A Fascinating Look at a Groundbreaking AI Startup

Intro 00:01
Welcome to Startup Success, the podcast for startup founders and investors. Here, you’ll find stories of success from others in the trenches as they work to scale some of the fastest growing startups in the world, stories that will help you in your own journey. Startup Success starts now.

Kate 00:17
Welcome to startup success today. We have Vladislav Voroninski in studio, who is the CEO and co-founder of Helm AI. Welcome Vlad.

Vlad 00:27
Thanks for having me, Kate.

Kate 00:29
We’re excited to talk to you, because you’re doing so much in the AI space. I want to get into your background in a couple minutes, but I think it would be best if you set the stage and gave us a brief overview of Helm first, and then we can talk about, you know, the inspiration for founding it and your background.

Vlad 00:53
Yeah, absolutely. Helm is an AI software company focusing on autonomous driving and robotics. We’ve been around for about eight years, and we made a decision early on to focus on high-end assistant driving as the initial go-to-market on the path to fully autonomous driving L4. And so we essentially build full-stack real time AI systems for autonomous driving that we license to the various automakers, and we also provide essentially development and validation tools powered by foundation models for tasks like kind of labeling, auto labeling, very large amounts of data, as well as AI-based simulation. We have made a lot of progress in terms of being an automotive supplier and currently on track for production in a partnership with Honda.

Kate 01:48
Congratulations. So I read a lot about what you accomplished, and it’s super impressive, especially this is a space that I think a lot of consumers are very interested in around AI, because they get it, right, self driving. So now, if you wouldn’t mind, give us a little bit about your background and what led you to this undertaking, because this is no small feat.

Vlad 02:16
Yeah, absolutely. I actually first got interested in self driving cars and AI and computer vision when I was an undergrad. So I was part of the UCLA Computer Vision Lab while they were competing at DARPA Grand Challenges. So the DARPA Grand Challenges were these competitions sponsored by the Department of Defense to basically benchmark autonomous driving technology. Of course, at that point, it was very nascent, you know, it was, you know, barely even scratching the surface of, really, the AI component of self-driving cars, but I found that to be an incredibly exciting application. And so that kind of sparked my interest. And you know, I did kind of two years of, you know, very intensive, you know, research into computer vision and kind of adjacent areas. And, you know, that was really kind of my first scientific passion, so to say. And it was very formative. But the conclusion for me was that actually it was too early to really jump into that space. At that point, the foundational technologies were not yet really there, but I did identify that applied mathematics was essentially the key to understanding AI better. And, you know, I figured that focusing on that would be the best way to make sure that I’m actually going to be able to come back to this space later on and not kind of take an early path in the wrong direction. So I was always intending to come back to the autonomous driving space and computer vision in particular. And so I spent, it took a little longer than I expected, but I spent about a decade in academia as a mathematician. So I did a math PhD at Berkeley, and then I was on the postdoc faculty at MIT for a couple of years. And, you know, I was always very interested in starting a company. Kind of knew that I would start a company from probably around age 20 or so. So I kind of, you know, spent my 20s thinking through various business ideas and never really had anything that I was passionate about enough to really kind of drop out of my academic track until I got to about the end of my postdoc, and, you know, I essentially I could either be a professor or can go start a company. Those were my two paths. But it became pretty clear quickly that really I should start a company. And you know, that’s what I’m really interested in. And actually, during my postdoc as well, I was actually a co-founding Chief Scientist of a company called Sift Security, that cybersecurity for the cloud, and it got acquired by Netskope. So I got a pretty good taste of entrepreneurship at those very early stages. I was there from, you know, before they even raised any money. So that gave me some nice experience that I applied toward Helm ai And so I guess, yeah, that kind of led to a situation where the technology was really kind of getting there in terms of deep learning taking off. I definitely felt like my scientific capabilities were kind of mature to where I could easily go and kind of branch off and run my own research program, essentially. And the industry was also maturing where there was quite a lot of interest in autonomous driving around that kind of 2015, 2016 time period. So it was kind of the optimal convergence of different factors that led me to start the company.

Kate 05:39
Impressive that you saw early on that it wasn’t the right time yet, and you went and did some really fascinating things in their own right, and built some experience here, and then saw the opportunity come back up and know that you always wanted to start your own business. That’s pretty impressive. So taking that, Okay, it’s time to take that first step. Like, what did that look like? Did you like … we have other founders on here, and they bring in a small team. They like, what were your first steps to get Helm kind of off the ground?

