What I Learned Building AI That Competes with Tesla

In this interview, Vlad Voroninski, CEO of Helm AI, talks about the decade of math and research behind Helm's approach to autonomous driving: spending years in academia before founding the company, two years of pure R&D with no product, and a contrarian bet on partial automation when the rest of the industry was chasing full autonomy. He shares concrete examples of how Helm tested its technology, built trust with automakers, and stuck with that strategy long enough for the market to catch up.

My name is Vlad Voroninski. I'm the CEO of Helm AI. Helm is an AI software company focusing on a unified approach to autonomous driving that goes all the way from L2 plus through fully autonomous driving L4. We have partnerships with companies like Honda and we allow automakers to compete with Tesla by bringing to market cutting edge autonomous driving systems. The end goal is achieving fully autonomous driving and there are a number of technological and commercialization challenges along the way. These days, one of the key areas that we're focusing on is AI based simulation, leveraging generative AI as well as our unsupervised learning IP in order to create a unified approach to solving autonomous driving that unifies both partial automation and full automation. First got interested in self-driving cars and computer vision during undergrad. I was part of the UCLA computer vision lab while they were competing in our grand challenges.

We're in the offhighway vehicle recreation area.

We did our first autonomous path tracking test and I thought that was a very exciting area and was focusing on computer vision at the time but decided to pursue mathematics in academia for about 10 years with the intent to come back to the space when it was more mature because I saw that as the key bottleneck to AI in the sense that the biggest challenge in reading research papers in AI or computer vision was ultimately a mathematical question of do you understand the equations in some sense. I looked at math as a tool to be able to solve AI down the road the goal was always to come back to the autonomous driving space. Around 2016 it became clear that the technology was really taking off in terms of deep learning. There was an inflection point in where AI technology was going with deep learning at the time and simultaneously there was a very clear opportunity that it was the right time to jump into that space because what I witnessed was a lot of companies making certain decisions that I didn't agree with. It was such an inefficient space at the time that it was very clear to me that with the right approach you can add a lot of value.

Because it's a strategy that's stood the test of time as opposed to many companies that peaked early and then died off or ran out of money and what that meant was that there was an opportunity. After my academic career moved back to California to start Helm. Covid was a very tricky time for everyone obviously but also in the automotive market in particular because it caused a halt in production. The Corona virus is idling one auto plant after another. All the different auto factories and all the automakers had to immediately start dealing with that issue versus everything else. And I think it caused a bit of a delay in the deployment of autonomous driving technology. But beyond that, I would say quite exciting. Ultimately, I don't know if it was I think the challenges were there, but they were outweighed by how exciting it was to start a company, make a truly deep tech bet in the space.

Our first 10 hires were basically all just very, very strong researchers. Any one of those people could have easily walked away and done something else.

Some of those people even made certain sacrifices of their academic career to come work at Helm because they were very excited about the vision. And I think that helped us really mold the engineering culture. When we were pitching Helm 2016, unsupervised learning seemed like a pipe dream almost. But we committed to that vision. We carried out the research and we were able to make that work. That was quite exciting. There was the fact that for 2 years we were developing that technology and there was zero product development during that time. It's pure R&D for 2 years very much all or nothing. Obviously there's risk involved in that but it was a very creative time I mostly appreciate it.

Back in 2018. foray into foundation models, even before that term was coined. What we used that foundation model for was to build an autonomous driving system that we were able to show outperforms the systems you were able to buy in the market by pretty wide margin. We conducted a series of tests where we put our autonomous vehicle on very steep and curvy mountain road scenarios where you have to make very rapid driving decisions as far as taking the various turns in a challenging landscape. And we were able to achieve much better disengagement rates up to a factor of 200 better than what was out there on the market. And that is how we got the attention of some of the brand name automakers in the world in the early days. In the last couple of years, we've been doing a lot of innovation in generative AI and combining that with our in-house technology which is called deep teaching in order to close the gap between AI based simulation and reality.

