The real hard part of self-driving isn't the technology, it's the last 5% of surprises
In 1995 a Mercedes drove 1758 kilometres of highway on its own; the machine could already handle 95% of the miles. But as long as the remaining 5% of surprises remain, someone has to stay in the driver's seat, and not a cent of the driver's wages is saved.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Write the rules, or let the machine learn for itself
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 the engineers could imagine into the system in advance: what surfaces are drivable, what to do when you hit an obstacle, one by one. Stanford's secret weapon in Stanley 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 argued over again and again afterwards — should the world be written as rules, or should the machine learn the world from data itself.
The race was lost to a loose filter connection
For the second race Carnegie Mellon prepared two vehicles, thinking it had double insurance, and in the end neither won. Highlander's engine developed a problem two hours into the race, it couldn't climb a slope, and the organisers had to make it pull over and pause. The fault went undiagnosed for years, 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 and 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 bad connection, and on race day the violent desert vibration finally cut the engine's power. The fault cost Highlander more than 40 minutes, and Stanley reached the finish only 20 minutes ahead of it. Even Thrun admitted they won that race purely on luck.
Employee test footage killed the L3 path
In 2012 the team handed cars with self-driving capability to Google employees for their commutes, and before setting off everyone was told they had to keep watching the road and be ready to take over at any moment. When the footage was collected, it turned out they were texting, opening laptops to do work, charging phones, putting on makeup, and some had simply fallen asleep. Once the car has the steering wheel instead of the person, and the first few times go well, people very quickly develop the illusion that ‘it probably won't go wrong’, and once attention has left the driving it cannot be reconnected in an instant. What this footage killed 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 most 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 a red line with a bonus
Levandowski had approached Google competitors under the name 510 Systems, pitching the sensors and positioning equipment then in use on Google's test cars; an internal investigation started, and an email titled ‘It's finally happening’ circulated among engineers. But Page made the completely opposite decision: not only did he not fire him, he ordered that 510 and another Levandowski company be bought together, paying more than 20 million dollars. His own words survive in an internal email: if the driver project succeeds, let Levandowski get rich. That line later became an institution — a bonus plan designed specifically for the self-driving team, not tied to Google's share price but to the self-driving project's own valuation, with an entry recorded for every technical milestone completed. Page managed him not by drawing a line that must not be crossed, but by binding him inside Google with money and interest.
The fuck-off money bought courage, not loyalty
This bonus plan, accumulating by milestone and recalculated as the project's valuation grew, was designed without anyone expecting the team to complete milestones that fast or the valuation to grow that much. One member's accumulated bonus over four years was multiplied more than tenfold at payout, and by the time executives realised the numbers would be astronomical 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 were quitting at any time. Levandowski's share of the plan was worth more than 120 million dollars, while Google had paid only a bit over 20 million dollars to buy all of 510 Systems. But when a colleague once said his life's goal was to earn 100 million, Levandowski thought the target was set too low; his own number was a billion. The money did not buy his loyalty; it gave him the courage to leave.
In 5.6 seconds every line of defence failed
On the evening of 18 March 2018, on a four-lane road in Tempe, 49-year-old Herzberg was pushing a bicycle across the road outside a crosswalk. When the Uber test car hit her it was still travelling at close to 63 kilometres per hour. The crash report reconstructed the 5.6 seconds before impact: radar detected her first, classified as a car; 0.4 seconds later lidar saw her and reclassified her as an unrecognisable object; after that her identity kept changing between car, position object and bicycle, and each reclassification threw away the motion track accumulated before. 1.2 seconds before impact the system finally identified a bicycle and calculated that a collision was imminent, but Uber's system had one more rule — when it judged that only hard braking could avoid the crash, it would not brake immediately but first wait one second to reconfirm. That second was originally meant to reduce needless hard braking. By the time that second had passed, only 0.2 seconds remained before impact.
Three chances to stop the testing, none taken
The first chance came in the autumn 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 crash Uber merged the two jobs into one, and although employees warned that one person would struggle to stay alert for long stretches, the company did not stop. The second chance was a number the whole company could see: Uber wanted its test cars to go 21 kilometres before needing a takeover, but internal documents showed that at the time they often couldn't reach 20 kilometres without one, while in the same period Waymo disclosed to California an average of 9000 kilometres per human takeover. The third chance was an email: five days before the crash, an operations manager warned that the test cars were having frequent damage crashes, that in February almost every other day a car was damaged, and recommended cutting the fleet by 85% and restoring two safety drivers; no executive replied. Three days later that manager resigned, and two days after that the fatal crash happened.
Cruise died on the 6 metres nobody mentioned
In October 2023 in San Francisco, a woman was thrown by a human driver into 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 moved forward again, the system misjudging the impact as a side scrape and initiating its post-collision pull-over procedure, dragging the injured person under the car about 6 metres forward at 11 kilometres per hour before stopping again. What really finished Cruise was not those 6 metres but how the company explained them to regulators afterwards: 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, yet the oral briefing did not include the 6 metres; at the meeting with the California DMV the company 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 crash report filed that afternoon still did not disclose those 6 metres.
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 let him feel he had started a company, and to let me feel I had 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 else was competing to get the technology on the road sooner; he alone never sat down at the table, and he 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 crashes, three kinds of mistake, and three companies chose the same way of responding: when the problem is in front of you, deal with the external narrative first, and reality itself afterwards.
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 is still zero. The driver has not disappeared, and not a cent of the driver's wages is 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 has not disappeared — they have just been moved further 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 at the second DARPA Challenge | 6 hours 53 minutes for the full course | 14:10 |
| Time lost to Highlander's fault | more than 40 minutes | 14:10 |
| The target Levandowski set for himself | 1 billion | 53:03 |
| Cruise's cumulative operating losses | more than 10 billion dollars, with revenue under 500 million dollars | 2:01:26 |
Glossary
- DARPA
- A frontier research agency under the US Department of Defense; the internet was built with its funding.
- L3
- Conditional automation: it lets the human leave the driving task but requires them to take it back within seconds at the most dangerous moment, the most awkward responsibility regime of all.
- LiDAR
- A sensor that fires lasers to measure the distance to objects around it; once the most expensive and most central component of a self-driving car.
- founder mode
- The practice of a founder personally getting into the details rather than delegating management.
How to listen
Founders and investors watching autonomous driving, AI deployment and hard-tech investment, who want to understand why the technology was good enough long ago but the business still doesn't work.
After 1:54 the section comparing Waymo and Tesla's approaches can be fast-forwarded; the crashes and lawsuits earlier are worth more.