Waymo's lead isn't about the best tech — it's that it never made a fatal mistake
Waymo started earliest, cycled through four leaders, and was written off by the industry — yet it ended up the only L4 company still leading the field, because it survived and didn't make the fatal mistakes Cruise and Uber made.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The ragtag team the DARPA Challenge buried
At the first DARPA Grand Challenge in 2004, the prize was $1 million and the course was 240 kilometers. The winner, CMU, died after just 11 kilometers, and not a single team finished. The second edition raised the prize to $2 million; of 20-plus teams, 5 completed the course. The most overlooked was the fourth-place finisher: a team from a Louisiana insurance company. Its two founders read about the event in a magazine and signed up, tapped the person at the company who knew computers best — the one who had originally built their website — to lead it, and gave him $650,000. No one on the team had a PhD, no one understood autonomous driving. They built the equipment in three months and spent six weeks learning AI while passing DARPA review, all while Hurricane Katrina flattened their school. Stanford finished in 7 hours; they finished in 7 and a half. Mengxing says this story tends to get buried.
— Meng XingThe history of autonomy is data-driven methods being slotted in, block by block
Mengxing lays out a clear arc of technical evolution: in 2009 the industry used traditional robotics rule systems; in 2012 AlexNet appeared, making end-to-end perception feasible, and Waymo introduced it on perception in 2014; in 2017 and 2018 the prediction module shifted from rules to data-driven, because they found people don't drive according to lane-line ratios; in 2021 even the hardest part, planning and control, went from hard code to collecting human driving data and using imitation learning. The whole process went from the most mature parts to the least mature, from partial introduction to the whole. Waymo's shipping product today is not fully end-to-end, but the senior technical people on the team hold extremely high ideals about end-to-end, believing it is the simplest, most direct, cleanest and most elegant approach — it's just that L4 has such a high bar for the floor, and end-to-end's floor control is inherently unstable.
— Meng XingFour months in Phoenix: treating the robotaxi like an Uber
Mengxing went to study Waymo in Phoenix in the second year of its public testing. The area was 7 miles by 10 miles, about 70 square miles. At midnight he drove to the Costco entrance and looked for the hardest parking spot, purely hailing Waymo as a user, watching the car come from afar with his initials on the light. The impact came from three things solved at once: high safety, not slow, and unlimited stops — no fixed stops, only a negative list, so you can stop anywhere except where you can't. He also tested the two hardest scenarios: in the Costco parking lot, people backing up, pushing shopping carts, kids playing, with no concept of right of way; and in a villa neighborhood on weekly trash day, bins all over the street, the car weaving around them like obstacles. He also strapped down the lidar, splashed water, and opened the door while moving for extreme tests, and found Waymo could self-diagnose and keep driving, whereas Cruise in the same situation had to send someone out to handle it.
— Meng XingWaymo's home base looks like a sci-fi movie mothership
Mengxing followed a car to Waymo's operations center, in a Phoenix suburb, a big warehouse of about five or six thousand square meters. The entrance looked like an abandoned car factory, with broken cars from the previous generation parked outside, lidar hanging off, unguarded. When the new-generation car arrived, an automatic door opened, and in the distance you could see charging piles and automated equipment. His metaphor is the sci-fi movies he watched as a kid: aliens attack Earth, the mothership door opens, out come very small spacecraft, and after a while they return to the mothership to charge and refuel. He explains the information asymmetry: in Phoenix's 45-degree summer, few people are willing to go out to the suburbs to do this, and there are security guards patrolling. His motivation for doing this research was intense curiosity — he didn't want to look at second-hand information, and the best first-hand information is usually what others won't let you get.
— Meng XingCruise's fall: the dragging incident and dishonesty
Cruise had just obtained its 24×7 paid operating permit in August or September last year when it had an accident at an intersection: a person was crossing the street, the Cruise car didn't move, but a human-driven car on the left knocked the pedestrian in front of the Cruise car, and the pedestrian fell under the car and got caught. By human driving behavior, the best thing would be to stop and not move, but Cruise's mechanism is to pull over in a safety scenario, so it dragged the person a few more meters and rolled them under the car. Mengxing points out two problems: first, the industry had only considered avoiding accidents, rarely how to handle an accident that has already happened, or how to know there was an accident; second, Cruise didn't mention the dragging in its briefing, and when it later came out, it was seen as dishonest, which is the bigger problem. After that, testing was suspended and investigated, Kyle and essentially all the executives left, OpenAI hired many of them, and the original team is basically gone.
— Meng XingCruise was aggressive because it had to prove itself to GM
Mengxing speculates on the mechanism behind the two companies' different styles: Cruise was a company acquired by GM, and its incentive mechanism is tied to performance; it is also one of the best-paying companies in the industry, and GM as a traditional automaker has to support a team more expensive than a tech company long-term, which takes a lot of courage, so it must have presentable results to continuously prove its rationale. This is pressure on Kyle, and pressure on GM CEO Mary Barra, who championed it. The upside is that GM's profits during the pandemic years were all above $10 billion, so it could afford to spend more than $2 billion a year. Waymo was incubated by Google itself, and the people it hired were initially on Google's payroll, so it didn't need to prove itself extra, and preferred to follow the laws of technical progress to set its operating pace, with less pressure.
— Meng XingIn 2020 everyone thought Waymo had fallen behind
Mengxing tells a small story: during the pandemic in 2020, he had breakfast with Sebastian Thrun in Silicon Valley, and in conversation he found that not only Thrun but many people at the time rated Waymo very poorly, thinking it was too slow, that its iteration speed represented the older generation of autonomous driving capability, that it wasn't using the latest technology, and that it hadn't done anything remarkable in the past three or four years; Tesla, by contrast, was updating and iterating every year. This was the common perception in the industry in 2020 and 2021, and even the most authoritative people said similar things. This only reversed recently, and the reversal was very thorough — within L4, Waymo is absolutely the best. Mengxing believes Waymo's model is not the core reason it does well or poorly; the real reason is that not many companies survive, and surviving is the most important thing, and among the survivors it didn't make the fatal mistakes Uber or Cruise made.
