Web Summit: AI in the Physical World with Raquel Urtasun and Sven Stroband, Khosla Ventures

Yeah, I think when you look at the application domain where this idea of AI doing tasks in the physical world is going to really impact and influence and change the world, self-driving is where we believe the big revolution is going to happen first. There's been a big shift in the industry in terms of how this was approached 20 years ago—I participated in and led one of the teams, the Stanford team, for the DARPA challenge—and where it is today, and the impact that this next generation of AI is bringing to finally building a scalable solution. Perhaps you want to drive us through your experience in the last 20 years.

Yeah, when I started the company four years ago, I saw a tremendous opportunity to bring the next generation of AI technology—this was before the ChatGPT moment—thinking that this would really revolutionize the way we approach self-driving, so that we have a much more capital-efficient approach, faster path to market, and at the same time save many lives. It was rooted in two differentiated technologies. One was that the traditional way of doing self-driving was this massive, I call it a spaghetti stack, hand-engineered by humans, very brittle, where you needed to drive millions of miles in the real world to try to see the next set of corner cases the system couldn't handle. I didn't believe that would provide us the opportunity to build something that can truly scale and solve what our customers really want. With our approach, it's an end-to-end system—a single AI system can do all the tasks for driving. It's able to interpret what it's sensing and use those interpretations to reason about all the things it can do and the consequences of its actions. This is very important because it can understand the risk of every possible maneuver and always choose the right action. This gives us the ability to generalize with very little data to things it has never seen before, which is a must-have when you have a robot driving an 18-wheeler with 80,000 pounds behind you. The consequences are tremendous. The second thesis for Waabi was that we will never drive enough to see everything that can happen in the world, no matter how big your fleet is. So the idea was to build the Matrix for self-driving—a simulator that is the same as the real world—so with no consequences, you can expose the system to accidents, safety-critical situations, all corner cases, things that would take many lifetimes to observe. This was before any neural simulator was built, and it turned out to be the right approach because in AI, data is more than half the equation. With that simulator, we can train the system, test it, and provide evidence that it is safer than humans. It was very different from anything else when I started the company, and four years in, you see that becoming the trend, but it was very contrarian at the time. Sven, what led you to invest in Waabi? You're also on the board.

Yeah, before Waabi, I was chief scientist and head of R&D for Uber, working on both robot taxis and self-driving trucks. That's where I really learned about how exciting the unit economics are for trucking. The way we work with Uber—they are a big investor in Waabi, but at the same time, we have a really exciting autonomous marketplace partnership with Uber Freight. Think of Uber Freight sitting in the middle of supply and demand: the carriers that own the trucks, the shippers that have the loads to go from point A to point B. We provide the virtual driver, and they do the matching in terms of getting our trucks driving 24/7, working with top Fortune 500 companies. With this exciting partnership, you will see a deep technical integration where it's seamless for both shippers and carriers to get access to Waabi's technology. The other thing you will see is that we have billions of miles of committed shipments within the Uber Freight network. Today, they have $20 billion of freight under management, which is massive—they have access to the majority of the market. So it's a very exciting go-to-market for us to work closely with Uber. We have a great relationship that started during the Uber days with Dara, CEO of Uber, as well as Lior Ron, CEO of Uber Freight. We continue to go through this relationship, but for us, it's about accessing a massive customer base without building a big sales organization.

Yeah, robot taxis and self-driving trucks have different challenges. On the robot taxi side, urban cores where humans do all sorts of crazy things are the major challenge. For trucks, from a technology perspective, you need to see much farther. The reason is that trucks, even driven by humans or robots, the maneuvers you can do are much smoother and take longer just by the physics of carrying 80,000 pounds—you cannot just swerve, you will flip the truck. So that's one difficulty. Also, controlling variable loads makes it more challenging from a dynamics perspective. But at the same time, it's a massive opportunity. If you look at the industry today for truck drivers, not everybody is a million-miler; there are a lot of novice drivers who have a lot of responsibility driving trucks that can have tremendous consequences. Self-driving trucks will be so much safer than any human, so from that perspective, it's a great opportunity to minimize the probability of any bad event and save many lives. They are absolutely superhuman, no doubt.

The regulatory framework in the US is actually really favorable for self-driving trucks. In the majority of states, we can deploy driverless trucks. I was just in DC with Secretary Duffy, and it was great to see the willingness to create a federal framework that will really enable this industry to expand to the entire United States, so you don't need to go state by state working with regulators for a path to deployment. In the past, it might have been a potential challenge, but right now we see a very open regulatory framework.

One of the things people need to realize is that it's not that suddenly you flip a switch and everything is autonomous. This is the physical world, so penetration, scaling, adoption is going to take time. We believe that any truck driver today who wants to retire as a truck driver will be able to do so. There was a recent study by the US Department of Transportation that shows self-driving will create more jobs than it displaces, and those jobs are much better. What people don't realize is that being a long-haul trucker is really difficult. You go weeks at a time on the road, and often for many days you don't have access to toilets, showers, etc. If you are a woman truck driver, there are huge safety issues sleeping in your truck. So with this technology, we are really enhancing the kind of jobs we can do. It will create things like remote assist terminals, enhanced inspection jobs, etc., that will be more exciting for humans to do. We see automation as an enhancer, not as removing the human from the question.