What I Learned Building AI That Competes with TeslaㅣHelm.ai, Vlad Voroninski

Name: What I Learned Building AI That Competes with TeslaㅣHelm.ai, Vlad Voroninski
Duration: 12 min 34 s
Description: Vlad Voroninski is the CEO and Founder at Helm.ai, an AI software company pioneering a new approach to self-driving. After spending 10 years as a mathematician,

EO12:34Added Aug 31, 2026

Vlad Voroninski is the CEO and Founder at Helm.ai, an AI software company pioneering a new approach to self-driving. After spending 10 years as a mathematician, he founded Helm.ai in 2016, and has since raised over $100M. He shares invaluable insights on AI-based simulation, the role of generative AI in self-driving, and what it takes to stand out in the competitive deep tech landscape.

Watch on YouTube →

Contributed by 刘嘉琪

Transcript

00:00The end goal is achieving fully
00:02autonomous driving. We allow automakers
00:04to compete with Tesla by bringing to
00:07market cutting edge autonomous driving
00:08systems. We've raised over $und00
00:10million. We'll see very very scalable
00:13commercialization happen. But I would
00:15say that starting a company felt like
00:17drinking from a fire hose. When we were
00:18pitching kind of like seemed like a pipe
00:20dream, pure R&D for 2 years. There was
00:23zero product development during that
00:25time. Everyone thought it was crazy. But
00:27we committed to that vision. We actually
00:29carried out the research and we were
00:31able to make that work very much all or
00:33nothing and it might take a long time
00:35for the world to adjust. My name is Li
00:38Berninski. I'm the CEO of Helm AI. Helm
00:41is an AI software company focusing on a
00:43unified approach to autonomous driving
00:45that goes all the way from L2 plus
00:48through fully autonomous driving L4. We
00:50have partnerships with companies like
00:52Honda and we allow automakers to compete
00:55with Tesla by bringing to market cutting
00:58edge autonomous driving systems. The end
01:00goal is achieving fully autonomous
01:02driving and there are a number of
01:04technological and commercialization
01:05challenges along the way. These days,
01:07one of the key areas that we're focusing
01:09on is AI based simulation, leveraging
01:12generative AI as well as our
01:14unsupervised learning IP in order to
01:16create a unified approach to solving
01:18autonomous driving that essentially
01:20unifies both partial automation and full
01:26automation. First got interested in
01:28self-driving cars and computer vision
01:31during undergrad. So, I was part of the
01:32UCLA computer vision lab while they were
01:35competing in our grand challenges.
01:38We're in the offhighway vehicle
01:40recreation area.
01:45So we just did our first autonomous path
01:47tracking test and I just thought that
01:50was a very exciting area and was
01:52focusing on computer vision at the time
01:54but decided to pursue mathematics in
01:57academia for about 10 years with the
01:59intent to come back to the space when it
02:01was more mature because I saw that as
02:03the key bottleneck to AI in the sense
02:06that the biggest challenge in in reading
02:08research papers in AI or computer vision
02:10was ultimately a mathematical you know
02:13just a question of do you understand
02:14sort of the equations right so yeah in
02:16some sense I looked at math as a tool to
02:19be able to solve AI down the road the
02:21goal was always to come back to the
02:23autonomous driving space in the war
02:25between humans and artificial
02:27intelligence this is for 33-year-old
02:29professional go gamer Lisa D Elon Musk
02:32he plans to equip all new Tesla cars
02:34with the hardware needed for full
02:36self-driving capacity around I would say
02:392016 it became clear that the technology
02:42was really taking off in terms of deep
02:44learning. There was an inflection point
02:46in where AI technology was going with uh
02:48deep learning at the time and
02:51simultaneously there was a very clear
02:53opportunity that it was the right time
02:55to jump into that space because what I
02:58witnessed was a lot of companies making
03:01certain decisions that I actually didn't
03:03agree with. Right? It was such an
03:04inefficient space at the time that it
03:07was very clear to me that with the right
03:09approach you can add a lot of value.
03:11Right? because uh it's a strategy that's
03:13kind of stood the test of time as
03:15opposed to many companies that peaked
03:18early and then died off or just kind of
03:20ran out of money and what that meant was
03:22that there was an opportunity. So after
03:24my academic career moved back to
03:26California to start
03:30Helm Co was a very kind of a tricky time
03:34for everyone obviously but also in in
03:36the automotive market in particular
03:38because it basically caused a halt in
03:40production. The Corona virus is idling
03:43one auto plant after another. All the
03:45different auto factories and all the