Vlad 06:18
Yeah. So this is a pretty interesting process actually. It’s kind of multi-scale in a sense, right? I mean, I think that about 2015, 2016 I think I had a very distinct realization that I definitely want to start a company. And then I think actually what came first was my kind of commitment to working in AI. That really was the thing that I was very sure about, that I had a very distinct kind of moment where I was like, Okay, the next 10 years I’m dedicating to working in AI full stop, like, I’m definitely doing that. And I knew I was going to do that in industry, not in academia. And it was kind of a mix of things. I mean, there were various events happening in the industry, you know, certain companies got funded. You know, like Zoox had some early funding. Then Cruise got some funding. You know, Cruise got acquired by GM for a billion. I would say that it became very clear that the autonomous driving industry was, you know, taking off. But also that there was still a very big opportunity because actually, the major challenges were still very much unsolved, like it was not to scale at that point. I think that is very different today. I think it’s now ready to scale. But in those early days, that wasn’t the case. So, yeah, I mean, I would say that it was kind of a lot of observation about this space, you know, kind of ideation about, Okay, what are going to be the key advantages to secure long term. And I guess, like once I started formulating those things, then indeed, I started to engage with kind of potential early team members as well as potential early investors. And, yeah, I mean, those early days are definitely much more chaotic, right? Where you’re you could sort of be, you can change direction kind of very rapidly. But actually, one of the first things that I did was basically work with my co-founder to, in the early days, to basically build a prototype. So yeah, because, I mean, you know, is one of those things where definitely wanted to not just kind of go out there and raise funding just based on like resumes or something like that, or verify that we had an advantage in the space. And so, you know, basically built an unsupervised learning pipeline that you know, had kind of like some proof points, some early proof points, that we could actually solve that problem. So that’s one important thing that we identified was basically that unsupervised learning had to be addressed in order to actually solve self-driving cars. And it was completely unaddressed at that point, like, if you know, in terms of what was out there in the literature, people were doing, you know, I think companies were trying to scale kind of purely supervised machine learning approaches, which means they were severely bottlenecked on human annotation, kind of really labeled data, which is very expensive, doesn’t really scale to the kind of data set sizes that you need. And I would say at that time, you know, saying something like, okay, we’re actually going to solve unsupervised learning to secure the key advantage, you know, I think that was actually a pretty ambitious goal at the time, but, yeah, you know, we’re able to make significant progress. And you know, that definitely leveraged my research background from academia. So I think there was definitely some kind of situation where we were leveraging tools from applied mathematics that were not typically being used in the AI space at the time to build these unsupervised learning pipelines. And so, yeah, I would say that was really kind of the key thing that we focused on early on, and then after two or three months of that, it became clear that we actually had something that we can go and pursue, the actual kind of fundraising and team building process.

Kate 10:08
There’s so many things that you shared that jumped out at me that really distinguishes what you’re doing. Like how you noticed the scaling wasn’t happening, how you wanted to build a prototype, how you focused on the unsupervised learning, the competitive advantage, like how you would differentiate yourself. I think for everybody listening, kind of how you walked us through that was fascinating. It really shows probably why you stand out now. So with that, give us a deep dive into Helm, like, talk to us more about this unsupervised learning. I think that’s so key. And like, where you’re different than everybody else. You know, we’ve all read about some of the bigger players out there, but if you could walk us through that, that would be really helpful.

Vlad 10:57
Yeah, absolutely. So just to kind of also give a little bit more of a backdrop in terms of context. You know, the reason that you need unsupervised learning and kind of very large scale learning to solve self driving cars is because there are many, many situations that can happen while driving that are actually very, you know, very rare.

Kate 11:18
Like, unique, right? Like, it’s in that moment.

