Inspired with Alexa von Tobel
00:02:32 · Boris Sofman, Alexa von Tobel
[00:02:32] S0: Pleasure. [00:02:33] S1: I wanna start with the beginning. And before we dive in, I'd I'd love to talk a little bit about kind of how you got here in life. You were born in The Soviet Union, later immigrated to The United States. I would love just to get a sense of, like, a formative moment in your childhood that you feel like launched you to where you are today?
00:02:51 · Boris Sofman
[00:02:51] S0: So I was born in Moscow, like kinda yeah, right in the heart of Soviet Union. My parents have been trying to leave for well over a decade and you just couldn't. That feels like such a foreign concept today. They were finally able to leave in 1989. Was actually one of the earliest families that were allowed to leave. And my dad was actually pretty deeply involved in kind of helping coordinate kind of families of Russian scientists and mathematicians and engineers that through kinda coordination with the US government were kinda negotiated to be able to leave Soviet Union, and I was six years old. And so we moved through a few countries in Europe and ended up in New York, lived in Brooklyn for a bit, and then through my dad's work ended up in Texas, where he was in telecommunications doing large scale optimization, and so I was in the Dallas suburbs for most of my childhood. And so I was always excited about kind of mathematics and engineering, and video games was actually one of the things that got me interested in AI, where I still remember there's a famous video game series, computer game series called Civilization, where you're basically kind of simulating the history of the world and kind of playing Civilization, thinking about research and units and everything. And the complexity was so immense, which was part of the appeal of the game. The AI was okay, but not great at all. And I started to think about how, what are the kind of algorithms that drive the kind of coordination of all these systems? And that actually got me interested in engineering and computer science, which took me to Carnegie Mellon. And it's funny because today, my 11 year old son, got him into civilization and normally parents would be like upset when their kids get into video games and I'm kind of like strangely proud that my son's kind of following in my footsteps.
00:04:38 · Alexa von Tobel
[00:04:38] S1: That's incredible. I love that. I wanna talk a little bit about your time at Waymo, where, again, you really led their autonomous trucking initiative. I could geek out on this just because I'm so fascinated by how that could transform society, safer, all of the things. But can we just get a sense of, like, first of tell us what you did there. Let's start there. And then I wanna talk a little bit about kinda where it's headed. So let's start there.
00:05:05 · Boris Sofman
[00:05:05] S0: It was such a wonderful time. So I was at Waymo for about five years, so I was an executive leading autonomous trucking, but then I was also leading various technology teams that supported kind of all platforms, cars and trucks. And so I was leading a perception team and then various other teams. And so we helped launch the Jaguars into San Francisco. We unified the tech stack to where cars and trucks were as unified as possible in terms of the overall architecture, where you'd have differences in the platform, sensor layouts, you'd have an articulated trailer, but we were using almost every single part of the autonomy stack, which was very powerful. And so we ended up leading both freeway driverless for both cars and trucks. Waymo ended up focusing much more on the car side where there was like a really a pressure to commercialize and accelerate that. And so we actually launched driverless on freeways for the cars at the end of twenty twenty three. But in a lot of ways that was actually the trucking kind of like kind of program, kind of unified into the car platform. And it was a fascinating problem because with trucking you had this like 80,000 pound trucks with an articulated trailer. You had like crippled vehicle dynamics where you know, there's like, so had so little flexibility on acceleration and deceleration because of the mass of these vehicles. You had these interesting challenges that because of that, we had to push the perception range out from a few 100 meters, which is what the car is mostly utilized to close to a thousand meters, you know, using camera and radar to go far beyond LIDAR range. You started to think about these extremely rare long tail kind of challenges that appear so infrequently on a freeway and completely changes the way you do a safety framework. It was just amazing problems.
00:06:46 · Alexa von Tobel
[00:06:46] S1: Can you give us an example of like one of those sort of real long tail challenges that you had to think through?
00:06:53 · Boris Sofman
[00:06:53] S0: Well, let me share one that's like actually a big class of problems, reliability. So reliability is effectively how resilient are you to failures in pieces of your software and hardware. And so what's interesting about freeways is that the speed at which you drive means that you cannot stop in a lane. Like it's just like your budget for that is just like tiny, absolutely tiny. Because if you stop in a lane on a freeway, you actually have a meaningful risk of having a rear end collision with a really, really serious, kind of high severity of collision. And as I think like something like six to 8% that integrates over time, the longer you're sitting there. And so you have this situation where on a surface street, when you're driving in San Francisco, you're at 25, 30 miles an hour, almost always. If you stop, it's a bad product experience or kind of operationally a headache, but it's not a safety risk in the traditional sense of safety. Here it is. And so not only do you have to reinvent the platform of redundancy for brakes, steering, actuation, electrical kind of anything that mechanically could fail. But by percentages, the largest frequency of what you have to deal with is everything else. So a sensor failing, your computer crashing, a process hanging, your localization failing, your router, your perception behavior, almost anything. And individually, things might be rare, like tens of thousands or hundreds of thousands or millions of miles in between. But when you compound all of them, you can't go driverless on a freeway without a solution. And so you end up having to have this incredible safety net of redundancy across everything where you have like catchalls with software that's constantly changing every single release and thrashing the reliability rates. And you have to have like a space shuttle level of reliability with a system that's evolving all the time. And so you have layers of primary and secondary systems and catchalls and budgets that you have to manage to deal with things that you will never see in our lifetime. So for example, a computer will have a kernel panic that just crashes completely out of your control every 500,000 miles, that'll eat up, that's way more than your entire budget of these sort of failures.