— Meng XingUber's exit: lawsuit, founder ousted, fatal accident
Uber exited autonomous driving for three overlapping reasons: first, the lawsuit with Waymo, with Anthony installed as Uber's head, creating internal questions about whether he was really in charge and whether to cut ties; second, founder Travis Kalanick was ousted, Uber became a company run by professional managers, and Dara is an excellent CEO but unwilling to invest heavily in a direction with no visible bottom; third, and most directly, Uber ATG was the only company in the US to have a fatal accident during testing, was suspended for more than a year, was badly damaged, and was judged to be at fault. Mengxing also mentions cost control: a typical US company spends three dollars where a Chinese company spends one; Waymo spends three dollars, Uber might spend five.
— Meng XingL4 has no ChatGPT moment, only a business moment
Mengxing gives two answers. First: this industry has no ChatGPT moment, because the eventual realization of L4 is a gradual process, not just a technical breakthrough but a business and product operation, a linear curve offered bit by bit to more people, not a hockey stick. If you look only at technology, the moment when one car is offered to one person may already have happened, for example Waymo running a 24×7 experience in San Francisco. The second answer is a limit he artificially creates: when there are 10,000 cars, whether deployed in one city or several, running 24×7, where people hail rides just like hailing a ride-hailing car, orders can fill capacity, and the unit economics work, that can be called the commercial ChatGPT moment. He stresses that what Waymo has validated so far is that single-vehicle technology passes, mass production is being tested, and unit economics and scaled operations have not yet been validated.
— Meng XingIn their own words · checked verbatim
As a result, a lot of people cursed him out, saying what does Sebastian count for, father of the self-driving car, you know this industry has been around for years, right, over 50 years of history, he just took the last baton.
结果 被很多人骂了 说斯巴森算什么 无人车之父 你知道这个行业 已经过号年了吧 50多年的历史 他只是最后 接了一棒
Meng Xing5:08
Are we crazy enough to participate in this, and even think about winning it, Probably
Are we crazy enough to participate in this, and even think about winning it, Probably
Meng Xing10:09
It's very much like the sci-fi movies I saw as a kid, where an alien comes to attack Earth, and there's a mothership, and the mothership door opens, and suddenly some very small alien spaceships come out, and after a while they go back to the mothership to charge and refuel.
特别像小时候看到这种科幻电影 有一个外星人来袭击地球 然后有一个母舰 然后母舰门一开 突然出来一些很小的外星人的飞船 然后一会儿就回到母舰去充电加油
Meng Xing29:40
By our human driving behavior, the best behavior should be to stop at this moment, not move anything, don't cause secondary harm.
我们按我们人类的驾驶行为来讲 最好的行为应该是此时停下来 什么都别动 不要造成二次伤害
Meng Xing44:51
I think it's because not that many companies actually survive, and he is one, so surviving is the most important thing first, and among the surviving companies he didn't make a well-known mistake.
我觉得是因为真正能存活下来的公司没那么多 然后他是一个 所以活下来首先是最重要 在活下来公司里面他没犯知名的错误
Meng Xing56:59
First, L4, I think there is no XGPT moment, I'll give you two answers, the first answer is that in this industry there is no XGPT moment, the reason is that L4's eventual realization is a gradual process.
首先L4我觉得没有一个XGPT时刻 我给你两个答案 第一个答案是这个行业里面没有XGPT时刻 原因是因为L4最终它的实现是一个循序渐进的过程
Meng Xing1:05:02
The premise of defeating Uber is that I have to do the same thing as you, and then do it, but if I'm anyway your downstream, I cooperate with you in the end, and I also fight you, then it's actually hard to say defeating is the thing.
打败Uber的前提是说 我得做跟你一样的事 然后就做 但如果说我anyway是你的下游 我跟你最后合作 我也打你也打 那其实很难说打败这件事
Meng Xing1:07:03
Figures
| DARPA Grand Challenge first-edition prize | $1 million | 7:09 |
| DARPA Grand Challenge second-edition prize | $2 million | 8:09 |
| Waymo team size | 2,000 to 2,500 people | 20:37 |
| Domestic L4 team size | about 1,000 people | 20:37 |
| Waymo Phoenix public testing area | 7 miles by 10 miles, about 70 square miles | 21:38 |
| Price Cruise sold to GM | $1 billion | 38:47 |
| Cruise team size at acquisition | 40 people | 38:47 |
| Cruise investment last year | about $2.5 billion | 37:45 |
| Estimated cumulative Waymo investment | $8 to $10 billion | 37:45 |
| Price Amazon paid for Zoox | over a billion dollars | 1:02:01 |
Glossary
- DARPA Grand Challenge
- An autonomous driving competition funded by the US Defense Department's DARPA; no team finished the first edition in 2004.
- AlexNet
- A deep-learning-based neural network that won the 2012 ImageNet competition, abandoning traditional expert rules.
- End-to-end
- Using a single deep learning approach to go directly from perception to planning and control, replacing a modular rule system.
- Moon Shot Project
- Google X's term for seemingly impossible projects that take a very long time to realize.
- Robotaxi
- A mobility service provided by autonomous vehicles, replacing human drivers.
How to listen
Founders and investors watching autonomous driving, mobility and hard-tech investment, and anyone who wants to understand the real operational details and competitive landscape of the L4 route.
The 2-minute show intro at the start and the music and subscription read at the end after 1:12 can be skipped.