials rendering so that the surfaces look about as close to accurate as possible and again show different levels of obscurity to try and give the AI model as much variance in the training data as possible and make it really robust. And yeah, we were, we were still able to run these pretty lightweight models at the end of the day on the Quest hardware and we ran into issues even then when we were at Dell World because the lighting conditions, you know, like it was, it was interesting seeing it happen in the real world.
- [35:18]
B
That's awesome. So I just want to say you're welcome. Because of my procrastination, it taught you a new skill which you're now going to play. I hope my check's in the mail, James. I hope my check's in the mail.
- [35:29]
D
Oh yeah, yeah, yeah, of course, yeah.
- [35:31]
B
I'm waiting, I'm waiting. One last question, James. And we're getting kind of towards it, but you know, you mentioned kind of the Pro Max, you know, in a. Not necessarily that synthetic training run, but using the Dell Pro Max with the Nvidia RTX card. You told me a pretty incredible stat when you were using the previous PC, like how long would it take you to complete a training run to running a very similar training run on the Delpro Max. The tower like you went from X time to X time?
- [35:57]
D
Oh yeah. I mean like so typically I was training so just to, you know, give a benchmark to some of these things. You know, I would typically train which with you know like a batch size of like 16 to and like 100 epics, which you know, is not a whole lot of training and not a, you know, like not necessarily great stats. And it would take me pretty close to 28:30 hours to complete a full training run on a new model as opposed to running on the Pro Max. You know, I was able to push the batch size up to 192 and I was able to process hundreds of thousands of images at a single time. I was able to store all of that on the, I was able to cache all of that on memory, you know, Due to The increase in RAM and each training epic was cut down from 15 minutes to maybe 2 minutes at most. And so I was able to conduct twice as much training, maybe even three times as much training in the same amount of time, you know, and so situations where I would normally be setting up training and then like, oh, I guess I'm gonna go walk my dog for like the next two weeks and then come back to maybe a trained model, it was. Instead I would go set up a training run, I'd make a cup of coffee and I'd come back and it was finished. And so it was incredible time saving.
- [37:19]
B
I love it. Well, I'll tell you a funny story because you're based in Hawaii, is that I was doing a training for an animation studio on all their artwork to try to replicate or whatever. It was about a million and a half images, which is huge. Like that's a massive training run. And I, I left for Hawaii and then I came back and it was done. It was about 10 days, but it was mad. It was absolutely massive. Yeah, trading run. But I love, I mean I love it. I think that's the end of the day is when it comes to AI and xr, it's having the right hardware to equip yourself ultimately to be able to do the task. And that time savings is what it really makes the difference. So we're kind of up against it, but I want to ask what I like to do at the end. And there's three of you. So we'll kind of go down through is, you know, starting with. Or like pretend someone, you know just started watching the episode. Now what is that 30 second sound bite? You would want them to walk away. And remember from this episode, it can be about Dauntless, it can be about you, it can be about how you need to. Guys need to pay me for the synthetic data capabilities you have because my procrastination. Whatever you want.
- [38:25]
C
Oh, that's a really good question. Well, other than go back to the beginning of the episode and watch it from the top, but thank you. That aside, I think it, the, the, the takeaway here is that, you know, these technologies are all complementary and they're all kind of converging. And the way that we use XR today is going to be changed by AI and the, the two might not be as separated as they are now going forward. And that unlocks a whole other level of capabilities. You know, we, we only like briefly touched on, on them. In this episode, we were talking about being able to identify Lego Bricks to build a little AI factory or using AI and XR to be able to identify, you know, like this is a zero offset butterfly valve and here's all of the data that you need to run an inspection on it. Or you know, you, you're doing your pre flight inspection incorrectly. Move your hand over here three inches to the left. You know, the, the way that we interact with AI right now is not the way we're going to interact with it forever. And to just, you know, keep in, keep your eyes open, keep your mind open to how that is changing and how you could, could apply that in your, in your industry and your job in your life.
- [39:41]
B
I love it. Sophia, same question to you.
- [39:44]
A
This is maybe more of a, a product oriented question because that is the lens through which I see the, the world. But as a someone who came up in software, product management, pre AI, the thing that gets drilled into you is that you want to compress your iterations as small as you can so you can learn as soon as you can. Because the sooner you have a hypothesis and you test it and then you look at the data, the faster you can respond to your market needs, your customer needs. You can validate your assumptions, right? And what's been so fun about being in tech the past really year and a half the most is how fast AI is compressing that iteration cycle. Right. I can tell if the user story is written poorly, if I kick it into a vibe coding app and it kicks back something that's not what I expected. Right. I don't have to wait a sprint to get it back from the developers, but that just when you think your iteration cycle is tight enough, it can actually get tighter. And we saw that happen right from James coming to stand up and being like, yeah, the model's still running. I don't have feedback yet on whether or not this training one was right. I don't have feedback yet on if I need x new data set or anything like that to going from. Well, since we've last Talked in standup 24 hours ago, I've run three different trainings and this is what I've learned and this is what I need you to go validate with our customers. And here's a new model for you to go load into your headset and test. And so if you're not looking at ways to keep driving that iteration cycle down, you should be. And it really does come down to having the right hardware in place and that's made a huge difference for us.
- [41:13]
B
I love that answer. That's Going to be a social clip. We're putting that on on the Grahams. All right, James, or rounded out for you, probably through the technical lens.
- [41:22]
D
Yeah. From a technical side, you know, as we've touched on before, it's really down to how are we interacting and interfacing with technology. And a lot of human computer interaction, I think, is a conversation that needs to happen more often when it comes to AI, because there is a world that's fast approaching where XR and AI are kind of on a collision course in order to make each other viable for everyday use. And that's kind of where Dauntless tries to put itself in terms of, you know, like the kind of software that we develop, because it's really about trying to create more intuitive interfaces to allow us to interact with AI in a context that makes sense and is relevant to what the user is doing. And it's about utilizing the data that your organization already has access to, you know, like time and time again. A thing that we, that I always bring up, you know, to Laura Lee and Sophia, regardless of who we're talking to in the situation, is that the real insights happen when organizations are allowed to play with their data. It's not so much about trying to understand what data you have at your disposable, but trying to create a sandbox in order to allow your organization to make real insights and make connections that they wouldn't have had before.
- [42:39]
B
I love it. I absolutely love it. I knew this was going to be a great episode. So with that, if you're watching this episode right now, which you hopefully are, and you should be, we will link to the Dauntless website as well as the information about Katana as well as Aura and the, you know, in the description below, and we'll definitely have them back on and we'll do like a 2.0 of more kind of XR heavy things because I think it's well worth it. And I think the key takeaway for me in the episode is that XR is kind of upon us. It's been around a while, but just with the use cases and everything that you've heard today, ultimately you can see how XR is impacting. You know, whether it's teaching someone how to fly a plane or how to fly it more safely, or helping people debrief or looking at coronal mass ejection in space weather. It's really upon us. So with that great episode, thank you guys for coming on. This is Logan with reshaping workflows with Dell Pro Max Nvidia RTX GPUs. Until next time, keep your workflows running locally and we'll see you on the next one.
- [44:01]
A
Do what you want.
- [44:02]
D
Do what you want. This podcast was produced in partnership with Amaze Media Labs.