Muhammad Umar argues that winning autonomous driving requires both hardware and algorithms, but also massive high-quality data. Tesla's Full Self-Driving codebase shrank from millions to hundreds of thousands of lines as neural networks replaced hand-coded logic. Vemo spent a decade collecting data in San Francisco; the real race is gathering that volume in a single year.
Chapters
Just to add on, like, of course it's a system race, but just one thing on top of it, like, it's a system race and it's also a data race. So we also need to, I mean, Andrej Karpath, he had this incredible talk where he explained how software 2.0 is eating software 3.0. And he gave the example of FSD test. FSD, like how I think millions of lines of code got replaced with few hundred thousand lines of code. The reason is that it's like that hardcore business, that hardcore logic is now being replaced by pattern recognition. So how this is possible, it's also a data race. Now you have so much data coming in that now your neural nets and all the, all those algorithms just perform better. So it's a system race, but it's also a data race. Always think of like, I mean, it's been 10 years, Vemo has been hovering around San Francisco trying to get all the data, get their perception, prediction right. And I think, is there another way to do it? Let's say, what does it take for us to do the same thing that, for example, VEMO has done for a decade, if we have to do it in a year?
