WEBVTT
00:00:00.160 --> 00:00:02.240
So we're team uh lidar and radar.
00:00:03.200 --> 00:00:04.320
LIDAR, radar, camera.
00:00:04.639 --> 00:00:14.560
It's not to say that like a camera-only system can't, you know, get the human relative safety, but when you talk about these orders of magnitude improvement, I think a lot of that actually comes from the complementary nature of a lot of these sensors.
00:00:14.960 --> 00:00:16.559
The bar for performance, right?
00:00:16.640 --> 00:00:19.600
At the human level is somewhat arbitrary.
00:00:19.839 --> 00:00:30.160
It makes sense as a minimum, but making sensor choices based on that bar, I think, is probably just the wrong framing of the whole problem.
00:00:30.559 --> 00:00:32.079
This is Unsupervised.
00:00:32.159 --> 00:00:34.320
I'm Dave Ferguson, co-founder of Neuro.
00:00:34.799 --> 00:00:39.840
Join me as we turn over the big questions in autonomy and physical AI.
00:00:42.799 --> 00:00:48.880
Today I'm joined by my good friend Boris Soffman, CEO and founder of Bedrock Robotics, and someone that I've known for a long time.
00:00:49.039 --> 00:00:50.640
Boris, welcome to Unsupervised.
00:00:50.960 --> 00:00:51.280
Thanks, Dave.
00:00:51.439 --> 00:00:51.840
It's a pleasure.
00:00:51.920 --> 00:00:52.719
Yeah, we go way back.
00:00:52.880 --> 00:00:53.520
Yeah, yeah.
00:00:53.600 --> 00:00:55.600
I was thinking back to when we first met.
00:00:55.920 --> 00:00:57.679
Back in the basements of Carnegie Mellon.
00:00:57.759 --> 00:01:00.960
You walk around the halls of CMU and you can't help but like trip over a robot.
00:01:01.119 --> 00:01:06.480
I remember getting really excited about um Tony was driving and started doing research as an undergrad in my junior year.
00:01:06.560 --> 00:01:15.840
And Dave, I was like your undergrad, basically working on the stuff that I thought was incredibly cool, but I really clearly now know in hindsight that you guys really didn't want to do most of most cool bars.
00:01:16.079 --> 00:01:17.680
We still had you doing the cool stuff.
00:01:18.239 --> 00:01:21.599
And uh I thought it was the most amazing uh kind of time in the world.
00:01:21.920 --> 00:01:26.879
Genuinely, I think you had an incredibly meaningful impact on my entire journey and my uh career.
00:01:27.040 --> 00:01:28.879
And so it's uh really grateful for that.
00:01:29.439 --> 00:01:30.959
Oh, thank you, Boris.
00:01:33.599 --> 00:01:40.000
So you were at Waymo for a while on the trucking program, leading the trucking program, uh, and then you left to create bedrock.
00:01:40.319 --> 00:01:40.799
Talk us through that.
00:01:41.040 --> 00:01:47.200
So the catalyst was a time at Waymo where there was a very intentional transition to embracing these more modern learning approaches.
00:01:47.439 --> 00:01:55.040
We started to see these like subsidies where you needed less and less incremental data in order to jump to what seemed like very different applications or even platforms.
00:01:55.200 --> 00:02:00.480
And so we thought like we can apply this sort of technology to really uh automating heavy machinery.
00:02:00.719 --> 00:02:05.840
There are the workhorse behind a good chunk of the GDP in the country, uh, but the labor shortages are quite stark.
00:02:06.000 --> 00:02:07.359
And so the physics just don't line up.
00:02:07.439 --> 00:02:10.719
You have supply that just can't meet demand, supply wins.
00:02:10.879 --> 00:02:16.800
And so you have this radiation on the whole space where costs go up, the school doesn't get built, the roads don't get replaced.
00:02:16.960 --> 00:02:22.560
Um, and so, like when we talk to our partners, this isn't about replacing a job.
00:02:22.639 --> 00:02:24.800
This is about taking on work that would be turned out.
00:02:24.879 --> 00:02:30.159
It's about utilizing their equipment more heavily and then actually starting to surge to meet that demand.
00:02:30.479 --> 00:02:39.199
When I think about mobility as an industry, it's not trying to have every existing ride on the Uber network be by a self-driving car.
00:02:39.439 --> 00:02:50.719
There will still be an enormous demand for human-driven rides that will be alongside all of the self-driving rides that we will have as a society.
00:02:50.879 --> 00:03:03.599
But overall, as we can get to this future where we have the improvement in safety, we have the improvement of cost, and we can make a personally owned vehicle drive itself for basically lower cost than you driving yourself.
