Transcript: Generative AI Will Revolutionize Autonomous Vehicles

Helm AI’s Vladislav Voroninski on Revolutionizing Autonomous Driving with Generative AI Simulation

Executive Summary

In this episode of The Road to Autonomy podcast, Vladislav Voroninski, CEO & co-founder of Helm AI explains why we are in the “golden age of AI” and how specialized generative AI models are the key to solving complex industry problems like autonomous driving. He details how Helm AI’s simulation platform provides a more scalable and efficient path to validation than the traditional large-scale fleet approach, creating a new paradigm that can help automakers level the playing field against pioneers like Tesla. Vlad also pinpoints the industry’s biggest bottleneck and shares his vision for the future of autonomous systems.

Key Topics & Timestamps

[00:00] The Golden Age of Generative AI 

Vlad describes the current moment as a “golden age of AI,” highlighting the momentum behind generative AI and the shift from general foundation models to more specialized, multimodal applications across various industries.

[01:13] Two Paths to Commercializing Gen AI Models 

Vlad discusses the two primary ways generative AI will be commercialized: general foundation models that face commoditization and specialized models for specific fields like autonomous driving, which are better positioned to capture long-term value.

[03:15] Developing Specialized Foundation Models for Industries like Autonomous Driving 

Vlad explains that developing specialized models starts with identifying a critical problem in a field, such as addressing corner cases and validation in autonomous driving, and then using generative AI simulation to close the gap between real and simulated data.

[05:58] Is Data the Real Bottleneck for AI? 

Vlad argues that data is not the primary bottleneck for training advanced AI systems, suggesting that humans learn to drive with relatively little data. He emphasizes that leveraging broad, internet-scale data is more effective than relying solely on specific driving data.

[08:00] Identifying the True Bottlenecks: Hardware, Validation, and Regulation

Vlad lists the main bottlenecks in autonomous driving as hardware, validation, and regulation. He asserts that while software and validation challenges are becoming solvable through AI simulation, the biggest hurdle is establishing a clear regulatory framework.

[11:30] How AI Systems’ Decision-Making is Validated 

The discussion covers methods for validating an AI’s decisions, moving from difficult-to-validate “black box” end-to-end systems to more modular approaches. Vlad notes that large-scale simulation is key to building confidence in a system’s decision-making properties across countless scenarios.

[13:58] Analyzing Tesla’s Large-Scale Fleet Approach

Vlad analyzes Tesla’s pioneering but “brute force” approach, which relies on a massive fleet for data collection and validation. He points out its core limitation: as the system improves, the rate of development slows down because finding relevant data for improvement becomes exponentially harder.

[18:20] How Helm AI’s Simulation Platform Helps OEMs Compete

Vlad explains that Helm AI’s generative simulation platform offers OEMs a way to catch up without needing a massive fleet. AI simulation provides a more scalable, cost-effective, and faster way to validate systems by focusing compute power on generating rare and interesting corner cases.

[23:18] The Origins of Helm AI’s Unsupervised Learning Technology

Vlad recounts how Helm AI developed its unsupervised learning technology, “Deep Teaching,” in-house starting in 2016. The technology was born from the realization that supervised learning alone could never solve Level 4 autonomy and was rooted in research from applied mathematics and compressive sensing.

[25:20] Detecting Edge Cases like Debris and Construction Zones

Vlad details how AI systems can handle unexpected obstacles, either through generic obstacle detection or more advanced open-vocabulary models that can identify specific objects. He emphasizes that generative simulation is crucial for training and validating these rare scenarios.

[30:30] The Evolving Relationship Between OEMs and Tech Partners

The conversation explores the trend of automakers, after initially trying to develop autonomy in-house, now recognizing the need for external partnerships. Vlad explains that this shift creates a “win-win” for OEMs to accelerate timelines and reduce costs by working with specialized suppliers.

[35:40] The Next Decade for the Autonomous Vehicle Market 

Vlad predicts continued industry consolidation, a clear focus on AI-first strategies, and a pivot toward high-end ADAS. He foresees a rapid evolution from Autopilot-like systems to truly L4-capable products, accelerated by AI simulation technologies.

[37:40] Helm AI’s Vision: Becoming the Independent AI Software Supplier

Vlad states Helm AI’s goal is to be the leading independent company that builds and supplies the critical AI software for scalable autonomous driving and robotics. He stresses the importance of being an independent partner that can serve many different OEMs.

Watch the Full Episode of The Road to Autonomy

Full Episode Transcript

Grayson Brulte: Vlad, Gen AI is in all the headlines. What are your thoughts on the current state of Gen AI? 