03:46automakers had to immediately start
03:49dealing with that issue versus
03:51everything else. And I think it caused a
03:53bit of a delay in the deployment of
03:56autonomous driving technology. But
03:58beyond that, I mean, I would say quite
04:00exciting. I mean, ultimately, I I I
04:01don't know if it was like I think the
04:03challenges were there, but they like
04:04were outweighed by how exciting it was
04:07to like start a company, you know, make
04:09a truly uh deep tech bet in the space.
04:12Our first 10 hires were basically all
04:15just very, very strong researchers. Any
04:17one of those people could have easily
04:19walked away and done something else.
04:21Some of those people even made certain
04:23sacrifices of their academic career to
04:25come work at Helm because they were very
04:27excited about the vision. And I think
04:30that helped us really mold the
04:32engineering culture. When we were
04:34pitching Helm 2016, right, unsupervised
04:36learning was kind of like seemed like a
04:38pipe dream almost, right? Um, but we
04:42committed to that vision. We actually
04:44carried out the research and we were
04:46able to make that work. So that was uh,
04:48you know, quite exciting. So I mean I
04:50guess like there was the fact that you
04:52know for 2 years we were essentially
04:53developing that technology and there was
04:55zero product development during that
04:57time. So it's pure R&D for 2 years very
05:00much all or nothing. Um so you know uh
05:05obviously there's risk involved in that
05:07but it was a very creative time so I I
05:09mostly just appreciate
05:14it. an Uber self-driving vehicle that
05:18flipped. Uber is now banned from testing
05:20its self-driving cars in Arizona.
05:22Tonight, Tesla confirming this car was
05:24in autopilot mode when it crashed in
05:26Northern California
05:29back in 2018. Kind of foray into
05:32foundation models, even before that term
05:35was coined. What we used that foundation
05:37model for was to essentially build an
05:39autonomous driving system that we were
05:41able to show actually outperforms the
05:44systems you were able to buy in the
05:45market by pretty wide margin. So we
05:48essentially conducted a series of tests
05:49where we put our autonomous vehicle on
05:52uh very steep and curvy mountain road
05:54scenarios where essentially you have to
05:56make very rapid driving decisions as far
05:59as you know taking the various turns in
06:01a challenging uh landscape. And we were
06:03able to achieve much better
06:04disengagement rates uh up to a factor of
06:07200 better than what was out there on
06:10the market. And that is how we got the
06:12attention of some of the brand name
06:14automakers in the world in the early
06:16days. In the last couple of years, we've
06:19been doing a lot of innovation in
06:21generative AI and combining that with
06:23our in-house technology which is called
06:25deep teaching in order to close the gap
06:28between AI based simulation and reality.
06:31So essentially that means how do you
06:33simulate driving data or driving footage
06:36sensor data from driving without
06:37actually having to get into a car. And
06:39there are many advantages to doing that.
06:41For example, very large fleets, right?
06:43They can be useful for collecting data
06:46in order to address difficult corner
06:48cases for autonomous driving, but the
06:50rate of occurrence of those corner cases
06:52basically goes down exponentially as
06:54your system improves. And so end up
06:57actually paying exponentially more to
06:59gather interesting data as you get
07:01further into the development process. So
07:02it's really not a good property. And
07:04what simulation allows you to do is
07:07generate all the interesting data
07:08without actually having to deploy a
07:10fleet. For example, Tesla that has a
07:12very large fleet. Other automakers don't
07:14have access necessarily to internal
07:17fleets that are that large. So even if
07:19they wanted to take the same approach as
07:20Tesla, they would not be positioned to
07:22do so. The only alternative to doing
07:24that is essentially AI based simulation.
07:27And until very recently, it wasn't
07:30possible to generate highly realistic
07:32simulation data. But that's very much
07:34changing these days due to the advent of
07:36generative AI and combining generative
07:38AI with technologies like deep teaching
07:40provides a highly scalable uh simulation
07:43platform that allows you to essentially
07:45deploy a virtual fleet so to say instead
07:48of a real world fleet you're just uh
07:50learning from existing data. So that's a
07:53recent inflection point that we're
07:54definitely you know proud to be part of
07:56and contributing to. Vid Gen 1 and World
07:59Gen 1 are foundation models for
08:02generative AI simulation. Vid Gen is a
08:04foundation model that creates highly
08:06realistic video data from a multitude of
08:09different cameras, essentially arbitrary