00:03:03.919 --> 00:03:04.960
Safer, it's more rustful.
00:03:05.840 --> 00:03:09.599
It's going to completely expand the market by orders of magnitude.
00:03:09.759 --> 00:03:18.319
And as a result, there's going to be more labor associated with mobility than there is today, and it's going to be a dramatically better product.
00:03:18.800 --> 00:03:25.039
We've seen this with like every technological boom, the industrial revolution, where people were rightfully worried about like what's the impact.
00:03:25.199 --> 00:03:26.319
Is jobs just going to disappear?
00:03:26.479 --> 00:03:27.840
Is going to be massive unemployment?
00:03:28.080 --> 00:03:31.039
And uh locally there were kind of transitions in jobs.
00:03:31.120 --> 00:03:37.599
When the dust settled, the uh productivity not surprisingly skyrocketed by just astronomical multiples.
00:03:37.840 --> 00:03:40.560
The number of jobs actually doubled.
00:03:40.639 --> 00:03:47.280
And then the average salary actually went up meaningfully because the training and the productivity justified the you know kind of extra spend.
00:03:47.360 --> 00:03:52.800
And so it's uh it's an interesting, you know, dynamic where it's always easy to see the immediate disruption.
00:03:52.960 --> 00:03:58.879
It's always really hard to imagine like how does the quality of life and the productivity of society increase on the other side.
00:03:59.199 --> 00:04:00.800
And the local disruptions matter, right?
00:04:00.960 --> 00:04:03.599
Like, I think it's really important that we try to get it all right.
00:04:03.919 --> 00:04:04.159
That's right.
00:04:15.520 --> 00:04:19.600
Particularly when it comes to the physical world, I think it just takes time to roll these things out.
00:04:19.680 --> 00:04:25.920
And it's our obligation to make sure that there's like a very natural kind of a handoff in evolution where nobody's left stranded in the middle of it.
00:04:26.160 --> 00:04:27.839
Yeah, and I think it's very, very doable.
00:04:27.920 --> 00:04:31.600
We just need to be we need to be thoughtful as we get this out there.
00:04:32.079 --> 00:04:33.439
Yeah, absolutely.
00:04:35.120 --> 00:04:42.160
We've been through in the last few years this massive inflection and a huge amount of progress on the digital AI side.
00:04:42.399 --> 00:04:44.959
And you know, this term physical AI has become popular.
00:04:45.360 --> 00:04:46.000
It's popular now, yeah.
00:04:46.319 --> 00:04:49.680
It feels like there is another sort of wave of of interest.
00:04:50.160 --> 00:04:56.319
You are perhaps acting upon some of the largest physical instantiations of AI that are out there.
00:04:56.560 --> 00:05:11.040
One of the questions that we often get is uh look, Chat GPT can basically replace me in terms of my job and my relationships, and why is it still that there seems to be a pretty significant gap on the physical AI side?
00:05:11.120 --> 00:05:13.040
Like, why is the physical world different?
00:05:13.199 --> 00:05:14.959
Why is it harder potentially?
00:05:15.360 --> 00:05:17.279
Yeah, this is a great uh topic.
00:05:17.439 --> 00:05:19.519
It's very clear that this is the next frontier.
00:05:19.680 --> 00:05:28.399
And there has been this energy behind the kind of generalizable models and you know, seeing what's happened with, like you said, with OpenAI and uh Anthropic and Gemini and all these approaches.
00:05:28.639 --> 00:05:30.079
This is fundamentally different.
00:05:30.160 --> 00:05:36.480
And you know, on the digital side, you have this benefit of language being a common feature across a lot of these applications.
00:05:36.639 --> 00:05:47.279
You have almost infinite data that's available on the internet and books and and journals, and there's an opportunity to like really scale in how you use that volume of data to create truly generalizable models.
00:05:47.439 --> 00:06:06.959
When you get into the physical world, the data doesn't exist, simulation's not at a point where you can just imagine all of it, and you really do need the volume of experience to deal with the long tail safety cases, and you have to go and not only figure out how to you know get that signal in order to learn you know these little patterns, but you also have for our types of applications huge safety bars.
00:06:07.040 --> 00:06:09.600
Like we're talking about giant machines.
00:06:09.759 --> 00:06:14.480
You see a lot of physical AI companies kind of brute forcing it with you know teleoperations.
00:06:14.720 --> 00:06:22.240
You know, I had a t-shirt uh back when I was at at Google before I started Neuro that said the revolution will not be teleoperated.
