00:01:44 · Ruby Justice Thelot, Brandon Basso
[00:01:44] S1: What is the hardest part about this? [00:01:47] S0: Oh my gosh. So many things. Well, I am the safety validation person. Nice. For job security reasons, I have to say that part. But there's so many hard aspects. I mean, as an industry, we've been at this problem for two decades going back to foundational research happening at Berkeley and Stanford and many other institutions around the world. And there's so many hard parts. I I would broadly split it into two big categories, and they're very much addressed in some of the the content that we put out today. So one is how do we make a driver drive better than a human, superhuman, an AI driver? And then two, how do we validate that it's doing what it's doing? Because it's, of course, not good enough to just say, well, we built this great model, and here you go. We have to prove that it works. And both of those parts have their own challenges. I personally focus a lot on the proof part, but they're both extremely hard when you think of the breath of driving that can happen scenarios, weather conditions throughout the entire world, which is what Wayve focuses on literally everywhere. [00:02:49] S1: Tell me a little bit more about safety. Why is safety such a difficult engineering problem?
00:02:49 · Ruby Justice Thelot, Brandon Basso
[00:02:49] S1: Tell me a little bit more about safety. Why is safety such a difficult engineering problem? [00:02:56] S0: Yeah. So, I mean, let's start today with the status quo. The status quo today for driving is not great. It is, a very risky thing that we do. It's costly in terms of time, but also just insurance costs for minor accidents. And then, you know, we don't even need to motivate the the major safety challenges and the populations affected by vehicular accidents in The United States everywhere throughout the world. So it is, I think, incumbent upon us as engineers, as scientists, as product managers, researchers, to think about how we improve all of those problems. So it kinda is one of those things that really, I think, needs no motivation, And why it's such an intent focus for Wayve, well, I mean, it's a focus of every self driving car company out there, whether you're working on l two or l four. We have this novel approach to self driving that we refer to as AV two point o or end to end systems, And there hasn't been a lot of rigorous thinking yet about how we prove the safety of those systems. It's exactly the same problem that folks at Anthropic and OpenAI think about when they are thinking about frontier safety. How do they prove that what the model is outputting is correct performance, not going to mislead you, not going to deceive you, aligned with human expectations. It's the same problem, but the physical interpretation of it.
00:02:56 · Brandon Basso, Ruby Justice Thelot
[00:02:56] S0: Yeah. So, I mean, let's start today with the status quo. The status quo today for driving is not great. It is, a very risky thing that we do. It's costly in terms of time, but also just insurance costs for minor accidents. And then, you know, we don't even need to motivate the the major safety challenges and the populations affected by vehicular accidents in The United States everywhere throughout the world. So it is, I think, incumbent upon us as engineers, as scientists, as product managers, researchers, to think about how we improve all of those problems. So it kinda is one of those things that really, I think, needs no motivation, And why it's such an intent focus for Wayve, well, I mean, it's a focus of every self driving car company out there, whether you're working on l two or l four. We have this novel approach to self driving that we refer to as AV two point o or end to end systems, And there hasn't been a lot of rigorous thinking yet about how we prove the safety of those systems. It's exactly the same problem that folks at Anthropic and OpenAI think about when they are thinking about frontier safety. How do they prove that what the model is outputting is correct performance, not going to mislead you, not going to deceive you, aligned with human expectations. It's the same problem, but the physical interpretation of it. [00:04:18] S1: So there isn't this golden North Star benchmark for what this means. What does that look like when you're trying to innovate at the frontier to define what safety means?
00:04:18 · Ruby Justice Thelot, Brandon Basso
[00:04:18] S1: So there isn't this golden North Star benchmark for what this means. What does that look like when you're trying to innovate at the frontier to define what safety means? [00:04:29] S0: Yeah. So many things. So we we break it down into four core concepts. So safety, of course, has to do with the safety of road users outside of the car, but also the people in the car. There's a comfort element, which is sometimes safety related or adjacent. You know, if you're heartbreaking or swerving or things like that, doing uncomfortable things, sometimes those are precursors of safety. Honestly, equally important is compliance. So there is a dizzying array of traffic rules throughout the world. How you behave with certain signs in Japan versus in Germany is quite different. [00:05:06] S1: Because you're doing London and all across
00:05:06 · Ruby Justice Thelot, Brandon Basso
[00:05:06] S1: Because you're doing London and all across [00:05:08] S0: the city. Literally everywhere. Literally... Quite literally everywhere. When we're operating like we've done with our road show last year in the Arctic Circle or just around the world, it's it's crazy how different it is. And regulators and our customers expect that we're performing safely, comfortably, compliantly, and competently. Meaning that when you ask the car to do a lane change, it does it in a way that people expect. Now your question is, well, how do we grade all those things? And that is the challenge, and there's no one answer. There's no one metric just in the way that you go to any AI leaderboard today. There's a ton of different metrics, but we try our very best to make sure that those metrics are aligned with human expectation, our expectation when we say what it means for a driver to be safe. [00:05:55] S1: Yeah. Something you alluded to at the start of our conversation is you mentioned a tough piece is gathering enough data to train how... These cars to drive autonomously. Famously, even a decade ago, I remember I grew up here in the Bay Area, and you would see the people inside the Waymo cars just gathering data about what's happening on the streets. Where are we now? What's different about 2026 about how we gather all the simulation data?
00:06:21 · Brandon Basso
[00:06:21] S0: Yeah. Absolutely. So one of the core advantages, and I I would say core aspects of Wayve's approach going back to a decade ago, is we very much recognize both the driving problem and the validation problem as a data problem, and we secured all sorts of interesting diverse data early on. We did this roadshow campaign where we drove well north of... I I forget if it's the Arctic Circle or not, so I'm not geography expert, where it was very cold, and collected diverse data. And that is so critical because our system is fundamentally one that gets better with data just like any frontier LLM today. And people in robotics, general robotics will tell you this. There is no general dataset for robotics today, for physical AI. There's no general dataset like there is for language. You know, humans have developed all this language data throughout the years, books and, you know, videos and stuff. And we don't have a corpus yet of driving data, so we work very hard on collecting that. That's both our first party collection, purchasing data going back to ten years ago, and then partnering with our fantastic OEM partners and working with them on getting the data so that we can deliver back to them the best the best products.