I Nailed a Robotaxi Forecast In 2013. Here’s Why Elon Keeps Blowing It

Waymo promotes their plans to serve riders in London in 2026

Waymo

Through a mix of luck and skill, I happen to have made the most accurate prediction of the timelines for robotaxis back in 2013. In this video, I show the prediction, and what factors led into it (and which didn’t) and I also show other’s predictions and explain what’s hard about forecasting revolutionary technology. I also examine what has made Elon Musk, who should know better, to make wrong forecasts about his own technology for almost as long.

(Here’s the video version of this article with video of the original prediction and many other predictions.)

The world of robotaxis is heating up in London. We’ve seen three companies announce plans to deploy robotaxis in London in 2026. First is Waymo, which is testing now with safety drivers. Next is the Chinese company Baidu Apollo, and local favorite, the London-based company Wayve. All announced they would like to deploy by the 2nd quarter of the year.

Back in 2013, a younger me was on stage in London at a conference hosted by Wired Magazine, giving a talk about the future of self-driving cars. I got a question from the editor of Wired UK that I would frequently get in these talks:

You will notice the audience laughed, because my answer is partly meant as a joke, to say that it’s foolish to name a date for something so far in the future. I obviously didn’t know, or think I knew, the exact date and time. I had done this joke before, and was just picking a random date, though deliberately a random date in the mid-2020s. While the message was that the technology doesn’t launch on a schedule, but rather when the people building it are convinced they’ve made it safe enough, I did have a sense of how long that might roughly take, and when I answered the question more seriously, I would say I suspected pilots would get on the road around 2020, and then there would be a scaling “land rush” as companies spread into new cities faster and faster around the middle of the decade.

That is indeed what happened, and not very many saw the same future, so I’m going to talk about how you go about making a prediction like that, what other predictions got made, and give a taste of what’s to come–though for that you will need to look at my other articles and videos for details.

The bad news is that though the companies are ready, the regulators in London are not, and it's now predicted to happen in early fall. So off by a few months.

As I said, there’s definitely some luck in being that accurate. Both in general, and because of survivorship bias. If 100 people make 100 different predictions, at the end of the day, one of them is going to look like a genius. All the people whose predictions were off aren’t making videos about them. I sometimes joke that I must be an amazing futurist because all of my predictions that I remember have been right.

Fortunately, I have other credits to back me up. I had left working for Waymo earlier that

year, so I had insider’s knowledge of the dawn of the industry, and had helped them create the robotaxi strategy they still follow to this day. I got that job by being good at predicting. In 2007, I wrote a book-length series of articles about the future of self-driving, which I shared with Larry Page, who created Google and Waymo. He hired me in to help. The founders of many of the other major players, including Cruise and Zoox and more, and many of the engineers and executives had also read those articles and asked me to advise.

While I also had decades of experience writing software and managing it, nobody could really be prepared for what was to come, or have the tools needed to make a precise prediction. In 2013, the revolution of deep neural networks was young. Alexnet, the first deep learning triumph, was just a year old. Waymo had only done their first experiments using this type of machine learning to help predict the actions of others on the road. Transformers, which would birth LLMs and the modern AI revolution, would not arrive until 2017 (created at Google) and 2018 (at OpenAI.)

In spite of those barriers, many predictions were far too optimistic. At the same time there have always been (and still are) predictions of things taking many decades or even “not in my lifetime.” (I usually responded to those by saying in jest that I was sad to hear they would die so soon, presumably in a fiery car crash.) That’s a classic application of a principle called Amara’s law, that we overestimate progress in the short term and underestimate it in the long term. That’s a principle we taught often at Singularity University, an institution I helped build, which is not an actual university but which helps people understand rapidly changing technology.

Building a self-driving car is one of the grandest software projects in human history. Some of the best have been working on the problem for decades. Full time continuous work began with the first DARPA grand challenge in 2004. The military sponsored these contests to promote work on self driving. The winners of the later contests were hired by Google to start Chauffeur, which became Waymo.

It became clear that a small but talented team could produce a car that did basic driving, starting with desert roads and then city streets. For city streets, if you put a human being behind the wheel not to drive, but to watch and intervene if something was going wrong–known as a safety driver–you could actually get out quickly on public roads. Every so often the safety driver might have to grab the wheel or slam the brakes, but it works. A year after starting, Chauffeur had racked up 100,000 miles in self-driving and 1,000 distinct miles from a selection of roads. It would not be the last time a team would produce impressive early results. Many have done poorly at estimating the gap between a car that gives impressive drives and one ready to be a robotaxi.

