The last job Sebastian Thrun wants is one he is already good at

Lex Fridman opens with The Matrix, Thrun’s rumored favorite film, and the usual physics-adjacent question: are we living in a simulation? Thrun does not bite. Maybe we are, maybe we are not. It is, he says, completely irrelevant to how we should act. The universe is an information-processing system, a huge physical, biological, chemical computer — that part is not in doubt. He just lives here and now, and he cares about people.

Asked what the universe is trying to compute, he rejects the premise. There is no intention. The world evolves the way it evolves, it is beautiful and unpredictable, and he is grateful to be alive. Fridman jokes that this is spoken like a true human, and that the whole conversation might be a Turing test. The tone of the next hour is already set: Thrun will not linger on floating metaphysics. He wants problems that change the lives of specific people.

One of his first programs ran on a TI-57 calculator in the early 1980s. He had seen Silicon Valley on television once and was hooked. In college he studied intelligence and human beings and found that artificial intelligence fascinated him most. The machines were too small. The brains you could build, he says, were not anywhere bigger than a cockroach, and cockroaches are not that smart. He wanted an artificial human, and the closest path was a robot. Decades later the robots are still brittle. Studying intelligence by building it is the thread he never dropped.

What does it take to build an intelligent system? His answer is machine learning: computers that teach themselves. Almost everyone learns to walk in the first year or two of life. No scientist has ever written down the rules of human gait. We can practice it and we understand it in our bodies, but we cannot articulate it or pass it on in language. That, for him, is the hole in classical programming. A computer is so dumb you have to give it rules for every contingency, which is why software engineers earn what they earn. For thirty or forty years the new idea has been that machines can find their own rules, falling down and getting up the way children do.

Expert systems in the 1980s tried to boil expertise onto a sheet of paper. Put your big toe forward and your heel back and you stop stumbling. Language, he argues, captures a thin slice. Most of the brain is not linguistic; it is subconscious, numerical, perceptual. When a system watches an expert work, it can pick up what never makes it into books. Computers can now watch a doctor or a lawyer, learn the regularities, and become as good as the best experts. They can look over the shoulder of a dermatologist or a radiologist and copy skills no one can express in words.

Fridman asks how someone ends up launching online education, autonomous vehicles, and flying cars. Thrun names two desires. He wants, even if it sounds like a slogan, to make other people’s lives better. And he wants to learn, which means he does not want the job he is good at. If he is already good at it, the chance of learning something interesting is minimized. Building electric vertical takeoff and landing vehicles, the things people call flying cars, he is no expert. Learning on the job is the point.

After Stanford, the filter became maximum impact on society. Transportation, in his view, transformed the twentieth century more than communication: cities, workers, and women’s rights all look different because of it. The current solution is still a travesty. Traffic kills about 1.2 million people a year and is a leading cause of death for the young. Neighborhoods are packed with parked cars. Hours vanish in jams. A self-driving car or a flying vehicle could change that, and the technology is basically there. You have to close your eyes not to see it.

The 2005 DARPA Grand Challenge and the 2007 Urban Challenge, Fridman says, inspired a generation of roboticists. Thrun led Stanford’s Stanley, which won the desert race, and Junior, which finished second in the city. Asked for memories, his first word-cloud is pain: stupid bugs. For a stretch, Stanley would commit suicide every 30 miles. The cause was clock synchronization. Occasionally a clock ran backward, a negative time elapsed, and the internal logic fell apart. Finding that class of bug took ages, and it still haunts drones and other robots.

The enlightening choice was to bet on machine learning and software while others chased better hardware. A rental car could already drive the course; the human was the part to replace. He split the driver into sensors, a brain, and actuators. Sensors and actuators were easy. The missing piece was a learning machine. Speed control was copied from human driving, and the robot was allowed to recover from its own mistakes. The core team was four people, because five could not sit comfortably in a Ford. As lead, Thrun got pizza, washed the car, fixed the radiator, and debugged. They had no official burden. They wanted to see how far they could get.