00:06:22.399 --> 00:06:28.160
I think it was I think it was one of the last robotics institute t-shirts that I got, but absolutely still still believe that.
00:06:28.639 --> 00:06:31.519
Yeah, the latest trend is teleoperation gets called semi-autonomous.
00:06:32.240 --> 00:06:37.680
So this would be the equivalent of OpenAI saying I'm gonna go hire a team of writers to go and train ChatGPT.
00:06:37.759 --> 00:06:43.680
It just it's almost like so insignificant relative to the scale of data that they've used that you know it doesn't compute.
00:06:43.920 --> 00:06:54.319
To me, it feels like there's a story that is easy to resonate with when these things aren't factored in, of like the the general foundation model for everything physical and every robot.
00:06:54.480 --> 00:06:56.560
But there's like physical challenges to this.
00:06:56.879 --> 00:07:04.000
Yeah, I think the same way is if you take someone who's incredibly good and has driven on multiple different platforms and you put them in a new one.
00:07:04.079 --> 00:07:14.480
Like if you or I, maybe you probably better, start driving an excavator, you're gonna have very strong like prior knowledge around driving and the basics of driving.
00:07:14.639 --> 00:07:22.560
There is going to be a significant period of additional learning that you're going to need to do to nail that and to get comfortable driving that.
00:07:22.720 --> 00:07:45.680
And then there is almost certainly going to be a massive additional period that you or whoever is managing you uh in a commercial operation is going to need to get comfortable that you're now performing at a level that passes all of the validation and that they can say, hey, we can guarantee or commit that Boris from a safety perspective is stronger than any of the other drivers.
00:07:45.920 --> 00:07:57.120
But for us, certainly the shared similarities between like a robotaxi that is carrying people and perhaps vehicles like the ones we've had a lot of experience with that are just carrying packages, right?
00:07:57.360 --> 00:08:07.279
And underneath all of it is obviously this shared driver AI intelligence, and then on top of that, there's all of the additional optimizations that you make for the different platforms.
00:08:07.519 --> 00:08:15.120
But the incremental data, but also validation and retraining and learnings become smaller and smaller.
00:08:15.439 --> 00:08:26.319
When all of us started working on autonomous driving, I think we were all like just excited about how cool the problem was, and it's like the holy grail of uh you know robotics in a in a lot of ways.
00:08:26.399 --> 00:08:31.360
But it was very easy to underestimate how hard the absolute long tail of that challenge ends up being.
00:08:31.519 --> 00:08:43.039
But in hindsight, there were like a couple of waves of kind of technological progress that really unlocked what now feels like this like upcoming kind of golden age of ability to solve these complex physical problems.
00:08:43.279 --> 00:08:53.440
Even though you are never able to get to absolute, you know, zero collisions of risk, we're starting to see that these technologies save lives and not just by a little bit, by uh by a lot.
00:08:53.519 --> 00:08:54.879
And that's actually quite exciting.
00:08:55.039 --> 00:09:06.000
So we know that you know 40,000 people die each year in the United States in traffic accidents, and construction's the number one most dangerous job in terms of injury and death in the country, even today.
00:09:06.080 --> 00:09:14.240
And so there's like genuine like public good that can happen, not even counting the productivity and time savings and everything, just in the safety aspect.
00:09:17.120 --> 00:09:21.759
Bedrocks go to market is taking these massive machines and retrofitting them.
00:09:21.919 --> 00:09:26.879
We've seen obviously there are many different players and approaches you can take in this space.
00:09:27.120 --> 00:09:38.240
From an efficiency angle, obviously, integration at the manufacturer is sort of the ideal, but that can require everything from partnerships to supply chains and whatnot that may be challenging.
00:09:38.320 --> 00:09:40.799
So, how did you guys land on that as a strategy?
00:09:41.519 --> 00:09:52.639
We studied the machines in this space, and we're able to safely go driverless with a retrofit solution because these machines have trended towards an electric architecture that is already drive by wire.
00:09:53.120 --> 00:09:58.240
And what you're integrating are sort of the full camera LIDAR, radar, sensor suite?
00:09:58.639 --> 00:10:00.159
It's uh camera and LIDAR systems.
00:10:00.480 --> 00:10:13.759
Yeah, you know, I think that the addition of the centers, camera, LIDAR, for us, also radar as well as the compute into sort of stock vehicle is something that we've seen an enormous improvement over the last 10 years.
00:10:14.000 --> 00:10:28.000
So now the NeuroDriver hardware that we integrate, so the one we're integrating into our lucid platform with Uber, as well as some other partners to provide full L4, the performance of that system and the cost has improved so much.