After the 3rd grand challenge, Chris Urmson, who is one of the best equipped in the world to answer this question, was interviewed, also by Wired, and asked “the when question” back in 2007.

My first task at Chauffeur in 2010 was to study the project and present my own impressions and forecasts to Chris and Sebastian Thrun. Chris wasn’t thrilled when I played his own clip, because he explained this date of 2020 was the company line of his sponsor, General Motors. He felt something was coming sooner. Much later, Chris became well known for saying his plan was that his son would not need to get a driver’s licence after turning 16 in 2019. Chris would leave Waymo shortly after this prediction, but Waymo would indeed put out its first cars with nobody on board in 2019, though we’re not quite ready in 2026 for teens not to need a licence, though they might in a few cities with a dad as successful as Chris has become as founder of the self-driving truck unicorn Aurora.

Back in the early 20-teens, a lot of folks predicted 2020 as a date for real commercial production. This was the nearly universal prediction of traditional car manufacturers, who were talking about selling you a car that could drive itself. These predictions did not come from their engineers, though. They came from marketing departments and had little basis in reality. The number became so common that the failure to deliver triggered the drop into the first “trough of disillusionment” on the famous Gartner hype cycle.

There have been other bad predictions from people with experience. Lyft President John Zimmer thought in 2016 that Lyft would be mostly self-driving by 2021 and most cars would be switched by 2025. Baidu executives thought they would be selling cars in 2020 and 80% would be autonomous by 2025.

The most famous, and most notoriously wrong predictor has been Elon Musk. He’s been fooled constantly by having an impressive drive in his car. All the way back to 2014, Musk has been predicting it will happen soon. In 2016 he declared his cars had all the hardware needed, and the software would be ready “next year.” He declared would have a robotaxi service running, unsupervised with nobody in the car in Austin on June 22, 2025 (exactly one year prior to my joke date, which is why some readers reminded me of it--I had actually forgotten.) but they were unable, and deployed that day with safety drivers (which they could have done any day for the last 5 years.) They have 17 cars now out with nobody in them, but somebody’s watching remotely, I can assure you. Next year!

So why did it take longer than many people thought, and less time than the skeptics thought? Well, the skeptics are always going to doubt, it’s their job. They look at the present and extrapolate it linearly when important things are changing exponentially. Our brains aren’t used to exponential extrapolation. Even the experts with all their expensive forecasts for things like smartphones and solar energy would make a wrong forecast every year, see it double when they expected 10%, and then make the same wrong slow growth prediction again. Skeptics of exponential trends, if they are smart, might be right, but it’s not usually the way to bet.

The reason so many got it wrong is the phenomenon known as the long tail. There’s just a huge amount of detail work in solving driving. I would not call it “easy” but it’s doable for a smart team to get a basic car on the road doing impressive things in a short time. Google Chauffeur pulled off a car that could go 100,000 miles and 1,000 different miles without interventions in around a year. It took 10 more years to put a car on the road with nobody in it, and 5 more years after that to get really scaling. It took several breakthroughs, none of which anybody could see. But you could see that with time and money, there was a lot of motivation for people to work on those breakthroughs.

That long tail is long. Now that Waymo has 6 years on the roads without supervision and is doing 500,000 rides/week, we still get reports every day of crazy stupid things they did, and even, more rarely, risky things. That will continue for decades to come. If they had to be perfect, they might never make it, but the reality is the human drivers they hope to replace are far less than perfect, so we’re already at the point where the robots, though always learning, can make the roads safer while they work and get better.

I mentioned the Gartner “hype cycle.” Even though it was created 30 years ago, and everybody knows about it, it still manages to surprise most of us every time. The peak of hype is exciting, and we want to believe there won’t be a backlash this time when reality sets in. But it does.

There’s also a lot of other things besides driving in the long tail. Deploying involves a lot of local work on the ground, and infrastructure. That part’s not rocket science, but it takes time and skill and effort.

The robotaxis are on the roads in London, awaiting approval. Waymo is rapidly adding new cities, and other competitors are chomping at the bit to get into the land rush and deploy in cities while they can still be early. They want to be the literal first mover, even though the metaphorical advantage isn’t all it’s cracked up to be. Even so, many of the things I talked about 13 years ago, and also 19 years ago, are still yet to come. Transportation is such a huge part of our lives and our world that there’s little doubt the changes will be dramatic. We love driving metaphors in this business, so strap in, it’s going to be an exciting ride.