Bloomberg – Why AI Investors Should Worry About the Self-Driving Car Crash
Editors note: this article is one of the best written on the hype of the robotaxi business, and its inevitable fall. It shows how all of the robotaxi companies (including Waymo, but especially Tesla) have been dependent on hype to make the false claim that robotaxis have a near-term financially robust future. Yet they are still shockingly dependent on support by a large number of actual humans. And now, not only do robotaxis appear to be post peak bubble, generative AI companies will surely be following the same trajectory.
See original article by Max Chafkin at Bloomberg
Robotaxis were supposed to be the easy part of automation. The failure of GM’s effort shows how far the industry is from living up to its wild promises.
In mid-December, Mary Barra, chief executive officer of General Motors Co., dialed into a conference call with analysts and announced a decision to “realign our autonomous driving strategy.” The company was shutting down development of its driverless cars—run by a subsidiary known as Cruise—and would fold the team into the part of GM that works on software for its regular lineup. Barra said this was about “accelerating the path forward, providing customers meaningful benefits along the way.”
What was presented as a strategy shift was also a profound admission of failure. For years, Barra—like many executives in the tech and auto industries—spun a fantastical vision of the future in which fleets of so-called robotaxis would imminently replace normal cars. The technology was already developed, according to GM’s boss; the only thing left to do was scale it up. “We’re here. It’s happening now,” she boasted at the 2023 South by Southwest Conference in Austin. She routinely claimed that GM, which had revenue of roughly $50 billion in its most recent quarter, would make an additional $50 billion per year from robotaxis by 2030.
These predictions turned out to be outlandishly optimistic, relying on questionable data and technical kludges that made the company’s software look more sophisticated than it actually was. Perhaps more unsettling, amid a boom in artificial intelligence technologies that has companies large and small contemplating replacing large numbers of human workers with modified chatbots, Cruise was hardly alone in overpromising. The company’s failure not only offers a cautionary tale for others attempting to sell robotaxis, especially Elon Musk’s Tesla Inc. and Google’s parent, Alphabet Inc., but it also suggests that the wild promises of operators of AI chatbots (and the companies that depend on these chatbots to justify their sky-high valuations) should be met with caution, if not outright skepticism. After all, autonomous driving was supposed to be the easy part of AI.
Despite its failure, Cruise got as close as almost any company has to operating a viable commercial driverless car service. The problem was, it wasn’t very close at all. At its peak, Cruise was a money-burning novelty, consisting of a few hundred cars overseen by a staff of thousands. The cars were kept off highways and difficult-to-navigate roads, yet they still managed to interfere with fire trucks and other emergency vehicles while causing a seemingly uncountable number of traffic delays in downtown San Francisco when they glitched out mid-drive. In the face of evidence to the contrary, the company ran ads in the summer of 2023 touting a study—naturally, one the company had conducted—that claimed its cars were superior to the ones normal people drive.
None of these ads made clear that Cruise’s “driverless” cars were in fact operating only partly autonomously; they relied on large teams of humans working out of call centers to monitor the vehicles and tell them what to do when necessary. This fact wasn’t exactly a secret within the industry: Waymo, Alphabet’s driverless car subsidiary, whose robotaxis are becoming increasingly ubiquitous in Los Angeles, Phoenix and San Francisco, also relies on so-called remote operators. But it helped create the impression that Cruise’s software was more sophisticated than it actually was.
Then, in late 2023, a Cruise robotaxi was involved in an accident in which a San Francisco pedestrian was struck by another car and landed in the path of the oncoming autonomous vehicle. The Cruise-operated car braked hard but still hit the woman. Rather than stopping to make sure she was OK—what a decent human driver would have done—the Cruise kept going, dragging her for 20 feet. The woman survived but was hospitalized with serious injuries. Cruise eventually settled a lawsuit brought by the victim for about $10 million, according to Bloomberg News, and also paid fines to state and federal regulators for withholding details about the crash. Cruise suspended operations (temporarily, the company said at the time), and its CEO resigned.
In the eight years since buying Cruise, GM burned through more than $10 billion operating the division. “The cash outlay has just been phenomenal for the incredibly low return on investment,” says Missy Cummings, director of George Mason University’s robotics center and a former adviser to the National Highway Traffic Safety Administration. Conventional wisdom about Cruise says that GM’s problems were singular, some combination of bad luck and corporate ineptitude. But Cummings says those who believe this misunderstand what happened to the company and what appears to be happening to its peers.
Like Cruise, Waymo spent enormous sums on the way to building a business that, while technically impressive, amounts to a modest fleet operating in only a handful of places at slower-than-normal speeds with no shortage of hiccups. For instance, in December, a Waymo customer attempting to use the service to go to the airport in Scottsdale, Arizona complained he’d been trapped in a robotaxi that spent five minutes going around in circles on the way to the airport. He called customer support, and an agent got the car to pull over. Alphabet doesn’t report Waymo’s losses, but its “other bets” division, which includes Waymo, has lost about $37 billion since 2016. Waymo is currently testing highway driving, but it’s yet to offer those rides to customers; Cummings says it’s because they can’t yet do it safely. The result is a service that’s popular with tourists in San Francisco but only commercially viable thanks to the enormous profits that Google’s search engine throws off. “What they’ve accomplished is tremendous,” Cummings says. “But they’re still limited to 45 miles per hour, and they don’t want to talk about that.” She says a fully featured robotaxi is still decades away.
The failure to successfully train computers to get anywhere close to the capabilities of any Uber driver (after 15 years of sending cars loaded with sensors onto millions of miles of road) should give pause to some of the same companies as they attempt to use a similar technology to supplant humans in performing more complicated tasks. Driving—unlike, say, writing news stories or doing customer service for a bank—is fairly straightforward, an activity governed by clearly defined rules that are more or less the same no matter where you are.
The early self-driving demos, which started in the mid-2000s, looked almost like the real thing. Company executives and venture capitalists confidently predicted that all that remained was to figure out how to deal with a handful of so-called edge cases, such as teaching the cars to follow the instructions of emergency workers and to handle foul weather. Much more than $100 billion has been invested since then, the edge cases aren’t solved, and no one is making money on driverless cars.
In retrospect, Cummings says, the early self-driving pioneers mistook demos for nearly finished products, a mistake that she says the chatbot purveyors are making as well. Large language models come close to approximating some types of human output, but they’re also prone to error. Their tendency to “hallucinate” facts, which roughly parallels a persistent problem in driverless cars known as “phantom braking,” hasn’t been fixed yet. And even the most sophisticated chatbots make mistakes at rates that make them unreliable for most kinds of work, at least without continuous supervision. As with driverless cars, you need humans to make sure that the bots aren’t inventing facts in your news story (a big problem for media outlets that have tried to deploy them) or to stop them from spouting obscenities or urging self-harm.
And like robotaxis, the chatbots cost more to run than anyone is willing to pay, causing some, such as Jim Covello, head of equity research at Goldman Sachs, to suggest that the AI boom is actually a speculative bubble. With an implied valuation of almost $160 billion, OpenAI is the richest startup of all time, but it’s losing billions of dollars a year.
Then there’s the question of the market: A robotaxi replaces something that most people find tedious. Today’s slow, somewhat limited driverless cars are clearly useful—at least at their heavily subsidized prices—if you happen to be an introvert or a tourist. But chatbots (think the AI characters Mark Zuckerberg has been inserting into his social network to keep people scrolling) take us further away from the parts of life that are actually, well, real.
See original article by Max Chafkin at Bloomberg