00:10:28.240 --> 00:10:30.559
And you know, we would love to claim credit for that.
00:10:30.720 --> 00:10:38.799
Most of it is that these are automotive components that have benefited from massive scale across consumer automotive.
00:10:39.200 --> 00:10:43.279
And the cost is now at a point where we can integrate this into a personally owned vehicle.
00:10:43.600 --> 00:10:54.320
So we truly have uh this opportunity to build personally owned fully autonomous vehicles, which I think all of us have been sort of dreaming about for many decades now.
00:10:54.399 --> 00:10:58.480
And so that is part of what keeps us so, so excited about this space in general.
00:10:58.960 --> 00:11:22.720
Yeah, we we talk a lot about the advances in AI and machine learning, but um that's just as important where when you started with Waymo in the early days of this push, a lot of the sensors had to get invented because they just didn't have the appropriate versions and the scale of effort that that drove in both time and cost, uh it's a is a gift that today um we can tackle these sort of problems much more quickly.
00:11:23.039 --> 00:11:28.320
Yeah, and I think that that's where we start to see the unlock across all the different industries and applications.
00:11:28.399 --> 00:11:30.559
And yeah, it's it's gonna be really, really cool.
00:11:30.960 --> 00:11:32.799
The next 10 years are gonna be amazing.
00:11:33.039 --> 00:11:33.120
Yeah.
00:11:33.360 --> 00:11:37.679
So we're team uh LIDAR and radar uh LIDAR, radar camera.
00:11:37.840 --> 00:11:39.279
But you mentioned you don't use radar.
00:11:39.600 --> 00:11:40.240
We don't right now.
00:11:40.559 --> 00:11:46.080
You don't have motorcyclists in black at night driving at 100 miles per hour through your construction site.
00:11:46.320 --> 00:11:56.000
I think this is one of the areas where you can rely a little bit on the like regulations of how you have to run a site and just you know, you can fail gracefully when you don't have those conditions.
00:11:56.240 --> 00:12:10.159
When it comes to autonomous driving, 100%, there's signals in these other sensors that are so complementary and to a point where you know you go through the response on every single sensor, and there just isn't a signal in camera that has to be picked up by one of the other uh sensors.
00:12:10.240 --> 00:12:20.799
And so um it's not to say that like a camera-only system can't, you know, get the human relative safety, but when you talk about these orders of magnitude improvement, I think a lot of that actually comes from the complementary nature of a lot of these sensors.
00:12:21.200 --> 00:12:25.840
The bar for performance, right, at the human level is somewhat arbitrary.
00:12:26.080 --> 00:12:36.399
It makes sense as a minimum, but making sensor choices based on that bar, I think, is probably just the wrong framing of the whole problem, right?
00:12:36.799 --> 00:12:41.840
And I think this applies very heavily to the camera versus camera and lidar and radar debate.
00:12:42.399 --> 00:12:54.720
And that's why a driver assist system that's level two is just such a different journey in development than something that's actually going into full driverless where you are 100% exposed to every element and you're expected to be safe and thoughtful.
00:12:55.039 --> 00:12:55.919
Yeah, I totally agree.
00:12:56.000 --> 00:12:58.720
And this has actually been something that we have debated over the years.
00:12:59.039 --> 00:13:05.519
Because is there is a shared foundational technology layer between driver assist versus full autonomy.
00:13:05.919 --> 00:13:11.759
But the end product and how you build it and how you validate and what you're optimizing for is so, so different.
00:13:12.240 --> 00:13:24.799
If you're trying to hedge between level two and level four, you can't afford the type of investment that's easily justified in level four if it's an assist system where it's only partially kind of capable of the sort of functionality that you need.
00:13:24.960 --> 00:13:47.919
And so when you go full level four in almost any of these applications, the value skyrockets so much that you actually get justified in investing enough to reach these sort of safety bars and capability bars, and it becomes, you know, in some ways more self-fulfilling versus trying to fight uphill from a place where you actually are handicapped from a unit economic standpoint, and therefore you have to compromise in a way that's not favorable at all.
00:13:48.159 --> 00:13:48.559
Love it.
00:13:48.720 --> 00:13:49.279
Totally agree.
00:13:49.360 --> 00:13:51.279
Uh Boris, thank you so much for your time.
00:13:51.440 --> 00:13:54.080
It's been uh a privilege uh and really lovely to have you here.
00:13:54.480 --> 00:13:54.960
It's a pleasure.
00:13:55.039 --> 00:13:58.559
It's been a fun multi decade journey with you on these windy roads.
