The Real Hard Part of Self-Driving Isn't the Technology, It's the Last 5% of Surprises
In 1995 a Mercedes drove 1,758 km of highway on its own; the machine could already handle 95% of the miles. But as long as the remaining 5% of surprises remain, a person must stay in the driver's seat, and not a cent of the driver's wages can be saved.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Write rules, or let the machine learn on its own
At the second DARPA Challenge in 2005, Carnegie Mellon and Stanford gave two completely different answers. Carnegie Mellon's approach was to write as many situations as engineers could imagine into the system in advance: what surfaces are passable, how to handle obstacles, prepared one by one. Stanford's Stanley had a secret weapon that wasn't hardware but machine learning, letting the machine watch how human drivers drive and then learn from driving data what kind of ground can be crossed. At the time it looked like just two teams' ways of solving the problem; looking back today, this is exactly the core question the industry has argued over again and again—should the world be written as rules, or should the machine learn the world from data on its own.
The race was lost to a poorly connected filter
For the second race Carnegie Mellon prepared two vehicles, thinking it had double insurance, but neither won. Highlander's engine developed a problem two hours into the race, it couldn't climb a slope, and the organizers had to make it pull over and pause. The cause of the failure was investigated for years without being found, until 2017 when the old car was pulled out of storage for a commemorative display; while cleaning the engine bay someone touched a small component on the wiring, engine power suddenly dropped, and pressing it again killed the engine outright—it was a filter between the engine control computer and the fuel injectors, with a poor connection, and on race day the severe desert vibration ultimately caused the engine to lose power. This failure cost Highlander more than 40 minutes, while Stanley reached the finish only 20 minutes faster. Even Thrun admitted they won this race purely on luck.
Employee test footage killed the L3 path
In 2012 the team gave Google employees cars with self-driving capability for their commutes; before departure everyone was told they must keep watching the road and be ready to take over at any time. When the footage was collected, it turned out: they texted, opened laptops to do work, charged their phones, put on makeup, and some fell asleep. As soon as the car takes over the steering wheel and the first few times go well, people quickly develop the illusion that "it probably won't crash," and once attention leaves driving it cannot be reconnected in an instant. What this footage negated was not self-driving itself, but the product design of "let the machine handle the vast majority of driving, then leave the most sudden and dangerous few seconds to the human." Google therefore cut the entire assisted-driving product line around 2013 and turned completely to full autonomy.
Page replaced the line with a bonus
Levandowski had approached Google competitors under the name 510 Systems, pitching the sensors and positioning equipment being used on Google's test cars; an internal investigation was launched, and an email titled "It's finally happening" spread like wildfire among engineers. But Page made a completely opposite decision: not only did he not fire him, he ordered the purchase of both 510 and another Levandowski company, paying more than $20 million. His own words remain in an internal email: if the driver project succeeds, let Levandowski get rich. This later became a system—a bonus plan designed specifically for the self-driving team, not tied to Google's stock price but to the self-driving project's own valuation, with a mark recorded for each technical milestone completed. Page managed him not by drawing a line that could not be crossed, but by binding him inside Google with money and interests.
The fuck-off money didn't buy loyalty, it bought courage
This bonus plan, accumulating by milestone and recalculated as the project valuation grew, was designed without anyone expecting the team to complete milestones so fast or the valuation to grow so much. One member's accumulated bonus over four years was multiplied more than tenfold when paid out; by the time executives realized it would compute to astronomical figures, the contracts had long been signed. By 2015 the team privately called it "fuck-off money"—with that money, you could tell your boss you didn't want to work anymore at any time. Levandowski's share in this plan was worth more than $120 million, while Google had spent only a bit over $20 million to acquire all of 510 Systems. But once a colleague said his life goal was to earn 100 million, and Levandowski thought that target was too low; he set his own number at a billion. This money did not buy his loyalty; it gave him the courage to leave.
In 5.6 seconds every line of defense failed
On the evening of March 18, 2018, on a four-lane road in Tempe, 49-year-old Herzberg was pushing a bicycle across the road at a non-crosswalk location. When the Uber test car hit her, its speed was still close to 63 km/h. The accident report reconstructed the 5.6 seconds before impact: radar first detected her, classified as a car; 0.4 seconds later lidar saw her and reclassified her as an unrecognizable object; after that her identity kept changing among car, position object and bicycle, and each reclassification lost the previously accumulated motion track. At 1.2 seconds before impact the system finally identified her as a bicycle and calculated an imminent collision, but Uber's system had another rule—when it judged that only hard braking could avoid the accident, it would not brake immediately but first wait one second to reconfirm. This second was originally meant to reduce needless hard braking. By the time that second passed, only 0.2 seconds remained before impact.
