On April 22, 2019, Tesla held an event it dubbed “Autonomy Day.” Intended to highlight the autonomous driving technology the company builds into every Tesla, the event featured presentations by CEO Elon Musk and other top executives and engineers. At the event, Musk said, “I feel very confident predicting autonomous robotaxis for Tesla next year.” He went on to suggest that Tesla would have a million such cars operating on public roads by the end of 2020.By “robotaxis,” Musk meant genuine self-driving cars, capable of operating with no one inside and able to pick up passengers and deliver them to random locations. In other words, a truly robotic version of Uber or Lyft.
This was an astonishing prediction: far out of line with the expectations of virtually every other expert I have talked to. A few days later, I appeared on Bloomberg TV and said that I was “astounded by” Musk’s prediction and that I thought it was “extraordinarily optimistic and perhaps even a bit reckless.” I said this because such an aggressive prediction would almost certainly result in market pressure on Tesla to deliver, and this combined with the company’s ability to provide new features to Tesla owners via software download could be very dangerous if unproven software that purports to deliver fully autonomous capability is suddenly put into the hands of drivers. While it may be fine for a company to have its customers test early versions of a new video game or social media application, this is not a responsible strategy for software that could clearly result in injury or death. Indeed, there have already been fatal accidents involving Tesla’s autopilot feature, which steers, accelerates or brakes the car to stay within its lane but still requires driver supervision. In addition, it seemed clear to me that even in the unlikely event that the company was able to perfect the technology within a year or so, it would take much longer to adequately test the cars and obtain regulatory approval. So, a million operating Tesla robotaxis by the end of 2020 was just not going to happen. Even a single truly autonomous car operating on public roads within that time frame would be astonishing.
Much of the Autonomy Day event was devoted to a discussion of a custom new self-driving microprocessor chip being developed by Tesla. Previously the company has used chips optimized for deep neural networks manufactured by Nvidia. Tesla claimed that its new chip offered unprecedented power, but executives at Nvidia quickly pushed back, pointing out that the latest versions of their AI chips were equivalent to or even faster than the product under development at Tesla.
Nonetheless, as I watched Autonomy Day unfold, it became clear to me that Tesla does indeed have a striking competitive advantage—something that ultimately could allow it to outpace its competitors and be the first company to deploy fully autonomous self-driving cars. This advantage is not a special computer chip, or even an algorithm. Rather—as is so often the case in the field of artificial intelligence—the advantage lies in the data that Tesla controls. Every Tesla is equipped with eight cameras that operate continuously, capturing images from the road and the environment around the car. Computers onboard the cars are able to evaluate these images, determine which ones are likely of interest to the company and then automatically upload these in a compressed format to Tesla’s network. Over 400,000 of these camera-equipped cars are driving on roads throughout the world, and that number is increasing rapidly. In other words, Tesla has access to a truly massive trove of real-world photographic data that none of its competitors can come close to matching.
Elon Musk’s promise of a million robotic taxis operating on roads by the end of 2020 is only the most recent example of in the autonomous vehicle industry. Perhaps because of the centrality of the automobile to our way of life, especially in the United States, no application of artificial intelligence has been subject to as much hype and hyperbolic enthusiasm as the self-driving car. Since the industry’s emergence following the Defense Advanced Research Projects Agency (DARPA) grand challenges in 2004 and 2005, the technology has achieved astonishing progress while at the same time regularly falling short of overinflated expectations. In 2015, it was widely predicted by the most knowledgeable industry insiders that fully autonomous vehicles would be on our roads within five years. Chris Urmson, one of the pioneers of the field, who was formerly the chief technology officer for Google’s self-driving car spinoff, Waymo, and is now CEO and founder of the autonomous driving startup Aurora, famously speculated that his then-eleven-year-old son might have no need to pursue a driver’s license when he turned sixteen. Major manufacturers including Toyota and Nissan likewise promised self-driving vehicles by 2020. All those predictions have now been rolled back. Urmson remains confident and said in 2019 that he expects at least “hundreds” of fully autonomous vehicles to be deployed on public roads within five years, and that there may be 10,000 or more such cars operating within ten years. My own view is that even those predictions could well turn out to be optimistic. I’d say there’s a real danger that truly autonomous cars are going to remain five years in the future for many years to come.
An infinite number of edge cases
The reality is that the routine operation of autonomous cars on both highways and in more urban environments—in other words, situations where things work more or less as expected— has largely been solved. If public roads were anything like the inside of an Amazon warehouse in terms of the overall level of predictability, self-driving cars might already be widely deployed.
The problem, of course, is in the so-called edge cases, the virtually infinite number of unusual interactions and situations that are difficult or impossible for a self-driving car to accurately predict or, in many cases, to correctly interpret. Most self-driving car initiatives depend on highly precise advanced mapping of the streets being traveled. Therefore, unexpected road closings, construction or traffic accidents can create problems. Inclement weather, especially heavy rain or snow, also produces major impediments. But the greatest challenge may be to safely interact with an ecosystem populated by unpredictable pedestrians, bicyclists and drivers. In cities like San Francisco, it’s not uncommon to encounter pedestrians who are distracted or drunk. Even those who are alert often act in ways that are a challenge to interpret, stepping tentatively off the curb in some cases, or in certain neighborhoods, and far more aggressively in others. In densely populated areas, much of the coordination between drivers and pedestrians relies on social interactions that would be very difficult for a self-driving car to understand or replicate. A connection achieved through eye contact, a wave of a hand, pausing midstride to wait for a driver’s acknowledgement and numerous other nuanced behaviors make up a kind of unspoken language that is somehow understood by nearly everyone who shares the road. I think it is quite possible that it may turn out that negotiating these types of interactions is simply beyond the capabilities of today’s deep learning systems. In other words, truly autonomous cars may require technology much further along the path toward general machine intelligence, and that could be a long wait.
Elon Musk and the rest of Tesla’s management team have clearly given a lot of thought to the prospects for full autonomy. Aside from the technology, they’ve also developed a potential solution to the business model problem. At the 2019 Autonomy Day event, Musk described a scheme in which Tesla owners would be able to have their cars participate in a robotaxi service run by the company. Tesla would get a cut of the ride-sharing fee in the same way that Apple generates revenue from its App Store. One interesting thing about this proposal is that it solves the ownership and maintenance problem that might eventually plague companies like Uber and Lyft. Tesla may have found a way to step into the role of a pure internet intermediary, while avoiding the need to own a fleet of cars. Most Tesla owners might not want to share their vehicles with strangers, but if the plan proves viable, many customers would presumably buy Tesla’s vehicles as a business investment, rather than as personal cars.
There is little doubt that self-driving vehicles will someday be one of the most tangible and consequential manifestations of the revolution in artificial intelligence. The technology has the potential to reshape both our cities and our way of life while saving many thousands of lives. However, I think we will need to wait a decade or more before the technology really arrives. Strong evidence of the AI revolution will first emerge in other areas—places like warehouses, offices and retail stores—where the technical challenges are more manageable, the environment is more controllable, the technology is less subject to government regulation and the consequences of an error are far less dire. It is very exciting to imagine, however, that a single software update from Tesla could prove me wrong.
Excerpted from Rule of the Robots: How Artificial Intelligence will Transform Everything. Copyright © 2021. Available from Basic Books, an imprint of Hachette Book Group, Inc.
Martin Ford is the founder of a Silicon Valley-based software development firm. He is also the author of Rise of the Robots, which won the FT Business Book of the Year Award, The Lights in the Tunnel; and Architects of Intelligence.