What is it all for
I started this newsletter because, after the sudden evaporation of the company I'd spent seven years building, I had more to say. Things I hadn't been able to say in that job. Things about the substantial problems with autonomous cars as a solution to mobility problems. Things about the substantial problems with CARS as a solution to mobility problems. Things about the substantial problems with autonomous cars as a realistic technological solution. Over the first few weeks of the newsletter's existence, I kept a fairly brisk schedule talking about all of those topics. Although it hasn't been very long, I think I have learned some things.
First of all, many of the things that I have been unable to say, but wanted to, are said very well by others. Advocates and writers like David Zipper and Jesse Singer, general interest writers like Alex Pareene, and the healthy community around podcasts like The War on Cars are eloquent, thoughtful and informed on the problems that ensue when cars are seen as a general purpose solution to transportation. Paris Marx, among others, is perceptive about the failures of Silicon Valley-style venture-funded startups, how they generally miss the mark on what kinds of solutions are actually useful for transportation. When I wrote about those things, I like to think I was basically coherent and informative, but a lot of what I was writing felt to me like received wisdom.
The other category of topics I wrote about in my first few weeks was somewhat deeper dives into the technical problems with autonomous cars. I wrote about the problems with how computer vision is defined. I wrote about the fundamental issues with how autonomous vehicles are defined and understood in the world. On these topics, I have an expertise that is probably less well represented in the world. But the material, while interesting and important, at least as a way to understand a technology that is likely to have implications for how the world looks in the future, didn't feel compelling. At least, it doesn't compel me as much as, say, ranting about the horrible infrastructure decisions that have led the U.S. transportation network to its current sorry state.
So that left me feeling a bit adrift, like I had yet to settle on the value I hoped this newsletter would bring to the world. And I think what ties the topics I've covered so far together is that they're things I've experienced firsthand and, well, it was really weird. I'd like to get across the sheer uncanniness of an entire economic sector where every single person seems somewhere between willfully and unintentionally oblivious to the screamingly obvious fundamental issues with what they're doing. The weightless feeling of being a first-time founder operating on luck and bluster trying to climb step by step into an industry where the entire superstructure is luck and bluster. The constant oddball dissonance of knowing that the deeper somebody's technical understanding—the more amazing the technical feats they were accomplishing on a day-to-day basis—the more insight they were quietly keeping to themselves about the sheer technical unreadiness of everything that was allegedly happening.
In some sense this pathology is pretty universal in tech. The idea that every tech company is teetering on the edge of catastrophe behind the scenes is a truism among software engineers. But the world of autonomous cars is weirder yet, because everybody is selling something that doesn't exist to help something else that doesn't exist become real on the as-yet-unprovable premise that anybody will want that thing when it does exist. It’s head-spinning, speculation piled upon speculation piled upon speculation, fueled by hundreds of millions of dollars and the breakneck, desperately inventive work of some of the most uniquely skilled, highly compensated and extravagantly employable engineers in the world.
Even at the time that we were starting Perceptive Automata, the idea of working with autonomous cars seemed like a little bit of a bank shot. They were too far in the future, too implausible. Speaking personally, I thought they were a bad idea. But as we iterated through other ideas, some seemingly kinda good, some kinda dumb, the world looked like it was disagreeing with us. Suddenly, there were dozens or maybe even hundreds of autonomous vehicle startups, and the venture capital—the smart money, so I heard—was flooding in. The advances, whatever my personal thoughts about how likely the industry was to succeed, had been material.
As we built our prototype in late 2016, the conversations we had internally were about whether we were too late. We assumed that by then, our potential customers would surely have solved perception: knowing what and where everything is. Obviously they would have solved prediction, as they defined it: knowing where everything was going to be based on its position and velocity. Surely behavior planning—knowing the right action to take predicated on a given road situation and taking it—was in good shape. So our internal debates were about whether our potential customers had already solved human behavior, and were just keeping quiet about it. I figured they hadn't, for what it's worth.
As it turned out, none of our assumptions were correct. Perception wasn't solved, and is now at the point of being maybe good enough, with some major caveats. Work on prediction had barely begun. Behavior planning, well, even if they had the capacity to build a rich understanding of the current state of the world, they didn't—and still don't—have a solution for how to know what the right driving decision was. Every single step of the way, as we learned more about the internal state of development at the level 4 autonomous driving companies we hoped to sell to, we were caught out by just how far behind they were. Every slick demo, every advancing deployment, every bold press release and billion dollar investment: behind all of them was a yawning abyss of fantastically difficult, unsolved technical problems.
Today, seven years later, there has been enormous progress. If you squint, and make some generous assumptions, the very most sophisticated players might be at the point where they can start to see whether anybody is interested in using their somewhat-close-to-existing product. That’s often taken as a given, but another fact that nobody is interested in talking about is that the answer could easily be no, all kinds of reasons. I’ll talk about that some other time, but for now it’s worth noting that the entire industry—the hundreds of millions of dollars invested, the incredible technical breakthroughs, the years of deep developed expertise—could evaporate into nothing, dissolve into air. Aurora, one of the very most sophisticated players, after pivoting away from autonomous urban driving, just had an email leak that indicated that they might be looking to unload the company. Even now, it could all disappear.
So it has been weird, being a part of that. It was extraordinary and fun and confusing and never, ever not weird. And if I can contribute anything in this newsletter I hope I can give you a sense of that, the surpassing strangeness of the whole AV world, from the very specific technical details of why these systems won't work right to the broader social question of what on earth the point is supposed to be.
