Google’s Waymo risks repeating Silicon Valley’s most famous blunder
By Timothy B. Lee
10 - 13 minutes
Everyone in Silicon Valley knows the story of Xerox inventing the modern personal computer in the 1970s and then failing to commercialize it effectively. Yet one of Silicon Valley's most successful companies, Google's Alphabet, appears to be repeating Xerox's mistake with its self-driving car program.
Xerox launched its Palo Alto Research Center (PARC) in 1970. By 1975, its researchers had invented a personal computer with a graphical user interface that was almost a decade ahead of its time. Unfortunately, the commercial version of this technology wasn't released until 1981 and proved to be an expensive flop. Two much younger companies—Apple and Microsoft—co-opted many of Xerox's ideas and wound up dominating the industry.
Google's self-driving car program, created in 2009, appears to be on a similar trajectory. By October 2015, Google was confident enough in its technology to put a blind man into one of its cars for a solo ride in Austin, Texas.
But much like Xerox 40 years earlier, Google has struggled to bring its technology to market. The project was rechristened Waymo in 2016, and Waymo was supposed to launch a commercial driverless service by the end of 2018. But the service Waymo launched in December was not driverless and barely commercial. It had a safety driver in every vehicle, and it has only been made available to a few hundred customers.
Today, a number of self-driving startups are aiming to do to Waymo what Apple did to Xerox years ago. Nuro is a driverless delivery startup that announced Monday that it raised $940 million in venture capital. Another, called Voyage, is testing a self-driving taxi service in one of the nation's largest retirement communities.
Right now, these companies' self-driving services aren't as sophisticated as Waymo's. Their vehicles have top speeds of 25 miles per hour. But Apple started out making under-powered products, too, then it gradually worked its way up-market, ultimately eclipsing Xerox. If Waymo isn't strategic, companies like Nuro and Voyage could do the same thing to the pioneering self-driving company.
Google founder—now Alphabet CEO—Larry Page, of course, knows all about Xerox's mistakes and is determined to avoid repeating them. "They weren't focused on commercialization," Page said of Xerox in a 2013 interview with Wired. According to The Information's Amir Efrati, Page has been pushing for Waymo to commercialize since at least 2016.
But Page may be drawing the wrong lesson from the Xerox experience. Xerox did try to commercialize its technology. It just didn't have a good strategy for doing so. And it's not clear that Waymo does either.
Xerox squandered a massive technology lead
Early computers were too expensive to be dedicated to a single user. But in the early 1970s, Xerox researchers realized that steadily falling costs would soon make personal computing economically viable. Xerox's deep pockets freed the researchers to temporarily ignore economic constraints and explore how these computers of the future might work.
By 1975, PARC had produced a futuristic personal computer called the Alto. It had a bitmapped display, a mouse, and a graphical user interface. Xerox researchers developed a word processor with a what-you-see-is-what-you-get interface. Now-standard operations like "cut," "paste," and "undo" were pioneered by PARC researchers, and the Alto had powerful networking capabilities.
Also in 1975, Xerox created a new Systems Development Division to commercialize PARC's computer technology. The team drew up plans for a sophisticated Office Information Systems Architecture. Every worker would have an Alto-style workstation with support for word processing, email, and other office applications. A fast network would connect the workstations to file and print servers.
This vision became a reality in 1981 with the introduction of the Xerox 8010 office system—known informally as the Star. It was a technological marvel, offering capabilities far beyond those found on other personal computers at the time.
There was just one problem. The workstation started at $16,595—more than $45,000 in 2019 dollars. A practical system needed several workstations as well as file and print servers, which could easily cost hundreds of thousands of dollars. Unsurprisingly, the new system didn't sell well.
Apple took a bottom-up approach
As Xerox engineers were developing the Star, a tiny startup called Apple was building much less impressive personal computers. Apple released its first product, a $666.66 kit computer called the Apple I, in 1976. The Apple II, released the next year, came pre-assembled and sold for $1,298 ($5,300 in today's money). It was cheap enough that small companies and even individuals could think about buying one.
Many computing professionals dismissed early "microcomputers" like the Apple II as toys, but they weren't completely useless. A key breakthrough came in 1979, when Dan Bricklin released VisiCalc, the first spreadsheet program, on the Apple II. VisiCalc sold hundreds of thousands of copies and helped make the Apple II an attractive choice for business customers.
Also in 1979, Apple signed a pivotal deal with Xerox. It gave Xerox the right to invest $1.05 million in Apple ahead of Apple's hotly anticipated initial public offering. In exchange, Apple got an in-depth look at the technology Xerox had sequestered at PARC. In a now-famous series of meetings, PARC engineers demonstrated the Alto's advanced capabilities to a group of Apple engineers, who took copious notes.