Vladislav Voroninski: I would say we’re kind of in the golden age of AI generally. And there’s lots of momentum on generative AI. We’ve now seen sort of the power of the relatively general foundation models, and we’re now seeing the development of many different kinds of specialized models, for different fields. I think we’re going to see a lot of different, types, including kind of multimodal models, right? So we’ve seen ones that are applied to text or video, we’re never going to see a lot of, multimodal applications, as well as, applications to, you know, various sciences and technology areas. I think there’s definitely, you know, kind of a lot of hype out there. And also maybe some skeptics, you know, that are kind of looking at, the costs involved, right? And sort of the, you know, the balance sheets versus revenue numbers, right? Kind of during this kind of R&D phase of Gen AI. But You know, and maybe they have some doubts about whether it will be properly commercialized, but I really think that that’s kind of, you know, missing the forest for the trees. I think that really the key two things that need to be addressed nevertheless are, certain aspects of compute efficiency, as well as kind of what’s the proper way to commercialize these models. But I think there are really good answers to both those questions.

Grayson Brulte: What do you feel is the best way to commercialize these models? Cause you’re right. You read all these analyst reports. You look at their, their CapEx spending that are in the tens of billions of dollars and the payback period. So what is the best way you’re putting to commercialize it? 

Vladislav Voroninski: I think that there’s going to be two different types of commercialization models. , there’s kind of, again, the very general, foundation models like LLMs, right. Or maybe, , say, you know, certain types of artistic applications of generative AI. I think that, you know, those obviously are, those are maybe the early movers in terms of commercialization. But they’re not necessarily going to capture the most value, because there’s going to be a fair bit of commoditization, right? So if if you’re kind of endpoint product is kind of close to being completed the second that you’ve trained a large model, and, you know, the barrier to entry is not necessarily that high, just need a lot of data, a lot of compute. I think there will be a certain amount of commoditization that will occur. I mean, it’ll lead to. Useful things, but, there’s going to be kind of a lot of competition for many different players. I think that where we’re going to see, a lot of really interesting commercialization that can capture a lot of value is more specialized foundation models, , that cater to the needs of certain specific fields and really couple with the kinds of modes that you can build in particular industries. I definitely think autonomous driving is an example of that. But there are other examples, maybe in the sciences, for example, in biology. Protein folding is kind of one, one interesting example where, just generally, if you’re coupling highly specialized insights about certain industries. With foundation models. I think you’ll be well positioned to capture a lot of value. So I’m not saying there’s kind of a right and right or wrong way, but that’s kind of my projection for where those different approaches will go.

Grayson Brulte: How are specialized foundation models developed? So you mentioned biology on one hand and autonomous driving another, two completely different aspects of the, of the equation. How are financial models developed for those different use cases? 

Vladislav Voroninski: , I think you have to start with asking sort of what’s the critical problem that needs to be solved, in in those fields, right? Or pick not saying there’s only one, but you kind of start with a very difficult technology problem. And, you see how, some of these tools can help. So, you know, with autonomous driving, I think, the key challenges have to do with kind of addressing the tail end of corner cases and also just validation, in terms of, , being able to certify that autonomous driving system operates at a certain level of safety. And, you know, kind of have an interpretable a way to look at that and not just quote a quote a number in terms of the average, you know, safety, the average safety, but in terms of interpretability across different cases. So, where generative AI can really help there is through AI based simulation, where it can really close the gap between kind of real data and simulated data. That really opens the door to a much more scalable way of, not only improving the neural networks to solve the talent of corner cases, but also, creating, you know, highly diverse, very large validation sets that can, Allow you to validate a system much more quickly than, and much more efficiently cost efficiently than you could with only leveraging a real fleet in other domains. I mean, I mentioned biology, right? I mean, so, yeah, I mean, protein folding problem is obviously a very important problem. I’m not an expert in that area. It’s more of a more of a curiosity for me, but, you know, there are data sets that have to do with . all the proteins that have been structured and you can use those data sets to actually build a foundation to basically, you know, leveraging many, many, you know, decades of research, and things that we’ve learned about proteins to actually build foundation models that can kind of short circuit, problems, solutions to certain problems. And can lead to entirely new paradigms development. So, I mean, yeah, again, those are those might look like different things and they are different things. But I think that, those are examples of, you know, much more specialized foundation models than sort of your typical, thing that I think most people are aware of. Your Chat GPTs s. and other types of general foundation models. I think We’ll find many use cases, for, for any given individual. But what I’m talking about are foundation models that can enable, enable you to solve very difficult technology problems, , that might not otherwise be possible that will have, you know, a lot of downstream value to individuals, but they’re not going to be things that are going to be used necessarily by kind of the average person, right? So we’ll only see. For the average person will see the value from that kind of trickling down.

Grayson Brulte: How much data is needed for AI based simulation models? Do you need to have massive amounts of data, or can you do it with limited amounts of data today? 

