
Rendered.ai CEO Nathan Kundtz on Using AI to Build Better AI - Ep. 177
NVIDIA AI Podcast
Using Physics-Based Synthetic Data
The industry has bifurcated somewhat where you have physics-based synthetic data being used for things like rare events, edge cases. And then in other circumstances, you can use generative algorithms to sort of create more data of things that you've already seen. But unfortunately, it has a limitation of you. You don't get that kind of rare event in edge case coverage.
00:00
Transcript
Play full episode
Transcript
Episode notes
Speaker 2
Like I don't know if people like seagulls. Oh no, like,
Speaker 3
I mean, one of the early days we went down to the strip near us, the touristy area with all the seagulls. You know, we got fish and chips and, you know, sitting down on a bench enjoying the sunset. And the seagulls just swooped us and took like all of the french fries and just knocked it over. And they're going out everywhere. Like ever since then, I'm always protective
Speaker 2
of my chips. Just seagulls. They're terrible. I mean, is there another,
Speaker 1
I guess, I guess in Austin, we have crackles, but the crackles are not really desperate forward. Yeah. Yeah, but the seagulls are just like, I mean, they're just like you would expect a sailor to be. They're like, hey, that's mine out of the way. Mine! And just barge into it. Yeah, yeah, yeah.
Speaker 2
So
Speaker 1
yeah, frogs. You have a lot of frogs there in Australia? Do they show up a lot?
Speaker 3
I think so. I haven't noticed a lot of frogs. Really? Yeah.
Speaker 1
Huh. We'll have to look. I mean, there must be frogs in Australia. I mean, by this point. But I wonder, are frogs like an indigenous species in Australia?
Speaker 3
I'm sure we
Speaker 1
have frogs. Yeah.
Speaker 3
I have ridiculous numbers of lizards. I'm sure they're frogs. I just don't live near the area. Well, that's the notion. So that's the problem for me.
Speaker 1
She also said that like, you know, when we live in Dubai, just like living with lizards in your house was just the way it was. There's always lizards. And spiders. Yeah. Yeah. So, you know, that's, you know, speaking of living with lizards, I think it looks like there were a few articles written this week. It looks like despite predictions that Kubernetes would be boring, we're going to be living with it in our house forever. I think it's always going to be there. And I actually had forgotten, not, I hadn't forgotten about it, but I don't think about it. And because it was, I think it was maybe, you know, Joe, Joe, Kubernetes, Joe, Joe Kubernetesium. I can't really make that pun well. Kubernetium, Joe, does that work? Keep going. I don't know. Yeah. Thank you. You know, his whole deal was that he wanted Kubernetes to be boring. And this was three, four years ago, which was just to say, like, you know, in the same way that like, aside from maybe listeners of this podcast, file systems are boring, right? It's kind of recede into the, the efficient lower levels of the stack,
Speaker 3
except when the maintainer, you know, murders his wife. But that's a different kind of excitement. Well, that's
Speaker 1
that's a different podcast world, the murder podcast. That would be us. I mean, listen, I am in no way making light or fun of murder, etc, etc. And so forth. But like, I feel like finally my wife, my wife, Kim, and my podcast universe could finally converge.
Speaker 3
They finally intersect.
Speaker 1
If there was a murder podcast, she listened to that was about whatever it is you just talked about.
Speaker 3
The rise are up us. Yeah. That would be they could they could go back and forth. It's something for everyone that could, you know, spend part of the time talking about optimizing B trees and the other time, you know,
Speaker 1
talking about burying people under trees. I think I think maybe I need maybe maybe maybe we
Speaker 3
maybe that that's one of the few podcast ideas that I don't think we're going to pursue. No,
Speaker 1
okay. Well, that's probably a good idea. Anyhow, there was I I read through the article and it made some points. But I think the headline is always with the headlines like this sums it up. Kubernetes predictions were wrong. We're still all interested in it. That is not become boring. It's it's more thrilling than ever. And people are into it. I don't know, Brandon, were were they wrong? Were there any predictions? What's the story here?