Apple's first computer with a graphical user interface, the Lisa, was released in 1983 for $9,995 ($25,000 in today's dollars). Like the Star, it was a commercial flop. But unlike Xerox, Apple quickly learned from its initial failure. Realizing that the high price was a dealbreaker for customers, Apple introduced the Macintosh in 1984. Its introductory price of $2,495 ($6,000 in today's dollars) was cheap enough to make it a commercial success.
In his influential 2011 book The Lean Startup, Eric Ries advises companies entering a new market to develop a "minimum viable product"—the simplest product good enough to attract paying customers. This strategy allows a company to begin selling real products to real customers—and get real market feedback—as quickly as possible.
The Apple I kit computer was a classic example of this approach. Only a small group of computer hobbyists could even put it together, and it was far less powerful than other computers already on the market. But the kits were simple enough to be assembled by a few guys in a garage. And creating the Apple I gave Apple practical experience that informed the introduction of the Apple II the next year. Brisk Apple II sales, in turn, created a market for software like VisiCalc, which further boosted demand for the Apple II.
By the time Apple introduced the Lisa in 1983, the company had more than six years of experience selling cheap personal computers. That not only helped Apple recognize the main problem with the Lisa—it was way too expensive—it also gave Apple the practical knowledge and experience to quickly design a much more affordable version.
By contrast, Xerox had little experience in the computer business and no experience selling low-cost personal computers. So when the company's initial product failed in 1981, it struggled to recover. Xerox did eventually introduce a lower-cost version of the Star, called the 6085. But it wasn't ready until 1985, and it cost $4,995—still about twice as much as a Mac.
How to build a minimum viable product for driverless cars
Some of the parallels between Waymo and Xerox are obvious. Like Xerox PARC, Google's self-driving car project created technology that was years ahead of its time. Like the Alto, Google's early self-driving cars were extremely expensive, with rumored costs as high as $250,000 per vehicle.
And like Xerox, Waymo has struggled to commercialize its technology. In November 2017, Waymo announced it had begun testing fully driverless cars on public roads, with plans for a commercial launch in 2018.
But Waymo failed to live up to its own hype. When Waymo launched its "commercial" service called Waymo One in December 2018, every car had a safety driver behind the wheel. That almost certainly means the company is losing money on every ride—which is probably why Waymo has only invited a few hundred people to use it.
On the other hand, there's an obvious difference between personal computers and self-driving cars: an underpowered personal computer isn't going to run anyone over. Obviously, no one should rush a cheap self-driving car to market if doing so would put lives at risk.
But there's more than one way to build a minimum viable product. The toughest problems in autonomy mostly crop up at higher speeds. Waymo's cars have struggled with merging onto freeways and turning left on high-speed multi-lane highways. High speeds mean long stopping distances, which requires expensive long-range lidar. And higher speeds are less safe in general because you're far more likely to die in a 65-mile-per-hour crash than a 25-mile-per-hour one.
25mph speed limit?
So one way to build a minimum viable product for self-driving cars is to build a car that never goes faster than 25mph. That's not fast enough for the kind of general-purpose taxi service Waymo is aiming to create. But it's not hard to think of other applications where low-speed vehicles could be useful: airport shuttles, package deliveries, or taxi services in controlled environments like college campuses or retirement communities.
Indeed, Google considered at least one of these options in the early years of the self-driving car project. Lawrence Burns is a car-industry executive who has been advising Google's self-driving car project since its early days. In a recent book, Burns mentions "large-scale private developments like retirement communities in Florida" as one potential market Google considered during early deliberations about commercialization.
Ultimately, Google didn't pursue the idea. But Sebastian Thrun, the first leader of Google's self-driving car project, didn't forget about it. Thrun went on to found Udacity, an online learning service that offered a course on programming self-driving cars. A few years later, some Udacity staffers decided to use the techniques they were teaching to start a self-driving car startup of their own, called Voyage.
In a recent episode of the Autonocast podcast, Voyage CEO Oliver Cameron credits Thrun with suggesting the idea of building a driverless car service in a retirement community. Cameron took the idea and ran with it. The result: Voyage is currently testing a self-driving car service in The Villages in Florida, one of the nation's largest retirement communities. It has a top speed of 25mph.
Other self-driving startups are working to build lower-speed services, too. In 2016, two engineers left Google's self-driving car project to found a self-driving startup called Nuro (this is another parallel to Xerox: companies like Adobe and 3Com were started by PARC engineers frustrated by the slow pace of commercialization). Nuro is already using small, fully driverless vehicles to deliver groceries in the Phoenix area. They're competing against Udelv, another driverless delivery startup that counts Walmart among its clients.
The Boston-based startup Optimus Ride has been operating a self-driving passenger service in Boston's Union Point planned community since late 2017. Recently, it announced a second deployment in a new mixed-use development in the Virginia suburbs of the District of Columbia. Drive.ai is operating two similar services in the Dallas metro area, while May Mobility is operating low-speed driverless shuttles in three different cities.