Vladislav Voroninski: Theoretically speaking, right? You can consider sort of what it what it what it takes for a human to learn right in terms of how much data we observe ., So a teenager can, learn how to drive within a matter of weeks, You know, so, and if you do the math, it’s not really that much data that we’re exposed to in our lives. So, I mean, theoretically speaking, you don’t need that much data. Now, of course, there’s a sizable gap between sort of the rate of learning of the human visual cortex and today’s kind of modern machine learning systems. But, there’s whatever that factor is. I mean, there’s definitely some big factor there, but I wouldn’t say that, data is really the bottleneck, right? In a sense that I mean, in terms of video data, for example, right? And text data. I mean, we have, quite a lot of that, right? And there’s ways to create more data, using simulation. So, , I think that, yeah, I don’t really think data per se is the bottleneck. If you’re taking If you’re taking the right approach, right? I mean, there’s always going to be a spectrum of how, different companies choose to attack these problems, right? There’s definitely a lot of variance, right? In, compute efficiency, right? Or in terms of how much accuracy you get per dollar, right? Per dollar spent on compute. There’s a lot of variance on that. I think that’s a very important thing to pay attention to . but I don’t think that it’s as much of a of a data bottleneck per se, just given the amount of data we have available to us. That’s actually quite broad. So, it depends on how you’re attacking these problems, right? So, for example, for autonomous driving, I would say that if you’re only working with driving data, You’re you’re missing out, right? You should really be working with, you know, kind of Internet scale video data, , at large, right? In order to learn all the different things you have to learn about, and then bring that knowledge to the autonomous driving domain. This is very similar to how humans learn how to drive. I mean, we don’t we don’t learn how to drive by sitting in a car and observing, you know, a billion miles or something, right? Like, obviously, we’re leveraging a lot of other experience.

Grayson Brulte: Where are the bottlenecks? Where do you see those bottlenecks currently? 

Vladislav Voroninski: There are kind of so I’ll list out the different things that you need, right? So there’s the software piece. There’s the hardware piece. There’s validation and there is regular regulation, right? So, with, the kind of approach that we take with at Helm AI and also, you know, Where you can leverage, generative AI is how to essentially build the optimal software stack, quickly and efficiently for any particular hardware constraints . so I think that that problem, to a large extent, we know how to solve right? And there’s a way to do that. That kind of unifies development from, L2 through L4 right? So we don’t have to look at those two problems separately . and, obviously on the hardware side, there are certain cost considerations, right? Just how much compute can you afford to put on the car? And, you know, what kind of sensor stack can you, can you afford to, to actually include as part of a, you know, package that you’re selling? And, once you’ve defined sort of the hardware for the system, that there’s going to be some kind of theoretical limit for what the software can achieve on on that configuration, right? But, yeah, I mean, it’s a very important and difficult problem to be able to then optimally build a software stack for that configuration. And. That’s, , that’s the problem that we’re that we’re addressing at Helm, the way that the hardware will evolve is a function of, those different industries, right? And, in terms of validation, right? It comes down to sort of where do you get your data set from, right? Is your data set coming from a very large fleet where you’re, you know, just gathering a ton of data? I mean. That’s like a pretty inefficient way to do validation, right? And that’s going to be a bottleneck sort of asked as this engagement rates improve, right? It’s going to take longer and longer to actually validate, systems that way and the other approaches via simulation. So, , I think that that’s where, , we’re seeing quite a bit of progress with, generative AI. And I think. That problem will be, you know, essentially addressed, using such methods. And then lastly, there is the regulatory aspect, right? So I think that, I think the biggest hurdle is regulatory, because essentially what you’re talking about is What’s the proper definition of a successful product that is a fully autonomous driving system? I think as we’ve seen, right, there’s, it’s, it’s not a simple, it’s not a simple thing to just say something like this system is safer than an average human driver, therefore, it’s a successful for problem L4 product. I think that there’s going to be, you know, there needs to be some kind of regulatory framework for, addressing those questions. And I think that, you know, different governments across the world are taking different approaches to that, but it’s incredibly, incredibly important to address that. Right? Because, ultimately, that bar of the average human driver is actually a very low bar, right? Because that includes distracted driving, drunk driving things that, you know, are very problematic. And and really, an autonomous driving system that you want to scale to be L4 capable, should be, you know, maybe even 100 times safer than that, right? And furthermore, it needs to be interpretable, right? We need to be able to understand. If the system made a certain decision, why did it make that decision? That’s, that’s going to be pretty important for, dealing with, you know, dealing with the way these systems are deployed in terms of safety. I would say the biggest bottleneck is regulatory on on the AI software side. Again, I think that it’s now possible to basically build the optimal software for a given hardware configuration and also . validate that system in a cost efficient way.

Grayson Brulte: How do you validate the decision making? Is it a process? Does it go through? What does that look like? 