08:11cameras, arbitrary locations. Worldgen
08:14is a foundation model that takes a
08:17further step in that it actually
08:18simulates the entire autonomous driving
08:20stack. You can actually use Worldgen to
08:22technically to drive a car because it
08:24does make predictions about what's going
08:26to happen next. So if you input data
08:29from your autonomous driving stack,
08:30it'll tell you what's going to happen in
08:32the next several seconds and that
08:34includes the path that the vehicle
08:36should take to perform certain actions.
08:38You can actually use it to drive. So
08:40it's technically uh you know a
08:42self-driving system that not only
08:44functions as a simulator or in a
08:45simulator environment, it also can
08:47function in the real world.
08:52Yeah, basically how to actually convince
08:54a customer, how to develop a
08:56relationship with a customer in
08:58autonomous driving space. Um, I think
09:00there's kind of two important aspects at
09:01least. So, one I think is, you know,
09:03seeing is believing. Not only having
09:06kind of marketing materials or video
09:08demos, right? actually being able to put
09:10somebody in a car and have that
09:12autonomous car navigate unforeseeable
09:14situations um I think is a very powerful
09:16sends a very powerful message about the
09:18robustness of the technology and the
09:20product and secondly I think that
09:23working on a production contract there
09:25are going to be kind of other contracts
09:27along the way right I think it's not
09:29really possible for a major car company
09:31to give a production contract to a
09:33supplier as the first contract so
09:35there's going to be some sequence of
09:37contracts along the way and so I think
09:38that being able to execute on those
09:41contracts and deliver exactly what you
09:43signed up to deliver is critical because
09:46ultimately what you're entrusted with as
09:48a supplier is providing not only safety
09:51critical technology but technology that
09:53is absolutely necessary to have in a
09:56certain timeline because you're talking
09:58about a production program that has to
10:00launch in a particular year where you
10:02know a lot of money is a lot more money
10:03is invested into that than just the
10:05money being paid to one supplier. So,
10:07it's incredibly important to be able to
10:09meet those deadlines. Maybe a third
10:11thing I would add is just not only
10:13demonstrating your current state of
10:15where your technology is, but
10:17demonstrating the the difference from
10:20one time to another, right? So, being
10:21able to show, okay, here's where we are
10:23at this point in time. And then in a
10:26month, we expect to be here, right? and
10:28actually showing, you know, showing that
10:30differential and allowing them to
10:32measure not only kind of your position,
10:34but also your velocity, so to say. When
10:36it comes to technology
10:40development, especially when you're
10:42going after such an ambitious play like
10:44autonomous driving, kind of requires
10:46quite a lot of conviction. The hardest
10:48part about it was just if you're a
10:50researcher, I think that your key job is
10:52really okay to to perform the research
10:54and then you put it out there. And sure,
10:56there is some marketing aspect to that,
10:58but it's not nearly as significant, I
11:00think, as what you have to do for a
11:02company. And generally, as a startup
11:04founder, you have to wear so many
11:05different hats. People say when they go
11:07to MIT, for example, that it's kind of
11:10like drinking from a fire hose. I never
11:12had that experience, but I would say
11:14that starting a company felt like
11:16drinking from a fire hose. You know, in
11:18our particular case, I think we were in
11:20some sense working against the grain in
11:22that the vast majority of the funding
11:24went toward companies taking a totally
11:26different approach in that they were
11:28pursuing pure play 4. And you know, when
11:30I first started to engage people about
11:33the fact that we're going to really
11:34focus on partial automation as the key
11:36market, you know, everyone thought it
11:38was crazy. But I I think that it
11:40basically just emphasizes more kind of
11:43this this notion of the importance of
11:45like having grit or something because
11:48you can't expect what you think to be
11:50the predominant world view and it might
11:52take a long time for the the world to
11:55adjust right like I think we only
11:57started seeing signs of the world
11:59adjusting to our point of view some
12:01number of years into the company right
12:03maybe four years in 5 years in right but
12:06then every year the our position has
12:08improved prod not only as a function of
12:11the technology and the product but also
12:13because our strategy was adapted to a
12:17certain worldview that we believed would
12:19essentially would eventually materialize
12:21and that is now happening.
12:24[Music]