The Operating System for Self Driving Cars (and Tanks, and Trucks...) With Qasar Younis and Peter Ludwig of Applied Intuition

Microsoft is a company that successfully executed this strategy. From 1975 to 1982, they developed tools that many people no longer remember. After that, they transitioned into operating systems, then applications like Office and Word. We are doing something similar, but our hardware focus is on cars and trucks, not PC manufacturers. I think one of Peter Thiel's old sayings applies here: 'What do you believe to be true that other people don't believe is true?' This was a key question when we started the company.

My slightly more critical view on Europeans is that they are a little bit asleep at the wheel. If I could inject something into the minds of European leaders, whether they are commissioners or industrial family heads, it would be this: you have to fight. There's almost this mentality that says, "Oh, it's going to be... The Chinese are so cheap." The reality is, that's not true. There's a finite amount of money required to make something.

If we look ahead, we can see that our manufacturing and industrial base was effectively exported to other countries due to labor arbitrage, or cheaper labor. This meant a lot of our know-how left the country and remained there. In some cases, it feels like we've lost some of that expertise over time. Can robotics, autonomy, and factory automation now allow us to bring that back, enabling a factory to be competitive anywhere, based on automation rather than labor costs?

We've been investing in synthetic data for quite a while, and the details of how that works have evolved. It has actually gone through several generations in our own tech stack. Traditionally, you had a very computer graphics-heavy approach. Now, approaches use things like Gaussian splatting and diffusion models to do really interesting things with synthetic data. You can also extend synthetic data into a bunch of other domains, including classified domains for defense, where you can actually gain some really interesting advantages.

One technical strategy that has been quite advantageous for us is that we always wanted to wait in the wings. We explicitly discussed waiting until the autonomy ecosystem converged on a handful of techniques, especially after this post-transformer boom. Now, you can experience this firsthand. You can ride in V13, the Tesla FSD product, or go to China and ride in these vehicles, and you realize, "This is it." This wasn't the case just three years ago. There was a huge debate about whether end-to-end, in the many ways it's marketed, was truly viable.

The key thing is the risk is lower. It seems many media headlines are the opposite, purposefully amplifying danger even when it has decreased. We've seen this in other instances too, so it feels like a deliberate approach by those writing the news.

There's huge value, especially when the company is young, in not constantly being out there. I'm sure there are downsides as well, like people not knowing you, which makes it harder to recruit or whatever. That's never really been a huge issue for us. However, a key advantage is that the moment you state, "We do X," there's an expectation that you will do X, even if X turns out to be incorrect. Therefore, I believe it's important to keep your identity small.

Much of the software embedded in these systems lacks intelligence. It's primarily focused on basic input/output functions. For instance, the seat warmer simply turns on and off. You can centralize all of this into a single CPU, which offers significant efficiency. Not only are all signals processed centrally, allowing for more advanced functionalities like those seen in Tesla vehicles, but it's also more cost-effective. By replacing redundant systems, which are often poorly managed and built with diverse software stacks across their subcomponents, you eliminate wiring harnesses and thousands of dollars in physical hardware. This removal literally leads to increased functionality.