A safety driver costs at least as much as a regular taxi driver, plus self-driving cars need expensive sensors and computers. Given that Uber and Lyft are already struggling to earn a profit, a self-driving service with safety drivers is almost guaranteed to lose money.
So Waymo needs to stop using safety drivers in order for its service to become commercially viable. But first, Waymo needs to be confident that its cars can safely handle any situation they might encounter. And given the range of situations Waymo's cars might encounter around the Phoenix area, that's a big challenge.
Voyage and its low-speed peers have a big advantage here. They don't need to master freeway merging, because their cars aren't designed to go on the freeway. Most of them avoid the kind of complex high-speed intersections that have given Waymo's cars so much trouble. So these low-speed services may be able to safely remove their safety drivers far earlier than Waymo can remove its.
In fact, Nuro has already taken this step; it is currently using two custom-built vehicles that don't even have room for a driver. Right now, Nuro has human-driven cars following them around for safety. But Nuro CEO Dave Ferguson recently told Ars he hopes to stop using these chase cars next quarter.
Removing the safety driver enables a self-driving company to make a profit on every ride, which in turn makes scaling up rapidly a possibility. And as a company scales up, it will likely discover myriad small ways—both technological and managerial—to make self-driving systems safer and more efficient. Developers will gain better insights about what features customers care most about and which ones to put on the back burner.
Most importantly, as a company gains more and more real-world experience—and accumulates more and more data—it will gain new insights about how to operate safely at higher speeds. If an initial 25mph service has an excellent safety record—and millions of miles of data to back that up—that could give a company the confidence to upgrade to 30mph. Data from operations at 30mph could lay the groundwork for services up to 35mph. And so forth.
Waymo’s commercialization efforts have been erratic
It will take years for any of these startups to reach fully driverless operations at freeway speeds. Waymo's executives are hoping—and probably expecting—to get their technology perfected much sooner than that, leaving the startups in the dust. But there's reason to be skeptical.
As I mentioned earlier, Waymo completed its first fully driverless trip on public roads way back in October 2015. The demonstration, involving a blind man riding through the streets of Austin, had all the hallmarks of a public media event. But then, bizarrely, Google kept the accomplishment secret for more than a year.
This was right around the time the self-driving project was changing leadership from computer scientist Chris Urmson to car-industry executive John Krafcik. Soon after the Austin demo, Google scrapped the adorable two-seat "Firefly" car used for the demo—it had no steering wheel and a top speed of 25mph—and negotiated a deal with Fiat Chrysler to use conventional Pacifica minivans instead.
Waymo started testing the Pacificas in the Phoenix area in early 2017. In November 2017, Waymo announced it would begin taking drivers out of the driver's seat of its cars. Waymo even paid Jimmy Kimmel to make an ad touting Waymo's driverless capabilities. Waymo promised a commercial launch by the end of 2018. Again, the company seemed to be gearing up to launch a driverless commercial service.
But the December launch of Waymo One was a disappointment. Far from being a driverless service, there were two Waymo employees in each vehicle. And rather than opening the service up to the public, Waymo limited access to people who were already participating in Waymo's existing testing program.
It's certainly possible that Waymo will prove its doubters wrong and start operating a fully driverless service in the next few months. But it's also possible that Waymo's development effort will drag on for years to come. Waymo's current approach might even prove to be a technological dead end. We don't know. And Waymo's leadership doesn't seem to know either.
A pilot program is no substitute for a commercial launch
Meanwhile, some startups might launch fully driverless services as early as this year. Then, like Apple 40 years ago, they could gradually move up-market to higher-speed services. If Waymo is slow enough, some startup rivals might even launch fully driverless Uber competitors before Waymo does.
An obvious objection here is that Waymo is getting significant real-world experience via Waymo's Early Rider testing program. The program, which gives a few-hundred handpicked Phoenix residents a chance to ride in Waymo's cars, resembles a real commercial service in many ways. It features real customers traveling on real routes, and in many cases they're even paying for the rides—though probably nothing like the full cost. This is an expensive way to gain real-world experience, but you might think Waymo can get the same kinds of customer feedback from this kind of pilot program as it would get from running a real commercial service.
But that's wrong. A pilot program is no substitute for an actual commercial product, as Xerox learned the hard way.
While Xerox didn't unveil a full-scale commercial product until 1981, Xerox did make some Altos available to the outside world in the late 1970s. Starting in 1978, Altos were shipped to potential customers in academia, business, and government—a few even wound up at the White House.