Vladislav Voroninski: There’s different ways of doing that. , so you can look at it from the perspective of, , kind of modularity, right? So there’s, that’s kind of one thing that can help you, right? So if you take, for example, on the extreme end. An end to end system, right? That goes from, let’s say, the RGB data or the raw sensor data all the way to the controls of the vehicle. You’re basically skipping, you know, all the way to the decision. Those systems are, you know, much more difficult to validate, right? Because, you’re basically dealing with a complete black box, right? then there’s ways to actually break up the stack into different pieces. There’s your, Perception piece in time prediction, path planning on. Then you basically do the controls, right? So that gives you more information about what’s happening in between and gives you some level of control. So you can actually take, for example, a perception stack and, you know, map out all the semantics that you want to understand and, you know, further validate that system to where you’re very confident, sort of where each piece is going to fit. And you can do that, you know, throughout the entire stack. Now, I don’t think modularity completely solves the issue, right? Because There’s still there’s still AI involved in between, right? So there’s gonna be, neural networks that are, effectively making decisions, maybe at a lower level, , in between as part of that stack, and you kind of have to get your hands on, what exactly those neural networks might be thinking. So there are approaches to essentially get information about, where neural networks are paying attention, right as they’re making decisions. You can actually, try to engineer systems that explicitly try to tell you sort of what is the system thinking. Of course, the systems have to be themselves validated. There’s a whole kind of framework for, safety certification, , that you have to follow, and, how that ties into interpretability for decision making is still is still, I would say, an evolving field. But in terms of generating, the kind of data that you would need to validate the system, I think that’s where, a lot of progress is being made. And, you know, if you can basically put a system through, a very large scale, simulation that puts it on all the different, scenarios and sort of certify that’s able to handle all those scenarios, it gives you quite a bit more confidence in terms of the decision making properties of that system.

Grayson Brulte: What are your thoughts on Tesla’s approach to cracking full self driving? 

Vladislav Voroninski: So I think Tesla, has definitely been a pioneer in sort of taking the kind of large scale fleet approach to autonomous driving, right? So they started, many years ago, right? As far back as I’m 2014 or something like that, , with the approach of basically, you know, unifying their stack across their entire fleet, the fully vertically integrated approach. And, , betting on sort of betting on autonomy in a big way, , putting a lot of hardware on these vehicles , sensors and compute that allow them to kind of like easily collect the data, that they need from those fleets and kind of really scaling that up. And I would say, you know, taking all that data and then crunching it through, their internal process to actually improve the machine learning models. So. I think that that’s an impressive setup in that it’s, now a large scale fleet that functions and that kind of, through that kind of feedback cycle. Of improvement, but it’s still, you know, relatively brute force approach, in the sense that, you’re basically talking about, you know, relying critically on that fleet. And so, you know, Elon Musk tweets about this, right? I think there was a tweet recently where he said something like, the main bottleneck to improvement is actually, like, how long it takes to validate improvement. So basically, , as the disengagement rates improve. You need to kind of wait longer for the fleet to correct, collect more interesting data or relevant data in order to, either improve or certify the system. So that’s not actually a good property to have, right? , because. That means essentially that your rate of development is actually going to slow down as you approach, better disengagement rates. So that’s, that’s basically like your, you know, classical, tail end of autonomy, sort of rearing its ugly head, right? And so I think that, it’s important to, , recognize that that approach, , has certain limitations. And that’s where I think, generative AI can really, come and can kind of change the paradigm entirely. Right? Because for, Tesla is kind of unique in that they’ve really bet on this, large scale fleet approach. And, have been doing that for many years. So, you know, other automakers are not positioned. If they wanted to compete with Tesla using that same approach, they’re not necessarily positioned to do so because they don’t necessarily have that fleet. , but there’s now a new paradigm in which you can attack the problem that doesn’t doesn’t necessarily require, , that very large fleet to make a lot of progress.

Grayson Brulte: As Tesla slows down just because of the curve, could they enter gen AI into that? Or are they so far down the track some? Their fleet approach that they would not be able to do that.