00:00:00.160 --> 00:00:02.240
So we're team uh lidar and radar.
00:00:03.200 --> 00:00:04.320
LIDAR, radar, camera.
00:00:04.639 --> 00:00:14.560
It's not to say that like a camera-only system can't, you know, get the human relative safety, but when you talk about these orders of magnitude improvement, I think a lot of that actually comes from the complementary nature of a lot of these sensors.
00:00:14.960 --> 00:00:16.559
The bar for performance, right?
00:00:16.640 --> 00:00:19.600
At the human level is somewhat arbitrary.
00:00:19.839 --> 00:00:30.160
It makes sense as a minimum, but making sensor choices based on that bar, I think, is probably just the wrong framing of the whole problem.
00:00:30.559 --> 00:00:32.079
This is Unsupervised.
00:00:32.159 --> 00:00:34.320
I'm Dave Ferguson, co-founder of Neuro.
00:00:34.799 --> 00:00:39.840
Join me as we turn over the big questions in autonomy and physical AI.
00:00:42.799 --> 00:00:48.880
Today I'm joined by my good friend Boris Soffman, CEO and founder of Bedrock Robotics, and someone that I've known for a long time.
00:00:49.039 --> 00:00:50.640
Boris, welcome to Unsupervised.
00:00:50.960 --> 00:00:51.280
Thanks, Dave.
00:00:51.439 --> 00:00:51.840
It's a pleasure.
00:00:51.920 --> 00:00:52.719
Yeah, we go way back.
00:00:52.880 --> 00:00:53.520
Yeah, yeah.
00:00:53.600 --> 00:00:55.600
I was thinking back to when we first met.
00:00:55.920 --> 00:00:57.679
Back in the basements of Carnegie Mellon.
00:00:57.759 --> 00:01:00.960
You walk around the halls of CMU and you can't help but like trip over a robot.
00:01:01.119 --> 00:01:06.480
I remember getting really excited about um Tony was driving and started doing research as an undergrad in my junior year.
00:01:06.560 --> 00:01:15.840
And Dave, I was like your undergrad, basically working on the stuff that I thought was incredibly cool, but I really clearly now know in hindsight that you guys really didn't want to do most of most cool bars.
00:01:16.079 --> 00:01:17.680
We still had you doing the cool stuff.
00:01:18.239 --> 00:01:21.599
And uh I thought it was the most amazing uh kind of time in the world.
00:01:21.920 --> 00:01:26.879
Genuinely, I think you had an incredibly meaningful impact on my entire journey and my uh career.
00:01:27.040 --> 00:01:28.879
And so it's uh really grateful for that.
00:01:29.439 --> 00:01:30.959
Oh, thank you, Boris.
00:01:33.599 --> 00:01:40.000
So you were at Waymo for a while on the trucking program, leading the trucking program, uh, and then you left to create bedrock.
00:01:40.319 --> 00:01:40.799
Talk us through that.
00:01:41.040 --> 00:01:47.200
So the catalyst was a time at Waymo where there was a very intentional transition to embracing these more modern learning approaches.
00:01:47.439 --> 00:01:55.040
We started to see these like subsidies where you needed less and less incremental data in order to jump to what seemed like very different applications or even platforms.
00:01:55.200 --> 00:02:00.480
And so we thought like we can apply this sort of technology to really uh automating heavy machinery.
00:02:00.719 --> 00:02:05.840
There are the workhorse behind a good chunk of the GDP in the country, uh, but the labor shortages are quite stark.
00:02:06.000 --> 00:02:07.359
And so the physics just don't line up.
00:02:07.439 --> 00:02:10.719
You have supply that just can't meet demand, supply wins.
00:02:10.879 --> 00:02:16.800
And so you have this radiation on the whole space where costs go up, the school doesn't get built, the roads don't get replaced.
00:02:16.960 --> 00:02:22.560
Um, and so, like when we talk to our partners, this isn't about replacing a job.
00:02:22.639 --> 00:02:24.800
This is about taking on work that would be turned out.
00:02:24.879 --> 00:02:30.159
It's about utilizing their equipment more heavily and then actually starting to surge to meet that demand.
00:02:30.479 --> 00:02:39.199
When I think about mobility as an industry, it's not trying to have every existing ride on the Uber network be by a self-driving car.
00:02:39.439 --> 00:02:50.719
There will still be an enormous demand for human-driven rides that will be alongside all of the self-driving rides that we will have as a society.
00:02:50.879 --> 00:03:03.599
But overall, as we can get to this future where we have the improvement in safety, we have the improvement of cost, and we can make a personally owned vehicle drive itself for basically lower cost than you driving yourself.