That means how do you simulate driving data or driving footage sensor data from driving without having to get into a car. And there are many advantages to doing that. For example, very large fleets.

They can be useful for collecting data in order to address difficult corner cases for autonomous driving, but the rate of occurrence of those corner cases goes down exponentially as your system improves. And so end up paying exponentially more to gather interesting data as you get further into the development process. It's really not a good property. And what simulation allows you to do is generate all the interesting data without having to deploy a fleet. For example, Tesla that has a very large fleet. Other automakers don't have access necessarily to internal fleets that are that large. Even if they wanted to take the same approach as Tesla, they would not be positioned to do so. The only alternative to doing that is AI based simulation.

And until very recently, it wasn't possible to generate highly realistic simulation data. But that's very much changing these days due to the advent of generative AI and combining generative AI with technologies like deep teaching provides a highly scalable simulation platform that allows you to deploy a virtual fleet instead of a real world fleet you're learning from existing data. That's a recent inflection point that we're definitely proud to be part of and contributing to. Vid Gen 1 and World Gen 1 are foundation models for generative AI simulation. Vid Gen is a foundation model that creates highly realistic video data from a multitude of different cameras, arbitrary cameras, arbitrary locations. Worldgen is a foundation model that takes a further step in that it simulates the entire autonomous driving stack. You can use Worldgen to technically drive a car because it does make predictions about what's going to happen next. If you input data from your autonomous driving stack, it'll tell you what's going to happen in the next several seconds and that includes the path that the vehicle should take to perform certain actions.

You can use it to drive. It's technically a self-driving system that not only functions as a simulator or in a simulator environment, it also can function in the real world.

How to convince a customer, how to develop a relationship with a customer in autonomous driving space. I think there's two important aspects at least. One I think is seeing is believing. Not only having marketing materials or video demos. Being able to put somebody in a car and have that autonomous car navigate unforeseeable situations I think sends a very powerful message about the robustness of the technology and the product and secondly I think that working on a production contract there are going to be other contracts along the way. I think it's not really possible for a major car company to give a production contract to a supplier as the first contract so there's going to be some sequence of contracts along the way. 

Being able to execute on those contracts and deliver exactly what you signed up to deliver is critical because ultimately what you're entrusted with as a supplier is providing not only safety critical technology but technology that is absolutely necessary to have in a certain timeline. It's because you're talking about a production program that has to launch in a particular year where a lot more money is invested into that than just the money being paid to one supplier. It's incredibly important to be able to meet those deadlines. 

Maybe a third thing I would add is not only demonstrating your current state of where your technology is, but demonstrating the difference from one time to another. Being able to show here's where we are at this point in time. And then in a month, we expect to be here. And showing that differential and allowing them to measure not only your position, but also your velocity.

When it comes to technology, especially when you're going after such an ambitious play like autonomous driving, requires quite a lot of conviction. The hardest part about it was if you're a researcher, I think that your key job is to perform the research and then you put it out there. And sure, there is some marketing aspect to that, but it's not nearly as significant, I think, as what you have to do for a company. And generally, as a startup founder, you have to wear so many different hats. People say when they go to MIT, for example, that it's like drinking from a fire hose. I never had that experience, but I would say that starting a company felt like drinking from a fire hose. In our particular case, I think we were in some sense working against the grain in that the vast majority of the funding went toward companies taking a totally different approach in that they were pursuing pure play 4.

When I first started to engage people about the fact that we're going to really focus on partial automation as the key market, everyone thought it was crazy. But I think that it emphasizes more this notion of the importance of having grit because you can't expect what you think to be the predominant world view and it might take a long time for the world to adjust. I think we only started seeing signs of the world adjusting to our point of view some number of years into the company, maybe four years in 5 years in, but then every year our position has improved, prod not only as a function of the technology and the product but also because our strategy was adapted to a certain worldview that we believed would eventually materialize and that is now happening.