Vladislav Voroninski: No, I think nothing prevents them from leveraging Generative AI at all. It’s more about what can everybody else do. , I think that, The point is that generative AI can sort of, , level the playing field in a certain sense, right? So, if it were the case where everybody had to take the same approach as Tesla, they, you know, everybody had to basically, has to build this kind of fleet that is kind of unified, unified across its stack, and, scale up to millions of vehicles and have this kind of, continuous feedback loop. Then I think Tesla is sort of like years ahead o of everybody else. But the point is that in the last, I would say, I don’t know, 12 to 18 months, it’s become clear that there is an alternative route, , and it’s via AI based simulation because you can now actually close the gap between simulation and real data. So that’s that’s a very new development, right? So, simulation has always been used in various ways, but, I would say it’s been quite limited. Okay. In its application, because ultimately, any simulators that we’ve seen before AI based simulation, we’re really, they were not really rooted in even training on real data, right? They were kind of just mathematical models and, essentially, what that means is that there’s always going to be some gap. And so from both a training perspective and a validation perspective, It had a very limited value, and now all of a sudden there’s this big sort of quantum quantum jump to where, okay, so if you can actually simulate, the fleet data, right, I mean, instead of having a real fleet, you have essentially a virtual fleet that totally changes the landscape, right? That changes the game. So, yeah, I mean, I don’t think that it precludes, I don’t think Tesla’s, has any limitation in terms of what they can do with generative But I think very importantly, we have to pay attention to now what everybody else is going to do.

Grayson Brulte: Can Helm’s AI sim platform help the traditional OEMs catch up? 

Vladislav Voroninski: Yeah, 100%. And, that’s that’s exactly what we’re doing. So we’re basically, what we offer right is both the real time, full stack software that actually goes on the cars, right? So that includes perception and some prediction and path planning software, , that runs in real time. We also offer, , basically a suite of foundation models for, the development and validation of autonomous driving systems, and that includes, generatives, generative simulation. And the aim there is to basically close the gap between, what OEMs can get from, , real fleet data and what they can get from, AI simulated data. Because. There are a lot of benefits to this, for example, as I mentioned, sort of this scalability question, right? If you have a, if you have a real fleet, you’re basically, you have this kind of exponential fall off in the occurrence of corner cases, right? So as your system improves, what becomes relevant data for that system to improve further becomes less and less frequent. At an exponential rate, that means that you’re effectively paying exponentially more for each new improvement, Whereas with AI based simulation, it’s effectively linear because you cannot focus the system always on generating interesting cases, right? You can avoid entirely. Simulating boring data, right? Whereas you cannot do that in the real world . and so, there’s that scalability advantage . and it also comes down to kind of a time. You know, you basically can do it in a lot shorter of a time span, right? If you need to go and validate, , an autonomous driving system and you have to go send cars out there for a certain period of time. You know, it could take quite a while. Whereas with the simulation, you just basically, if you have the right software, you can throw a lot of compute at it and generate the data that you need pretty quickly. So you can actually accelerate time to market substantially. It’s also less liability ultimately. Right. So I think, putting a self driving system on the road . that isn’t, you know, that’s kind of a constant state of improvement. And, you know, I think that there’s, , there’s a spectrum of risk that different companies take on that front, right? But ultimately, there is the liability aspect of doing so, right? So if you can basically generate all of your important data for where your your system might fail and you can learn how to improve. You can do all that in simulation. That’s a substantial advantage in terms of, not having to actually, take any risk, and deploying those fleets.

Grayson Brulte: How much compute’s needed to scale the Helm AI system? Or does it just depend on what your customers want to do with it? 

Vladislav Voroninski: On the compute side, we have, yeah, so it’s actually a very important question, right? So, , if you take, , your kind of vanilla generative AI technologies, And you, try to scale those up. , I think a lot of people talk about scaling these systems, right? It can be definitely quite compute intensive. And, you know, where we found is that actually by, , combining our unsupervised learning technology that we call deep teaching. With, , kind of additional innovation and generate architectures, we’re able to, , effectively get better scaling laws so we can actually get, more accuracy per compute dollars spent. Then, kind of like the more typical Generative AI systems. As I mentioned before, there’s, there’s always going to be a spectrum of compute efficiency for all these different approaches. And, you know, we’ve learned a lot from both, the kind of optimizations you need to do for real time deployment, of autonomous driving systems. Take those lessons, , to kind of scaling, offline models. But beyond that, just kind of applying, our technology for unsupervised learning we’ve been developing for many years now ., it turns out that that enables us to build essentially a more capital efficient version of generative AI. For autonomous driving, and, , and not only autonomous driving, actually, I would say, even for for other markets as well. I think those kinds of examples are important in the sense that, that came from, you know, certain specific needs, right of the autonomous driving industry, and we’re now able to leverage those to build, highly, compute specialized foundation models, but it could also have applications to, you know, to other fields. So, yeah, I do think that’s like a very important, , question because you have companies that are just kind of throwing quite a lot of compute with these problems, and then they’re able to showcase certain capabilities, but when push comes to shove, there’s going to be a commercialization question, and obviously, the more compute efficient your systems are, The better positioned you are to, offer competitive pricing and also generate a better profit for yourself.

Grayson Brulte: I would say you’re offering a balanced approach. That’s grounded in economics. On the unsupervised learning lane, how did you develop that? 