00:03:03.919 --> 00:03:04.960
Safer, it's more rustful.
00:03:05.840 --> 00:03:09.599
It's going to completely expand the market by orders of magnitude.
00:03:09.759 --> 00:03:18.319
And as a result, there's going to be more labor associated with mobility than there is today, and it's going to be a dramatically better product.
00:03:18.800 --> 00:03:25.039
We've seen this with like every technological boom, the industrial revolution, where people were rightfully worried about like what's the impact.
00:03:25.199 --> 00:03:26.319
Is jobs just going to disappear?
00:03:26.479 --> 00:03:27.840
Is going to be massive unemployment?
00:03:28.080 --> 00:03:31.039
And uh locally there were kind of transitions in jobs.
00:03:31.120 --> 00:03:37.599
When the dust settled, the uh productivity not surprisingly skyrocketed by just astronomical multiples.
00:03:37.840 --> 00:03:40.560
The number of jobs actually doubled.
00:03:40.639 --> 00:03:47.280
And then the average salary actually went up meaningfully because the training and the productivity justified the you know kind of extra spend.
00:03:47.360 --> 00:03:52.800
And so it's uh it's an interesting, you know, dynamic where it's always easy to see the immediate disruption.
00:03:52.960 --> 00:03:58.879
It's always really hard to imagine like how does the quality of life and the productivity of society increase on the other side.
00:03:59.199 --> 00:04:00.800
And the local disruptions matter, right?
00:04:00.960 --> 00:04:03.599
Like, I think it's really important that we try to get it all right.
00:04:03.919 --> 00:04:04.159
That's right.
00:04:15.520 --> 00:04:19.600
Particularly when it comes to the physical world, I think it just takes time to roll these things out.
00:04:19.680 --> 00:04:25.920
And it's our obligation to make sure that there's like a very natural kind of a handoff in evolution where nobody's left stranded in the middle of it.
00:04:26.160 --> 00:04:27.839
Yeah, and I think it's very, very doable.
00:04:27.920 --> 00:04:31.600
We just need to be we need to be thoughtful as we get this out there.
00:04:32.079 --> 00:04:33.439
Yeah, absolutely.
00:04:35.120 --> 00:04:42.160
We've been through in the last few years this massive inflection and a huge amount of progress on the digital AI side.
00:04:42.399 --> 00:04:44.959
And you know, this term physical AI has become popular.
00:04:45.360 --> 00:04:46.000
It's popular now, yeah.
00:04:46.319 --> 00:04:49.680
It feels like there is another sort of wave of of interest.
00:04:50.160 --> 00:04:56.319
You are perhaps acting upon some of the largest physical instantiations of AI that are out there.
00:04:56.560 --> 00:05:11.040
One of the questions that we often get is uh look, Chat GPT can basically replace me in terms of my job and my relationships, and why is it still that there seems to be a pretty significant gap on the physical AI side?
00:05:11.120 --> 00:05:13.040
Like, why is the physical world different?
00:05:13.199 --> 00:05:14.959
Why is it harder potentially?
00:05:15.360 --> 00:05:17.279
Yeah, this is a great uh topic.
00:05:17.439 --> 00:05:19.519
It's very clear that this is the next frontier.
00:05:19.680 --> 00:05:28.399
And there has been this energy behind the kind of generalizable models and you know, seeing what's happened with, like you said, with OpenAI and uh Anthropic and Gemini and all these approaches.
00:05:28.639 --> 00:05:30.079
This is fundamentally different.
00:05:30.160 --> 00:05:36.480
And you know, on the digital side, you have this benefit of language being a common feature across a lot of these applications.
00:05:36.639 --> 00:05:47.279
You have almost infinite data that's available on the internet and books and and journals, and there's an opportunity to like really scale in how you use that volume of data to create truly generalizable models.
00:05:47.439 --> 00:06:06.959
When you get into the physical world, the data doesn't exist, simulation's not at a point where you can just imagine all of it, and you really do need the volume of experience to deal with the long tail safety cases, and you have to go and not only figure out how to you know get that signal in order to learn you know these little patterns, but you also have for our types of applications huge safety bars.
00:06:07.040 --> 00:06:09.600
Like we're talking about giant machines.
00:06:09.759 --> 00:06:14.480
You see a lot of physical AI companies kind of brute forcing it with you know teleoperations.
00:06:14.720 --> 00:06:22.240
You know, I had a t-shirt uh back when I was at at Google before I started Neuro that said the revolution will not be teleoperated.