Xerox researchers presumably got valuable user feedback that helped further refine the Alto user interface. But these market probes didn't help Xerox answer the most fundamental question: was the Alto compelling enough that Xerox could sell a lot of them at a profit? You can sell a few dozen units of any new technology based purely on its novelty. But that didn't help Xerox figure out how to turn the Alto into a viable commercial product.
Google made a similar mistake earlier this decade with Google Glass. Instead of releasing an actual commercial product, Google released the Explorers Edition, charging early adopters $1,500 each for access to a half-baked version of the technology.
Once again, the Explorer program may have provided Google's engineers with valuable feedback on the user interface. But it didn't help Google figure out how to make a version of Glass that ordinary customers would buy. In fact, Google never found a very compelling answer to that question. Google eventually pivoted to selling Glass for use on factory floors, but the technology has clearly fallen far short of early expectations.
Waymo has fallen into the same trap with the Early Rider program (Waymo One, with its safety drivers and invitation-only customer list, is best thought of as a continuation of the Early Rider program). In an ambitious project like Waymo's, there's an almost infinite list of problems for engineers to work on. Launching a real commercial service is helpful for figuring out which of these problems deserves priority.
Once a company has real paying customers, it's likely to find that most are concerned about one or two issues and asking for one or two main features. There are likely to be a few problems that have a disproportionate impact on the cost of the service or a few bottlenecks that prevent the service from scaling up.
And crucially, developers will discover that many of the challenges that seemed important pre-launch actually are not. They'll find that some engineers were working on features hardly anyone used. They'll learn that others were working on complex, expensive solutions for problems that rarely arise in the real world. As customer needs become clearer, some projects may prove to be completely unnecessary. Management can shelve those projects and re-assign engineers to projects that are addressing more urgent customer needs.
Skip the pilot
This kind of learning rarely happens in a pilot program like Waymo's Early Rider program, because no one knows how closely the final commercial product will resemble the pilot. Mainstream customers might care about different issues than early adopters. Running the service at scale might expose issues that weren't obvious in small-scale tests. Knowing this, the company has to work on many different problems simultaneously. So making rapid progress on any one problem is difficult.
Xerox spent six years perfecting Star features that turned out to be irrelevant, because the product was just too expensive. Bringing a more limited product to market quickly would have given Xerox much-needed clarity about what was actually standing between Xerox and commercial success: the high price.
The same seems to be true for Waymo. The company is spending millions of dollars figuring out how a driverless taxi service should work. But what if the technology to operate a city-wide taxi service is still five or 10 years away? Then Waymo would not only be wasting millions of dollars on unnecessary work, it would be distracting its engineers from launching a less ambitious service. And starting with a less ambitious service may turn out to be the fastest way to get to a city-wide taxi service down the road.
“They keep doing these moonshot projects”
In writing this story, I drew heavily on Dealers of Lightning, Michael Hiltzik's classic 1999 history of Xerox PARC in the 1970s.
In the book, Hiltzik points out that some Xerox insiders pushed for the company to take a more Apple-like approach. In 1976, an executive named Robert Spinrad suggested to his Xerox bosses that the company start with a simple, low-cost version of PARC's technology aimed at the clerical market.
"My notion was that we could build small and develop up," he told Hiltzik in the 1990s. "But I lost every one of those fights. There was no way the Xerox Corporation of that era was going to do anything but full-scale product development."
The problem, Hiltzik wrote, is that "Xerox headquarters expected a product development program to possess a certain minimum heft." Or as a Xerox veteran put it: "Xerox had a hard time understanding anything that wouldn't be a $100 million business."
After reading the book, I asked Hiltzik whether he thought Waymo was falling into the same trap. But Hiltzik was optimistic about Waymo's chances.
"Google is less oriented to its single business model" than Xerox was, Hiltzik argued. Xerox executives, he said, "really wanted PARC to come up with new ways to come up with office equipment"—ideally office equipment that Xerox could rent out the way it rented out photocopiers.
"I think Alphabet's orientation is a little different," Hiltzik said. He argues that Google has a broader portfolio of products and is more savvy about adopting the right business model for each technology.
But Eric Ries disagrees. "I think your analogy is completely dead-on accurate," Ries told me in a recent phone interview.
Google, Ries said, "never developed the discipline of learning how to build new category businesses in non-adjacent spaces. They keep doing these moonshot projects where they hire the smartest people on the planet [and] give them unlimited resources, then hit a stumbling block turning it into a commercial product."
Alphabet executives like to think they're different from the bumbling bureaucrats at Xerox, but it's not clear how different they really are. Like any big company, Alphabet likes to pursue big markets. As Voyage CEO Oliver Cameron put it on the Autonocast, a company of Google's size "might get laughed at" if it initially used its much-hyped self-driving technology to provide a taxi service for retirement communities.
Xerox executives thought that building a big, impressive office computing system was the way to pursue a big market. But in retrospect, we know they were wrong. Often the best way to get big is to start small.