Vladislav Voroninski: That’s a great question. We started working on supervised learning very early on back in, 2016. That was really kind of the 1 of the 1st questions, , that I asked was, what’s going to be our advantage in the space? Right? And kind of extrapolating out, to what it’s going to require to actually scale all 4. It was very clear that, L4 was never going to scale using . supervised machine learning only. And so there was this kind of glaring open problem right in the space. And I think I generally at the time, , that needed to be solved. And, you know, one of the core strengths of Helm is kind of the, you know, our R and D capabilities. We kind of, focused entirely just for the first two years on addressing that problem. And, you know, we used, various insights from, research and applied mathematics, and the area that I worked in previously called compressive sensing. So. It was, , I would say kind of a different, different take on on a I that I think maybe was, being explored as much at the time. But, yeah, it was really kind of developed fully in house, at Helm and, essentially leveraged mathematical modeling in ways that were. More informative and more flexible,, than other competing methods. So, and, you know, we use that, technology to actually build the world’s first foundation model for semantic segmentation. So it’s basically, understanding what every pixel in an image means and actually scaling that. Right. So, you know, we trained on quite a lot of data for, for that time, like up to a hundred million images. You know, back in like 2017 or something like that, and that’s how we got, very accurate perception models that were better than the state of the art. And, we brought that, brought those to our, OEM customers, and that’s how we got our foot in the door and how we scored some of the major partnerships, , including the partnership with Honda. But, I would say it was definitely, rooted in, that research background that some of us had and applied mathematics.

Grayson Brulte: How does the Helm AI system detect or approach decision making such as a construction zone or a downed tree in the road or an object if you want to use the term edge case that normally wouldn’t be there? 

Vladislav Voroninski: So there’s there’s a number of ways. One approach is to basically have, a generic obstacle detection system, right? So you can actually train a machine learning model to detect potential obstacles on the road, , the same way that you would train, a model to, detect a specific object, right? So that is a, well posed problem, from a machine learning perspective. So you can actually train such systems and, They’ll essentially surface for you potential obstacles. And, you know, that information can be, further validated by either other sensors or, generic depth, machine learning, , systems, right? So, in terms of predicting depth from from vision, for example, and, you know, beyond that, I think where we’re going to see things go, in the medium to longer term is machine learning models that actually have. A fairly comprehensive understanding of the different types of objects, so kind of an open vocabulary type of machine learning model where it actually, knows, all the objects that you can possibly enumerate, right? It actually can classify those, , and tell you that’s actually a construction zone. That’s a tree. That’s a ladder that fell off a vehicle, et cetera. So that’s, definitely totally within scope, right of, machine learning models . but yeah, this is actually where generative simulation comes in very handy, right? Because those situations are just going to be extremely rare in, actual fleet deployments. But, the objects that we’re talking about, anything that can fall on the road, even if it can be a rare thing to happen, you can get plenty of data, right? About those different objects from different situations, from other types of video data. So from a simulation perspective, you can, if you can properly leverage those data sources and simulate those scenarios in a highly realistic way, then you basically say yourself, yeah, I mean, you get something that’s much more scalable in terms of addressing those corner cases.

Grayson Brulte: How do you teach the system? I know it’s unsupervised, but I’m curious now. How do you teach the system to understand hand gestures if somebody’s waving you or their flag in the road to tell you to go into this lane instead of that lane? 

Vladislav Voroninski: There are a couple of ways to do that. So for example, those hand gestures, They have some meaning, in terms of, how we interpret those hand gestures, as a function of sort of our mental model of other humans. Right. And what are they seeing from their perspective, what are sort of, colloquial, means that we have communicating certain things. Right. So we’re kind of leveraging some prior history that shared history that, two people might have. And then there’s the actual environment that you’re in, right? So, I mean, in terms of the environment that you’re in, that’s your standard sort of autonomous driving task of interpreting the environment, understanding what everything is. There’s sort of, you know, your position in that environment, your state. There’s a state of the other person, let’s say, right? And machine learning model can understand both and understand the relationship between both. in terms of that, you know, shared history. Of what hand gestures might mean, you know, that’s another example of where you want to learn from, data sets that are not just autonomous driving data sets. You can’t possibly learn about, , all the different hand gestures, that might come up by just leveraging fleet data. I mean, That’s yeah, you might get like some, some bad data there, I think. So I think like there you would want to leverage, probably multimodal foundation models, that understand video data. They understand actions. Maybe they have text descriptions for those actions, and you can basically figure out, what, you know, what associate what people are doing with, what are they actually intending to communicate. Right? So that’s definitely a totally tractable problem. And then I think you would combine that with, some amount of, structured data. You might produce that specific to your application, as well as, you know, simulations thereof. So, yeah, I mean, it’s a totally, totally tractable problem to solve. , I’m not trying to trivialize it, but I don’t think that there’s anything you can do. About that problem that cannot be approached properly from a modern perspective.

