Image by author, using Stable Diffusion
Recent advances in AI, and specifically in Generative AI, with systems like DALL-E, Midjourney, LaMDA, BLOOM, and more (ChatGPT anyone?), are reaching higher and higher milestones at a neck-breaking pace. So much so that it’s becoming hard just to keep up with the news frenzy about it.
AI is at the peak of new results and investment: in 2021 alone, global AI investment approached the $100 billion mark.
I’ve been myself an AI researcher for a long time (30+ years, since AI wasn’t cool yet), and I can tell you it’s very far from the “interesting experiments” phase of the past, which includes, for instance, intelligent players for Chess, Go, etc. It was precisely because AI got stuck in toy problems –like playing Chess, that albeit they could be immensely interesting– that it failed to deliver in terms of Return Of Investment. Then the gap between hype and practical results produced the so-called AI-Winters when funding dried up and research efforts moved to more promising areas like biotech.
But now AI is back on track –I would say “back on steroids.” Now it’s the time to deliver widely usable products like ChatGPT –or autonomous cars.
Ops. Not autonomous cars.
Self-driving AI gets stuck
Meanwhile, after having poured $4 billion into Argo, a joint R&D between Volkswagen and Ford, this October, they decided to shut it down. Apparently, they prefer to invest in stuff they can offer right now to their customers instead of betting on an (increasingly) distant future. And as we know, in 2020, Uber abandoned its efforts in self-driving taxis after expensive litigation with Google, which sued its former engineer Lewandowski for IP theft for giving it to Uber. Things weren’t pretty.
All in all, over $100 billion have been invested in self-driving, with meager results. Bloomberg cries, “Self-Driving Cars Are Going Nowhere,” Investor Monitor asks, “Have the wheels come off investment in autonomous vehicles?” and The Guardian says that “Self-driving cars got stuck in the slow lane.” The most definitive statement is given by TechCrunch: “It’s time to admit self-driving cars aren’t going to happen.”
The sentiment is unanimous.
How can we explain such a big contrast in AI's state of progress? Is AI advancing, or is it stuck? Could it be both?
In my view, the struggles of self-driving are not intrinsically related to AI –they are more related to the real world. In particular, I see the following reasons:
1.- Cars and driving are heavily regulated.
Driving regulation comes from early in the XXth century, and laws tend to accumulate over time in order to cover more and more possible situations. There could be some laws specifically for self-driving cars, but the general laws for any car are already challenging enough.
2.- Self-driving is mission-critical, but AI is unreliable
The first part is self-explanatory, as accidents could be –and often are– fatal. The lack of reliability of AI is less known but not less accurate. Believe me, I was a researcher in AI for more than 30 years, and I’m familiar with the not-so-glorious aspects of AI. Recently, I wrote a post about this unreliability. An unreliable technology is perfectly fine for creating gorgeous graphics but not so much for accomplishing critical tasks.
3.- The economy is going through tough times
In times like these, it’s not just you and me who watch every expense. Most companies are worried about their bottom line and are consolidating their money and other resources while the waters get clear. Of course, most companies are avoiding risky investments and even less spending on nice-to-have developments. Autonomous vehicles seem to almost everybody like a “nice-to-have” thing, as opposed to a “must-have.”
In short, an over-hyped IA is not a magical recipe for self-driving success but rather the opposite.
So, is self-driving doomed?
Not in my opinion, but it will take time because it’s not just technology; it’s a real-world development.
Recently I was at a conference by Steven M. Miller, author of the book “Working with AI: Real Stories of Human-Machine Collaboration.” He argues that while the technology itself can sometimes advance very quickly, real-world applications are much slower because they have to take into account myriads of practical details. This applies, of course, to self-driving cars. One of the mundane details that have to be worked out is sorting legislative hurdles, as commented above.
In my view, the most critical aspect of self-driving is that it must give a return on investment at some point, and the sooner, the better, especially when strategic investment is drying up. There are several possible ways to do that:
1.- Charge car owners for self-driving capabilities, as Tesla, which is starting to offer a so-called “Full Self-Driving” to “select customers,” people willing to pay $12,000 for such a blessing. As usual, in this case, Musk is overpromising with “Full Self-Driving” capabilities, which hardly reach Level 3 of the self-driving standard scale, where Level 5 is complete autonomy. Tesla’s Full Self-Driving doesn’t even comply with Level 4, as this one requires that the human can leave complete trips to the autonomous car without paying attention to the road or to eventualities.
2.- Charge robotaxi passengers for the ride. This option implies bleeding money for a long time before making profits, but it’s a viable option, nevertheless.
3.- Get government funding, given the many cost savings that could be achieved with autonomous vehicles, like, for instance, a reduction in roadwork costs, as well as a radical cut in road fatalities. Well, this one is not really revenue, but it could keep the wheels turning (no pun intended).
4.- Use self-driving capabilities for long-haul freight trucks. This makes sense now there is a truck driver shortage, and the cost of human driving time (when the driver has to stop to rest) or driving pay is a high percentage of transportation costs. Tesla has announced its “Semi” truck, though its autonomous driving capabilities are not clear yet.
Startups like Waabi, Embark, TuSimple, and more are approaching or starting business operations because freight transportation is a highly simplified scenario when compared to general-purpose cars. Established truck manufacturers like Volvo, Kensworth, and Navistar are working to provide higher levels of autonomy than their current level 2.
And what if nothing works for autonomous vehicles?
I confidently predict that at least one of the options above for profitability will work. But in case they won’t, human transportation for the next decades could be one of these:
- Mostly the same as today 🙁
- Walking in redesigned cities for walkability. Not a solution for cities in extreme weather.
- Use VTOL taxis instead of cars. Easier said than done.
- Use public transportation combined with last-mile personal vehicles, though this option has been heavily criticized.
AI at large is finding profitability and is doing it in a big way, while autonomous driving is struggling. This is a fact. What’s up to debate is the explanation. I’ve done my best to present the possible causes.
But it’s hard to tell if the current slowdown of self-driving is more related to the economy, unrealistic expectations, or the technological problems in the real world. I think it’s a combination of all of them.
Personally, I bet on trucks for having the first real success stories in self-driving. One underestimated advantage of trucks is that they are less “sexy” than cars, so they are under lighter media scrutiny.
You see, media coverage can be great, but sometimes it is a liability, in particular for car accidents. From the perspective of the tech giants, one fatal accident involving one of their autonomous cars raises a colossal backlash, damaging their public image.
But honestly, making completely safe self-driving cars means going far beyond human abilities: we humans manage to get involved in many million cars crashes per year, most of which don’t raise a single headline but result in 1.35 million deaths worldwide. Alcohol and fatigue are significant factors of this toll, and obviously, these ones are immaterial for self-driving cars, but then again, statistics will only be seen many years from now.
We are in a very tricky stage of self-driving, where its initial dreams are vanishing, but its actual advantages are not yet visible.
What are your thoughts on this? Please comment.