Vlad 11:21
Yeah. And you want to be able to be prepared for any of those situations. And really, you kind of need a combination of things. Not only do you need unsupervised learning, you also need essentially AI based simulation. So kind of applying generative AI in conjunction. But you know, to start off with, if you just have a lot of real data that you’ve observed, you want to be able to learn from all that data, and not only from driving data, but also, say, from, you know, internet scale data that’s out there, and there’s just way too much data to actually label by hand, right? And so the typical machine learning pipeline for supervised learning, basically means you just have a human perform the task and provide that to you as annotation, and then you have a neural network emulate that capability using that data. And if you’re doing unsupervised learning, you don’t have the labels. So that means that inherently, you have to replace that with something else, and that something else has to be some kind of mathematical modeling that you’re injecting into your deep learning pipeline, and so that’s really kind of what we focused on is manufacturing the appropriate kinds of mathematical models that are both sufficiently informative about the task so they actually allow you to to learn a lot from every, say, image that you put into the to the system. And these models need to be sufficiently easy to actually create, right? Because, I mean, an example of mathematical models that are too burdensome is something like traditional simulators, right? So traditional simulators in some sense are trying to manufacture data that you could, in theory, use to train on. But there’s two issues there. I mean, first of all, there’s a big gap between traditional simulation quality and real data. So for machine learning purposes, it’s not really going to transfer well. And second of all, it’s just very arduous to build. I mean traditional simulators, mean they’re, you know, multi billion dollar efforts to actually build those systems, right? Whereas, you know, clearly, we have to do something very different than that. And so that’s where certain expertise from the research that I was doing and in applied mathematics came in very handy. There’s an area called compressive sensing that I was working in for many years, which effectively has to do with solving certain kinds of inverse problems, and the imaging sciences using a lot less data than you might expect by exploiting the structure of the data that you’re working with. So, you know, we had a team of mathematicians essentially developing the core IP in the early days, and we were able to create these kinds of models that allowed us to build unsupervised learning pipelines that actually scaled very well.And so we actually were able to, for example, train the world’s first foundation model for semantic segmentation, which is a difficult computer vision problem that essentially involves understanding what every pixel means in an image, right, including all of the boundaries of all the different objects, kind of localizing everything very precisely. And we were able to train our models on something like 100 million images back in like 2017 which for that time period, that’s a huge amount of data. Clearly, we were not doing that using any significant amount of annotation, because we just simply wouldn’t be able to afford that. And the result was a much more robust perception stack for autonomous driving purposes than was currently available on the market. And that’s really allowed us to kind of get our foot in the door with the various automakers as potential partners. So that’s kind of the unsupervised learning piece. And the other piece of it has to do with addressing the fact that, you know, data from these rare corner cases, right is just so difficult to get right. So if you’re driving around, your machine learning models will actually want a lot of data from every situation, right? But certain situations are so rare that you’re just not really getting much data even if you have a large fleet, like a very large fleet, will still not get you some of these really rare corner cases. Where AI based simulation becomes very important, and you know, what’s changed in the last couple of years is, you know, with advances in generative AI, it’s become possible to actually close the gap between AI based simulation and real data. So we can actually create, you know, AI based data that a human can’t tell is AI generated. (That’s pretty cool.) Yeah, and mathematically also good enough to actually train these models on the AI generated data to get the results that you want. And so we’ve done actually, is we took that unsupervised learning technology that we call Deep Teaching that we’ve been developing for many years, but actually combined it with innovation on generative AI architectures that we also pursue internally to create a more scalable version of generative AI. So basically get you know by combining these technologies, we actually get better scaling laws, which mean for a certain amount of compute, we get higher accuracy in the generative models, and we’re now deploying that in a pretty wide way to the autonomous driving space, in terms of AI based simulation. So creating kind of camera data, LIDAR data, modeling these things simultaneously, or using camera data to simulate LIDAR data, you know, to actually create models that can effectively do the driving right, actually predict what the what the ego vehicles should do and what the all the other agents are going to do next. Yeah, so there’s kind of a pretty significant inflection point currently in the use of generative AI for AI based simulation to really kind of make autonomous driving, development and validation highly scalable.

Kate 17:28
Wow, that is a significant development. That’s so neat. I mean, I can see where that would make a huge difference with autonomous driving, right? That’s what everyone was afraid of back in the early days. So I know, Honda made this significant investment. Congratulations. Can you talk a little bit about that and what initiatives you’re working on?

Vlad 17:53
Yeah. I mean, I have to be a bit careful talking about that.

Kate 17:54
Yeah, I know. Or maybe just another use case or something, yes.