Three chances to stop the testing, none taken
The first chance was in the fall of 2017: previously every test car had two safety drivers, one in the driver's seat watching the road and one in the passenger seat watching the screen; five months before the accident Uber merged the two jobs into one, and although an employee warned that it was hard for one person to stay alert for long, the company did not stop. The second chance was a number visible across the company: Uber wanted its test cars to drive 21 km independently before needing a takeover, but internal documents showed that at the time they often needed a takeover before even 20 km, while in the same period Waymo disclosed to California an average of 9,000 km before a human takeover. The third chance was an email: five days before the accident, an operations manager warned that test cars were frequently having damage accidents, with a damaged vehicle almost every other day in February, and recommended cutting the fleet by 85% and restoring the two-safety-driver configuration; no executive replied. Three days later this manager left, and two days after that the fatal crash occurred.
Cruise fell on the unspoken 6 meters
In October 2023 in San Francisco, a woman was thrown by a human driver and landed in the path of a Cruise self-driving car; Cruise braked hard but did not stop in time, and its wheels ran over her. The car stopped for a moment and then continued forward; the system misjudged the impact as a side scrape and initiated its post-collision pull-over procedure, dragging the injured person under the car forward about 6 meters at 11 km/h before stopping again. What really killed Cruise was not these 6 meters, but how the company explained these 6 meters to regulators after the accident: at a meeting with the National Highway Traffic Safety Administration the next day, more than 100 people inside the company already knew the vehicle had moved forward after first stopping, but the oral report did not include these 6 meters; at a meeting with the California DMV it chose to let the video speak for itself, and the video the DMV saw stopped at the point where the vehicle first braked. The formal accident report submitted that afternoon still did not disclose those 6 meters.
In their own words · checked verbatim
I wanted to hire Levandowski. He wanted to start his own company. So I came up with a way to make him feel like he'd started a company, and to make myself feel like I'd hired him.
我想招莱万多司机 他想要自己开公司 于是呢我就想了一个办法 让他觉得自己创业了 也让我觉得自己招到了他
Kalanick1:04:15
In all the time Levandowski worked for Uber, all Uber got from him was this one lawsuit.
在莱万多斯基为优布工作的全部时间里面 优布从他那里得到的只有这一场官司
Uber trial lawyer1:31:37
He didn't call an Uber, he didn't get into any test car, and he didn't agree to take part in this self-driving race. Everyone was competing to get this technology on the road sooner; only he never sat down at the gambling table, yet paid the most irreversible price.
他没有叫一辆优步 没有坐进任何的测试车 也没有同意参加这场自动驾驶竞赛 所有人都在争着让这项技术更早上路 只有他从来没有坐上赌桌 却付出了最无法挽回的代价
Three accidents, three kinds of mistakes, three companies chose the same way of responding: when the problem is right in front of you, deal with the external narrative first, then deal with reality itself.
三次事故 三种错法 三家公司选择了同一种应对方式 当问题摆到面前的时候 先处理对外的叙事 再处理现实本身
So 95% is 95% technically, but in business it's still zero. The driver hasn't disappeared, and not a cent of the driver's wages can be saved.
因此95%在技术上是95% 在生意上却依然是零 司机没有消失 司机的工资一分钱都省不下来
The driver may have left the driver's seat, but the person hasn't disappeared—they've just been moved farther away.
司机可能离开了驾驶座 但人并没有消失 只是被搬到了更远的地方
Figures
| Winning time of the second DARPA Challenge | 6 hours 53 minutes to complete the course | 14:10 |
| Time lost by Highlander's failure | more than 40 minutes | 14:10 |
| Levandowski's target for himself | 1 billion | 53:03 |
| Cruise's cumulative operating loss | more than $10 billion, with revenue under $500 million | 2:01:26 |
Glossary
- DARPA
- Advanced Research Projects Agency under the U.S. Department of Defense; the internet was created under its funding.
- L3
- Conditional automation: allows the human to disengage from driving but requires them to take back control within seconds in the most dangerous moments—the most awkward responsibility framework.
- LiDAR
- A sensor that fires lasers to measure the distance to surrounding objects; once the most expensive and central component of self-driving cars.
- founder mode
- The practice of founders personally getting into the details rather than delegating management.
How to listen
Founders and investors watching self-driving, AI deployment and hard-tech investment, who want to understand why the technology has long been sufficient but the business still doesn't work.
After 1:54, the section comparing Waymo and Tesla's approaches can be fast-forwarded; the accidents and lawsuits earlier are worth more.