Grayson Brulte: Helm AI is advanced. You’re far down the road of what you’re developing. You’re developing a model that is highly scaled by a little term used earlier, the golden age of AI. You have a specialty in this. You have a, you have a mathematical background. You have a really great team of scientists that working with you. You have the relationship publicly with Honda overall OEM wide. When do OEMs realize that they need individuals and companies like Helm to help them achieve autonomy because you have the specialties and expertise where they don’t have it necessarily 

Vladislav Voroninski: I think that that is already happening, right? Of course, every company has its own kind of arc, right? Their own trajectory. But, , I with the example I provided earlier, right? Back in, 2017, we provided technology demonstrations to OEMs. That showed clearly our advantage over other, you know, other competing solutions and, that was well received right there. That got quite a lot of interest from the various automakers. And we definitely got, partnerships out of that. I think that. You know, you’re right that, there it’s, it’s, , it’s not as simple. It’s not as simple as. Okay. If you have the best tech, therefore, everyone will work with you. I think that there are other, questions that come into play in terms of just how certain companies are oriented decisions they’ve already made in the past. Just in terms of being able to properly communicate the information, et cetera. But, yeah, I mean, I think that the industry is, is definitely very adaptive, right? It might take, you know, it might take a lot of time, but there’s definitely a trend that I think, where OEMs are leveling up substantially in their understanding of autonomous driving technology, right? So, in the early days, I think there was a lot of concern from automakers that. L four was right around the corner. I mean, it’s a sort of both opportunity and concern, but I think it was more concern because, if that’s the case, then, they might end up being sort of like the commoditized component of that solution. And I think it led to a lot of investments. There’s a lot of, a lot of hype out there about L four timelines that just didn’t come to fruition. And, you know, I think to anybody, technical. It should have been relatively clear that that was not going to come to fruition, but, nevertheless, it led to all these different investments that didn’t, didn’t really pan out and we’re just, we’ve seen, we’ve seen that play out. Right. But what that’s resulted in is, I mean, it’s an important lesson, right? For, for everybody in the industry, including automakers, and we’re now seeing automakers essentially pivot to kind of high end ADAS commercialization. As the sweet spot of where they want to focus and yeah, we’ve seen that now in many cases And also oems where initially they wanted to do it fully in house and you know Okay, we’re going to do everything ourselves We’re going to hire all these people and deploy billions of dollars and then years down the road Oh, it turns out not as much progress was made as we thought They had hoped, and there’s no public statements from those OEMs that, hey, we actually need partnerships. So, I mean, look, I think that, , it’s, it’s happening. , I think that there is, there’s always some kind, it’s not going to happen instantaneously, but I think it’s definitely now going in the right direction. So, I don’t, yeah, I don’t see that as, nearly as much of an issue as in the past.

Grayson Brulte: Do you see oems you’re right about the focus on the a desk the high end a desk that’s mistakes and consumers are going to want Level four as we’re kind of seeing certain polling data Do they get to the point in your opinion where they approach Helm and your competitors and license a stack? From you instead of spending the billions of dollars in, in house developments, because as you and I both know, as public reported, they’ve already made that decision and didn’t work out well, 

Vladislav Voroninski: Yeah, absolutely. I mean, I, I think that, OEMs will be well served by partnering externally with the right suppliers, whether it be to accelerate their time to market, or just to reduce their costs, right? There’s going to be some inherent, R and D cost to developing these technologies. And there’s also kind of the uncertainty timeline wise, right? If you’re, if you’re an automaker and you. Are trying to attack these problems. , there’s going to be that question mark around the timeline that, you know, is also a big concern . so if there’s an external party, right, an independent company that’s already built all the critical software and can offer it to you at a better price than what you can afford to build it for yourself. It’s it’s like a win win situation, right? , so that’s kind of how I see it. Is that, , what we’re trying to build is basically the best option for automakers in terms of, helping them catch up to and potentially leapfrog companies like Tesla. By partnering externally. , I, I don’t think that it’s going to be, such a binary thing where an automaker will just say, okay, we’re doing everything in house and we’ll never work with anybody externally. I mean, of course, it’s on us to offer them the right kind of product, that, provides a better option. But I think we’re well on our way.

Grayson Brulte: Do we ever get to a point where there’s a Windows, Apple OS, Linux of autonomous driving systems? 