Speaker 2
Well, I don't think
Speaker 1
there were in the sense
Speaker 2
we've been pursuing the same goal probably for all the time that we've been
Speaker 1
professionally working and computing and maybe, you know, long before that is just like,
Speaker 2
hey, then the day, we're trying to make things simpler to build, you know, application logic and deploy it quickly. And I think, you know, with every wave, there's this like, Oh, it's going to be simpler. And then it's never really simpler, right? And we just keep making it more of our complicated. So, you know, to me, it's like looking at reading the whole thing. It's sort of it's not so much like Kubernetes, the container orchestrator probably is relatively simple. And if you will, boring, if you contain it to that, what is not boring is the CNCF, you know, cloud native landscape, where, you know, whatever, however, many logos were up to. And that seems to be like, in some degree, like I think we should all blame ourselves. This is kind of what we want, right? Like it's a very, if you will, engineering and software architecture has sort of just kind of just driven the bus for the last 10 years. And it's like, yeah, everything should have an API and we should have multiple implementations of the API and we should have multiple subsystems. And we just want ultimate flexibility. And I think, you know, to some degree, it's like, that's what we want in computing, right? But at the end, it doesn't, you know, a decade plus into this is like, we're definitely not getting any faster at like deploying, you know, application logic, nor is it getting any easier for like, you know, large corporations that have hundreds of applications to like deliver new functionality, right? I think in some ways, it's just a minor revision of things we've seen in the past. What do you think, Matt?
Speaker 3
Well, it's the constant tension between specialization and generalization, right? We now have a platform, Kubernetes, that anyone can run just about any workload on for good or bad, right? You can, you can run bare metal, you can run GPUs, you can run at the edge, you can run in space,
Speaker 1
on planes, you know,
Speaker 3
Kubernetes is everywhere for good or bad, right? And the, the, the reason that the CNC of landscape looks like that is because everybody has some sort of specialized use case that they want to bring to this. For me to say that I've got, you know, the one true path, well, as soon as I run into the next organization, they're going to be like, yeah, but what about this? And that, right? So it's, we're, it's the same story forever, I think. I mean, you know, it might not be Kubernetes for, for some organization. It might be like, oh, our, you know, instead of Kubernetes being our lingua franca, our lingua franca is, you know, Linux, but then, you know, we're deploying across this. We've got the same sorts of issues. We have to maintain applications across lots of platforms. Oh, wouldn't it be great if we had APIs for everything? Wouldn't it be great if we had something that applies consistency? You know, so a lot of these deployment and usage patterns, it doesn't matter if it's Kubernetes. It's just for, you know, good or bad, we've mostly adopted this as a lingua franca. But there's plenty of people succeeding without it. You know, you got our friend, DHH, he's got his, you know, MERSC rename. He's, he doesn't need Kubernetes, but you know what? He's read, he's re-implementing the same patterns. Yeah. And so. Well, maybe what I'll at least it's open source. Yeah. At least it's standardized.