00:06:22.399 --> 00:06:28.160
I think it was I think it was one of the last robotics institute t-shirts that I got, but absolutely still still believe that.
00:06:28.639 --> 00:06:31.519
Yeah, the latest trend is teleoperation gets called semi-autonomous.
00:06:32.240 --> 00:06:37.680
So this would be the equivalent of OpenAI saying I'm gonna go hire a team of writers to go and train ChatGPT.
00:06:37.759 --> 00:06:43.680
It just it's almost like so insignificant relative to the scale of data that they've used that you know it doesn't compute.
00:06:43.920 --> 00:06:54.319
To me, it feels like there's a story that is easy to resonate with when these things aren't factored in, of like the the general foundation model for everything physical and every robot.
00:06:54.480 --> 00:06:56.560
But there's like physical challenges to this.
00:06:56.879 --> 00:07:04.000
Yeah, I think the same way is if you take someone who's incredibly good and has driven on multiple different platforms and you put them in a new one.
00:07:04.079 --> 00:07:14.480
Like if you or I, maybe you probably better, start driving an excavator, you're gonna have very strong like prior knowledge around driving and the basics of driving.
00:07:14.639 --> 00:07:22.560
There is going to be a significant period of additional learning that you're going to need to do to nail that and to get comfortable driving that.
00:07:22.720 --> 00:07:45.680
And then there is almost certainly going to be a massive additional period that you or whoever is managing you uh in a commercial operation is going to need to get comfortable that you're now performing at a level that passes all of the validation and that they can say, hey, we can guarantee or commit that Boris from a safety perspective is stronger than any of the other drivers.
00:07:45.920 --> 00:07:57.120
But for us, certainly the shared similarities between like a robotaxi that is carrying people and perhaps vehicles like the ones we've had a lot of experience with that are just carrying packages, right?
00:07:57.360 --> 00:08:07.279
And underneath all of it is obviously this shared driver AI intelligence, and then on top of that, there's all of the additional optimizations that you make for the different platforms.
00:08:07.519 --> 00:08:15.120
But the incremental data, but also validation and retraining and learnings become smaller and smaller.
00:08:15.439 --> 00:08:26.319
When all of us started working on autonomous driving, I think we were all like just excited about how cool the problem was, and it's like the holy grail of uh you know robotics in a in a lot of ways.
00:08:26.399 --> 00:08:31.360
But it was very easy to underestimate how hard the absolute long tail of that challenge ends up being.
00:08:31.519 --> 00:08:43.039
But in hindsight, there were like a couple of waves of kind of technological progress that really unlocked what now feels like this like upcoming kind of golden age of ability to solve these complex physical problems.
00:08:43.279 --> 00:08:53.440
Even though you are never able to get to absolute, you know, zero collisions of risk, we're starting to see that these technologies save lives and not just by a little bit, by uh by a lot.
00:08:53.519 --> 00:08:54.879
And that's actually quite exciting.
00:08:55.039 --> 00:09:06.000
So we know that you know 40,000 people die each year in the United States in traffic accidents, and construction's the number one most dangerous job in terms of injury and death in the country, even today.
00:09:06.080 --> 00:09:14.240
And so there's like genuine like public good that can happen, not even counting the productivity and time savings and everything, just in the safety aspect.
00:09:17.120 --> 00:09:21.759
Bedrocks go to market is taking these massive machines and retrofitting them.
00:09:21.919 --> 00:09:26.879
We've seen obviously there are many different players and approaches you can take in this space.
00:09:27.120 --> 00:09:38.240
From an efficiency angle, obviously, integration at the manufacturer is sort of the ideal, but that can require everything from partnerships to supply chains and whatnot that may be challenging.
00:09:38.320 --> 00:09:40.799
So, how did you guys land on that as a strategy?
00:09:41.519 --> 00:09:52.639
We studied the machines in this space, and we're able to safely go driverless with a retrofit solution because these machines have trended towards an electric architecture that is already drive by wire.
00:09:53.120 --> 00:09:58.240
And what you're integrating are sort of the full camera LIDAR, radar, sensor suite?
00:09:58.639 --> 00:10:00.159
It's uh camera and LIDAR systems.
00:10:00.480 --> 00:10:13.759
Yeah, you know, I think that the addition of the centers, camera, LIDAR, for us, also radar as well as the compute into sort of stock vehicle is something that we've seen an enormous improvement over the last 10 years.
00:10:14.000 --> 00:10:28.000
So now the NeuroDriver hardware that we integrate, so the one we're integrating into our lucid platform with Uber, as well as some other partners to provide full L4, the performance of that system and the cost has improved so much.
