Robert Scaringe56:15
Andres, this is by a significant degree one of our biggest focus areas. I alluded to this and talked about this before, but just to be very explicit, the approach that we're utilizing on autonomy represents a significant shift, a significant step. You often hear it called AV 2.0 relative to what was done prior to 2021, 2022. So what we launched with on R1 with our Gen 1 vehicle is a completely different topology, different sensor set, different compute stack, completely different software architecture to what we now have on our Gen 2. And so what we designed our Gen 2 vehicle with was a much higher level of compute, so it's inference on vehicle, much more capable cameras. We brought all that perception stack, the camera set, in-house. Of course, we designed the compute platform to support that really rich camera set that I talked about earlier. In the case of R1, that's 55 megapixel cameras with outstanding breadth of performance. And really the goal of that, which launched a little more than a year ago now with the Gen 2 vehicles, was to create a data platform to train a very capable model. And that very capable model of course gets better as the fleet size grows. And when we say train, I want to be specific here. It's not just having vehicles with lots of sensors and a lot of inference on board. The way that we trigger what data gets sent back to us, what are the triggering events? Of course, it's when the vehicle encounters an issue, when it disengages from its autonomous driving mode, but importantly, when you're not in autonomous driving mode, what's triggered? What are we doing to find interesting situations that fill in and add robustness to this large model we're building? So we've really spent a lot of time defining a whole array of different triggering events that allows us to pull data off of our deployed fleet, use it to build robustness into the model, and then redeploy that model in a distilled version back down to the vehicles. And that's what customers are starting to see in terms of improved feature set. And we are deeply of the belief that this approach of an AI-centric approach, what you may often hear called an end-to-end approach where you're training based upon the behaviors of the vehicles, is the approach that's going to win. But it does require you to control the perception stack. It requires you to have a very robust data triggering architecture. It requires a very robust data movement architecture, meaning your vehicles are ideally connected to Wi-Fi, which allows much lower cost movement of lots of data. And so the incentive structures we've put in place around having customers on Connect Plus and having them Wi-Fi connected helps drive that. But that feeds our offline training that's occurring on lots and lots of GPUs, which then continues to get better and better. So that set of ingredients we think very few manufacturers have. We're investing tremendously into it. And what we'll start to see in terms of customer-facing features is, first and foremost, we'll see the vehicles be able to operate under a much wider range of roads. So today we limit our feature to a highway feature. In the not too distant future, we'll go map-free. And when we say that, what we really mean is it can operate essentially anywhere. And we've already gone hands-free in certain conditions on highways. We'll broaden that to be more conditions, so eyes on, hands off. And of course, the next step as we look into 2026 is identifying specific areas where you'll go hands off, eyes off, where you'll truly get your time back, where you can be looking away from the road, not an active participant in the operation of the vehicle, and coupling that with a turn-by-turn capability. So you get into your car, you can imagine the future state where the car, you give it the address, and it takes care of getting there. So that is our roadmap for the next couple of years. That's what we're driving towards. Core to that is this building of a large model. Of course, that model has applications beyond just R1, R2, and beyond, as you referenced. There's applications in the commercial space and with our commercial vans. So we're very, very excited about what this represents in terms of the business opportunity it creates, but also the customer experience it forms.