Speaker 2
The part I guess I keep thinking about maybe this is just wrong, but so but I'll throw it out there was it does feel like in computing at some point, like a level of abstraction sort of emerges and becomes good enough that we don't think about it. And I think we were kind of joking earlier about like the file system. I don't know. It just works, right? The operating system, it just works. You can go deeper, like, you know, not mostly, mostly no one writes assembly. There's all these languages and these compilers and we just, you know, pick your high level language and for the most part, you just go with that. And it's, I guess that's kind of where I just feel like we've just been stuck for a long time around. It's like, well, I want to build a, you know, you know, if you will, like some type of scale application, let's call it that I'm going to cloud. Just sort of something with scale. And, you know, and I know, you know, Adam Jacob is trying to make a run at this. It just feels like, you know, like the Kubernetes and the CNCF, like we're just kind of stuck, we're just kind of stuck with like more and more abstraction. So I've just been like kind of looking around as like anything else, even emerging, like kind of joke, I like to run as I love the marketing, even the company went out of business, like the data center operating system, DCOS, right? So the thing is like, yeah, like there should just be some operating system kind of thing that it all just kind of works. If you kind of, if you will live with this level of abstraction. Now I watched the Adam Jacob presentation he did at one of the conferences. It's on YouTube. We'll have a link to it. But he mentioned this thing I never heard about is that the wing programming language, right? And it was sort of this, the short story is, is wing combines infrastructure and runtime code into one language. So I was like, oh, this is kind of interesting where it's like, maybe this is like, instead of having dev ops, right? The two groups to work together, it's like, no, everything is specified in kind of this programming language. And then if you will, when you deploy it, it's not really a compiler in the sense of a traditional compiler, but it sort of just figures out all the underlying infrastructure for you. And, you know, it just kind of just works, right? So you don't have to know about Terraform or CloudFormation or Pileming or whatever, all the other things. And so I don't know, like, you know, man, I really want to draw opinion because it's like kind of your old world and kind of your new world. It's like, is something like this like, you know, is the world ready for something like this? Is this a way out of the CNCF landscape? Can we just like get above it? And if you will, the compiler for the cloud, if you will.
Speaker 1
No. I mean, when so I
Speaker 3
looked at it and I was like, that sounds great. And then, you know, there are two dozen options to alternatives, right? I mean, we could stick with Terraform. We could stick with Pulumi. We could, you know, implement something in a CDK or an SDK or, you know, there's just so many ways to skin this cat that I don't, I mean, you know, good luck wing. I think that it'll have some traction, but I, it just didn't feel that compellingly different enough to me to want to spend more time with it. And I think that's, that's a problem for like anything new, right? Is you're in this crowded landscape where you have a thousand choices. You know, you have the CNCF landscape in front of you. And how do you differentiate yourself?
Speaker 1
I do.
Speaker 2
What do you think you sort of made a run at this with your time and pivotal and different, but similar ideas, but different implementations, like, you know, are you ready to dust off the pivotal messaging and your own? So
Speaker 1
CEO, the former CEO, right? He's not trying something new. I don't know. Yeah, modernizing mainframe stuff with AIs or whatever. Yeah.
Speaker 2
That, you know, sure.
Speaker 1
Well, you know, I have two thoughts branded. One is, I don't know. I mean, I mean, coincidentally, I guess to a pivotal way of thinking about it. And, and then all, but then also if you look at like many of the configuration things out there, like, I feel like maybe you need to just like pick what infrastructure you want and stop dicking around and dedicate your life to that infrastructure. Right. And instead of all these, like, different configuration things that work with all the stuff, just be like, no, we work with this stuff. This is what we do. Right. Yeah. And like all that time that we're going to save, trying to figure out, you know, being, you know, we just want to be mono cloud and, you know, that'll either work or it won't work, but the chance that it won't work is probably as good as the chance of other things that won't work. Right. Like you just, you go into that platform and you're fine and then sure you've got all these companies that are 200 years old that have all sorts of stuff. And you got to manage that. And that's a thing. But like just sort of stick to the configuration that's built for the platform that they're using. I don't know if that's good, but it seems like maybe we should just try that and see what happens. Now, now, now the second, well, the, the, the, the middle thing until I get to the second is like, I, I, you know, y'all were making me think. What sort of, what sort of constant. Specification, not programming language, but sort of like specification config kind of format has like lasted the longest. You know, there's programming languages. But if you think about it, like HTML and JavaScript have kind of like resisted anything else, like sure, you've got like Objective C and whatever Android is, but like those two things, whatever it is, they are, it's very confusing. But and they've evolved and they've been in place. And so something about how they've existed and that whole ecosystem, like, you know, they haven't gone away and they. Well, that's, that's the.
Speaker 3
I mean, JavaScript is evolved like crazy.
Speaker 1
And right. But it's like, it's like still there.