Vlad 18:00
I mean, I think that in terms of what’s out there that’s, you know, publicly announced, right? Yeah, there’s a new electric vehicle that Honda is developing – the Zero series. And, you know, the initial production year is 2026. It’s a very exciting vehicle that they’re working on, and they’ve announced us as a core AI technology provider for that production program. And, yeah, I mean, I think that, you know, we, generally speaking, our goal is to accelerate the roadmap for, you know, any particular automaker in terms of deploying these kinds of high-end assisted driving systems. So in Honda’s case, they’re targeting actually all three capabilities. So it’s a very, you know, it’s a very high level of autonomy. And in order to do that right, we’re providing both, generally speaking, as a product, the real time software that actually runs on the vehicle. So things like, you know, perception, intent prediction, path planning, so kind of understanding all the sensor data, kind of where everything is, where might all the different agents go next, right, the vehicles and the pedestrians. And then lastly, what should the ego vehicle do in order to complete its goal? But equally as importantly, we’re providing these foundation models you know, that have been trained using unsupervised learning, and also provide AI based simulation capabilities in order to, you know, essentially, accelerate the training and validation process for self driving. And you know, importantly, this is actually we have a unified approach for both high end assistant driving and fully autonomous driving. That means that once we’re provided with the hardware, the camera system, the other sensors involved, the compute stack and the operational domain where it’s supposed to operate, then our development process to create the optimal software stack is the same, regardless of sort of whether it’s an L2 system or an L4 system. And that’s a very important property, because, you know, there are approaches to fully autonomous driving, like L4 where there’s no steering wheel, that don’t actually scale down to high-end assisted driving. Yeah, if they’re too highly reliant on maps or LIDARs or teleoperation, and simultaneously, there are approaches to assisted driving that don’t scale up to fully autonomous driving. If you’re relying too much on unsupervised learning, if you don’t know how to address the rare corner cases at the tail end. And so importantly, we have an approach that actually unifies development for those actually different – kind of pretty different – product categories.

Kate 20:53
That is fascinating. Can’t believe we’re at time, because I could ask you so many more questions. But what Helm is doing is so impressive. The fact that you have achieved what you have so far already is really remarkable.

Vlad 21:13
Yeah, thanks. I mean, we’re definitely really excited about the progress. And, you know, I mean, it has been eight years. Things are, things are definitely heading in the right direction,

Kate 21:22
I mean, and the way you explained it all, it really puts in perspective all the different inputs that happen with self driving and where you’re playing a role. It’s very exciting. We always wrap up the show with just, you know, any general advice you could give other startup founders listening that just from your experience would be helpful?

Vlad 21:47
Yeah, I mean, I think, and again, this is, I think that, you know, advice that I think can be applied in general, but the way that I think about it is, it’s very important to think about the long term advantages, right? If you’re starting a company, or at any given point of running a company, keeping a very strong viewpoint regarding the asymptotics of where the industry will go. And the reason that’s important is because there will always be trends. Always be kind of what do investors think today? Right? What are they on today versus what potentially could be different, and what is maybe the important thing longer term? If your goal is to really build a long lasting company, I think it’s critical to really stay true to that, having a strong sense of what are going to be the long term advantages, and really focusing on that, even if it’s not the popular thing that’s happening in a given moment, right? And so, you know, in our case, it was the fact that we focused on, we made the decision back in 2017 to not go and pursue pure play, L4 fully autonomous driving, because we did not think that anybody would be able to do that on the timelines that were being projected. Really glad that we made that decision, despite the fact that a lot of people thought we were just flat wrong. Most people thought we were flat wrong about that. And the other thing, I would say is kind of, especially with AI, in my opinion, there are a lot of foundation models being developed today, right? And I think that it’s really important to consider markets where kind of AI technology interfaces in sort of a sophisticated or complex way with the actual, you know, market dynamics, right? In a sense that not something that you want to be working, I think, on foundation models that not anybody could build, because if you’re if you’re working on something that technically anybody could build or anybody could train, you’re going to be competing with all sorts of really big companies, and then you might end up getting into a price war over a somewhat generic capability. So I think that finding the right kind of market to funnel AI technology into, and really leveraging market knowledge or certain unique market advantages is going to be very important for capturing the value of that AI technology. My hypothesis is that, you know, the kind of foundation models we’ll see down the road that’ll capture the most value will be far more specialized and not necessarily these more generic models that we see today in you know that people are more aware of today.

Kate 24:34
That’s so interesting, because I had another VC in the space on here, and he said the same thing, absolutely. And I like how you started with too, not getting caught up in the trends, which I think founders can do, right? You kind of lose sight of what you’re doing in that regard. Super fascinating to talk to you. So interesting. If I’m listening and I want to find more information about Helm, where do I go?

Vlad 25:03
The best way to learn more about Helm AI is just to visit our website Helm.ai. We actually just launched a new website, pretty comprehensive information about our product, about our technology, sort of career information, and press and all that stuff. So definitely, the website’s the best place to go.

Kate 25:24
Excellent. Thank you so much for being here. I really enjoyed the conversation.

Vlad 25:29
Same here. Thanks for having me.

Outro 25:30
You’ve been listening to Startup Success to make sure you don’t miss out on future episodes. Subscribe to the show and your favorite podcast player. Like what you hear? Tap the number of stars you think the show deserves in Apple Podcasts. For more tools and resources for your own startup success, check out burklandassociates.com. Thank you so much for listening. Until next time you.