00:10:28.240 --> 00:10:30.559
And you know, we would love to claim credit for that.
00:10:30.720 --> 00:10:38.799
Most of it is that these are automotive components that have benefited from massive scale across consumer automotive.
00:10:39.200 --> 00:10:43.279
And the cost is now at a point where we can integrate this into a personally owned vehicle.
00:10:43.600 --> 00:10:54.320
So we truly have uh this opportunity to build personally owned fully autonomous vehicles, which I think all of us have been sort of dreaming about for many decades now.
00:10:54.399 --> 00:10:58.480
And so that is part of what keeps us so, so excited about this space in general.
00:10:58.960 --> 00:11:22.720
Yeah, we we talk a lot about the advances in AI and machine learning, but um that's just as important where when you started with Waymo in the early days of this push, a lot of the sensors had to get invented because they just didn't have the appropriate versions and the scale of effort that that drove in both time and cost, uh it's a is a gift that today um we can tackle these sort of problems much more quickly.
00:11:23.039 --> 00:11:28.320
Yeah, and I think that that's where we start to see the unlock across all the different industries and applications.
00:11:28.399 --> 00:11:30.559
And yeah, it's it's gonna be really, really cool.
00:11:30.960 --> 00:11:32.799
The next 10 years are gonna be amazing.
00:11:33.039 --> 00:11:33.120
Yeah.
00:11:33.360 --> 00:11:37.679
So we're team uh LIDAR and radar uh LIDAR, radar camera.
00:11:37.840 --> 00:11:39.279
But you mentioned you don't use radar.
00:11:39.600 --> 00:11:40.240
We don't right now.
00:11:40.559 --> 00:11:46.080
You don't have motorcyclists in black at night driving at 100 miles per hour through your construction site.
00:11:46.320 --> 00:11:56.000
I think this is one of the areas where you can rely a little bit on the like regulations of how you have to run a site and just you know, you can fail gracefully when you don't have those conditions.
00:11:56.240 --> 00:12:10.159
When it comes to autonomous driving, 100%, there's signals in these other sensors that are so complementary and to a point where you know you go through the response on every single sensor, and there just isn't a signal in camera that has to be picked up by one of the other uh sensors.
00:12:10.240 --> 00:12:20.799
And so um it's not to say that like a camera-only system can't, you know, get the human relative safety, but when you talk about these orders of magnitude improvement, I think a lot of that actually comes from the complementary nature of a lot of these sensors.
00:12:21.200 --> 00:12:25.840
The bar for performance, right, at the human level is somewhat arbitrary.
00:12:26.080 --> 00:12:36.399
It makes sense as a minimum, but making sensor choices based on that bar, I think, is probably just the wrong framing of the whole problem, right?
00:12:36.799 --> 00:12:41.840
And I think this applies very heavily to the camera versus camera and lidar and radar debate.
00:12:42.399 --> 00:12:54.720
And that's why a driver assist system that's level two is just such a different journey in development than something that's actually going into full driverless where you are 100% exposed to every element and you're expected to be safe and thoughtful.
00:12:55.039 --> 00:12:55.919
Yeah, I totally agree.
00:12:56.000 --> 00:12:58.720
And this has actually been something that we have debated over the years.
00:12:59.039 --> 00:13:05.519
Because is there is a shared foundational technology layer between driver assist versus full autonomy.
00:13:05.919 --> 00:13:11.759
But the end product and how you build it and how you validate and what you're optimizing for is so, so different.
00:13:12.240 --> 00:13:24.799
If you're trying to hedge between level two and level four, you can't afford the type of investment that's easily justified in level four if it's an assist system where it's only partially kind of capable of the sort of functionality that you need.
00:13:24.960 --> 00:13:47.919
And so when you go full level four in almost any of these applications, the value skyrockets so much that you actually get justified in investing enough to reach these sort of safety bars and capability bars, and it becomes, you know, in some ways more self-fulfilling versus trying to fight uphill from a place where you actually are handicapped from a unit economic standpoint, and therefore you have to compromise in a way that's not favorable at all.
00:13:48.159 --> 00:13:48.559
Love it.
00:13:48.720 --> 00:13:49.279
Totally agree.
00:13:49.360 --> 00:13:51.279
Uh Boris, thank you so much for your time.
00:13:51.440 --> 00:13:54.080
It's been uh a privilege uh and really lovely to have you here.
00:13:54.480 --> 00:13:54.960
It's a pleasure.
00:13:55.039 --> 00:13:58.559
It's been a fun multi decade journey with you on these windy roads.