Vladislav Voroninski: Yeah, that’s a good question. I think that, maybe for certain like layers of the stack, , but I don’t know. I don’t know. But but but in terms of like the most differentiated. Aspects of it. I mean, I think every automaker, will want to have obviously software differentiation, right? And so that is the key. That is the key thing. So in the coming years, vehicle sales, one of the biggest factor in vehicle sales, will be software differentiation. And that’s also where the margins come from, So from that perspective, every automaker I think is incentivized to produce autonomous driving features that have, certain advantages over other systems. And I do think that they’re all positioned to, potentially produce something like that, provided that they partner externally, with the right suppliers. And so, the way I think about it is, we build foundation models, we build, real time, ,real time software that’s geared toward production. But all those things can be optimized for any particular automaker, And that’s what allows them to actually get software differentiation. Well, I think there could be a world in which there are maybe multiple systems that are shared across different OEMs. I think there will definitely be an incentive for customization, , based on kind of every automaker’s strategy for, what kind of vehicles they want to sell to their, customers.

Grayson Brulte: You’re seeing it today. They don’t want Apple to take over the car or Android to take over the car. They’re all GM’s trying to develop their own system and other OEMs as well. Vlad over the next decade, how do you see the autonomous vehicle market evolving? 

Vladislav Voroninski: I think that there’s, , going to be kind of like some continued consolidation, right? Where, I mean, we already saw a pretty, , Extensive period of consolidation, but I think there’s going to be a bit more. , I think the, you know, there’s still like. I think that the companies that went after kind of pure play L4 are going to have a very difficult time staying independent . so I think that’s, you know, that’s definitely area. We’re going to see consolidation. We’re already seeing this kind of pivot to. a more ADAS centric approach, from the automakers . definitely, an AI first strategy, makes the most sense. I think that is now very clear. So, in a lot of ways, I think that, you know, what’s happened in the last 10 years, is, you can basically see it as like a reset in some ways, right? I think what happens in the next 10 years is far more relevant than what happened in the last 10 years. And, yeah, I think we’re going to see, substantial commercialization of high data software, as automakers, try to catch up to and potentially surpass Tesla. We’re going to see an evolution of these autonomous driving systems where they go from. More like autopilot type, systems to L4 capable autopilot systems, and then actually L4, , truly L4 systems. Right? So, and that’s really more of a product, question than a technology question in some sense. I think that technologically we’re going to see, you know, Continued improvement, of autonomous driving software to where, you know, certainly will go beyond sort of, you know, being safer than the average human driver and all that. But from a product perspective, there has to be a lot of clarity about, what it means to actually, offer an L4 product I think we are going to see quite a bit of, , acceleration in the development of these systems and a validation of these systems using, developments in AI such as generative simulation. And, yeah, I mean, I think that, you know, essentially the future is bright when it comes to autonomous driving and also, adjacent areas and robotics.

Grayson Brulte: You’re taking the tagline away from me. The future is bright. The future is autonomous. What role do you want Helm AI to play in that market? 

Vladislav Voroninski: So I mean, this has been the goal from the very beginning. We’ve, and we’ve structured the entire strategy around this goal. We want to be an independent company, truly independent company that builds and supplies all the critical AI software required for scalable autonomous driving. And beyond that, , we also want to do the same for other areas like, you know, essentially robotics. Robotics at large. And, you know, we already have actually, traction and some of these robotics markets, you know, for example, in the mining market. But there are, of course, many other applications in both industrial and consumer robotics . so, yeah, I think that’s that’s one important thing, though, is, yeah, I mean, it has to be an independent company, right? In the sense that, that’s also why we didn’t go after, you know, pure play L4 mean, there were many other reasons, but that’s another reason. And I think the industry will be well served by having an independent party that builds the critical software and can offer that software, to many different, many different OEMs, because there are there are advantages to that right from, you know, both kind of time to market and cost perspective. So I think that yeah, there’s that. Kind of gap in industry that we’d like to fill.

Grayson Brulte: I like that. You have ambition to stay independent. I like that. You’re going to move into robotics. You’re going to stay independent. You’re going to help autonomy scale and you’re going to help potentially humanoid robots. I’m getting ahead of myself there. Helm has a very bright future. I’m excited to watch and see where you go is it I’m sure I’ll guarantee it and let him go on the record and say I’m sure I’m gonna be impressed. Vlad as we look to wrap up the conversation for today. What would you like our viewers and listeners to take away with them? 

Vladislav Voroninski: What I’d like viewers to take away is that, , despite sort of, , all the different ups and downs of the autonomous driving industry that we’ve seen over the last 10 years. , I think it’s now, , a very important. Moment for the space where there’s kind of it’s a bit of a moment of clarity, I think it’s going to be, maybe even surprising for people to see the amount of progress that happens in the coming years . so, yeah, I guess, as I said before, I think the, you know, the future is bright for autonomous driving and, especially generally applications, , to self driving cars, robotics and beyond.

Grayson Brulte: The gen AI and unsupervised learning what Helm is building is ushering in the future of autonomy The future is bright the future autonomous. The future is Helm AI, Vlad thank you so much for coming on the road to autonomy today 

Vladislav Voroninski: Thanks for having me.

Key The Road to Autonomy Episode Questions Answered