Speaker 3
Like, yeah, well, I mean browsers aren't going away. And, and I think, you know, Terraform had the window to be that. And now people are actively trying to tear it down or replace it with Opetofu and other alternatives. Sure. Sure. I mean, that's, that's the closest thing I could think of in this space is, you know, something that became the lingua franca. And then people are like, I'm tired of speaking French.
Speaker 1
And then so that gets me to my second thought, which I was, you know, kind of now that I've rediscovered that there's people actually on Twitter and it's not just a wasteland of Bitcoin and vaporware or, you know, that kind of aestheticy stuff. I was saying, like, I don't think there has ever been a configuration language that anyone has liked.
Speaker 3
Like, like, whether it's like, I mean, that's, that's just like Kubernetes, right? It's, it's going to have its warts and there are people who are like willing to go with, with what they have. You know, they're like, fine, this is the worst best alternative.
Speaker 1
And so like, you know, I'm sure there are people who complained about like dot files, although I think people like dot files, they seem pretty into that. Like, I think, I think as always with things that people don't like, what they don't like are things they don't like. Now, if they, if they, if they, they don't like bad things, but I'm sure there are plenty of, you know, there used to be this whole thing of sharing your dot emacs file and your bash file. You know, like, mine's actually do like dot files. They just don't like other people's dot files. Maybe that's the rule of configuration stuff is like everyone likes configuration, just not other people's configuration. But, you know, and then, and then I was never big into the windows world, but did people like dot I and I files? I'm not sure. Like people, someone did point out that you used to have the, what was it? Auto exec dot bat? I think people like the fuck out of that. It will really end of those files. And, and then, but then you had, I think, I think in Java, we had dot property files, I don't know what Pascal was using or what they use and see. And then you had, you had XML, which turned into WS star. We're not allowed to like that. That was apparently terrible. Nobody liked that. Yeah. And then you had, I don't remember. Like, I don't remember what like what rails used.
Speaker 2
No, you're forgetting your favorite. Cote is all the, the network configuration, right? You know, the, the files. Yeah. Yeah. One that, I mean, you know, you named a bunch there, but like, you know, your first one, I mean, not a config file, but HTML, it's like, you know, like, like is a strong word, but it's like it's stood the test. I mean, it's probably going on 30 plus years. I mean, if the web's around, it's like, and I
Speaker 1
don't know, you know, there's been
Speaker 2
a few attempts, I guess, so W flash to the kind of recreate the front end client, but really nothing has even come close to like displacing it. So do we like it? You know, maybe that's a two, two strong word, but it's like it does fine work. Yeah. And then we would like the fact that like, yeah, just, you know, at the end of the day, you know, you can just spit out the smell and it generally works for the stuff. I guess, I guess
Speaker 1
maybe what I was wangling to in my mind without knowing it with the HTML is like no one owns it. There's no commercial entity that owns it despite the attempts of like Microsoft early on and Netscape. Yeah. But and so therefore I wonder if it's the case, if we were to look at the other configuration stuff that kind of comes and goes all the way up to your, your wing programming language, there
Speaker 2
could be a theory that
Speaker 1
like when it comes to configuration and automation scripting, if someone owns it, eventually it goes away. Where it was like, if there's no commercial interest around it, it can just live forever because like, you know, you don't have to worry about making money and then making those decisions. I don't know if that's true, but like it seems like. Like no one owns a gamble, right? So maybe that's not forever.
Speaker 3
Nobody trust me. Nobody owns you. Maybe the different
Speaker 2
question, I guess I'm posing.
Data is the fuel that makes artificial intelligence run.
Training machine learning and AI systems requires data. And the quality of datasets has a big impact on the systems’ results.
But compiling quality real-world data for AI and ML can be difficult and expensive.
That’s where synthetic data comes in.
The guest for this week’s AI Podcast episode, Nathan Kundtz, is founder and CEO of Rendered.ai, a platform as a service for creating synthetic data to train AI models. The company is also a member of NVIDIA Inception, a free, global program that nurtures cutting-edge startups.
https://blogs.nvidia.com/blog/2022/08/31/rendered-ai/