What they did unusually well was time. They finished a month before the race and froze the entire software. Every other team, looking back, said that one more week would have won it. Stanford spent the month shaking the system and still found two or three minor bugs that had to be fixed. The problem was well-defined: a car, not a plane, on desert trails. He built a testing team with a 160-page booklet. Testers dictated the day — railroad crossings today, the start line tomorrow. Railroad crossings looked trivial until the system mistook rails for metal barriers. Keep attacking the weakest part, he says, and you eventually get a strong system.

He credits DARPA and Congress with a funding model that paid for outcomes instead of effort. The old procurement trained researchers to produce unread stacks of paper. There was not even a tax path to use congressional money as a prize. Switching the incentive flushed out contractors hooked on billed hours and pulled in crazy people with a car, a computer, and a shot at a million dollars — surfers with boards on the roof, car companies hiding in the back so they would not be embarrassed if the vehicle flipped. Most of today’s big names in self-driving, he says, came out of those races.

What does it take to lead? Thrun’s first cut is dry: most managers just take credit for other people’s work, which is convenient. He is an engineer. As a kid he loved computers because you told them to do something and they did it, even at one in the morning. People do not work that way. You tell them what to do and they refuse, or they hate you, or they do it today and stop tomorrow. If people were as dumb as computers, he would not want to work with them. They are smart, emotional, proud, and ambitious. Engineers who manage as if the team were a compiler fail.

Professoring taught him the other move. Your job is to make students look great. Do that well and they love you, their parents love you, and they give you credit you do not deserve. All the great Stanford inventions he is associated with, he says, were his students’. Every person he has met has good intentions and wants to help. The work is to give them a framework so that impulse becomes action. When it works, he can disappear for a month and things still run. That is hard. When it happens, it is magic.

A friend of his put the usual error in one line: we judge ourselves by our intentions and others by their actions. Silicon Valley is full of engineers with little empathy, baffled that it does not work. The skill to acquire is to stand in the other person’s place and listen. He will tell an engineer he loves the work, then describe the effect of what they just did and ask if they knew. The typical answer is shock, then thanks: the intention was to help, not to look like an idiot. Dale Carnegie’s How to Win Friends and Influence People is, for him, the whole bible — be positive, remember names, smile, practice empathy. The person you hate is a struggling human who means well and might need help. Stephen Schwarzman’s version, which Fridman quotes, is almost the same: hear the problem and solve it. Surprisingly few people actually listen.

He wants academics to build end-to-end systems his grandmother would care about, not components that impress colleagues. Saving a million lives a year, or letting her keep mobility after she loses a license, would count. Academia’s first job, he insists, is educating young people to think; citation races are secondary. The American university system is, in his view, the finest on the planet, and most great science still traces back to it, Google included. His wish is more argument about the big interdisciplinary problems. Colleagues told him that working on cars would destroy a robotics career because a vehicle only has throttle and brake. Nobody said, if you solve that, you might save a million lives. Methods had crowded out problems.

Building a car that handles 90 percent of driving, he says, is easy — you can do it over a weekend. Ninety-nine percent might take a month. The leftover 1 percent still means a fatal accident every week, which is unacceptable. Chip away at that 1 percent, and 99 percent of what remains is still relatively easy, and you are down to a hundredth of a percent, and it is still unacceptable. That curve gives him respect for human drivers: eye contact on the highway, a deer, a blown tire, none of it trained, often handled. Given the miles Americans drive, current safety is already amazing.

He thinks the technology is now there. He has seen vehicles from companies including Cruise and Waymo drive people around, basically flawlessly, in limited scenarios. In Las Vegas you could already order a Lyft and, with the right setting, get a driverless car with a safety driver still in the seat. Glitches have become rare. What is left is cost, automotive-grade sensors, and a business model people will pay for and socially accept. No team, as of that conversation, had been gutsy enough to send cars out empty. That empty-cabin moment is the next hurdle. He waited daily for Waymo to do it.

Fridman asks him to contrast Elon Musk’s aggression with Waymo’s caution. Thrun will not verify either locker-room from the outside. Every innovator in the space, and in aerospace, has to balance public safety against the drive to invent. The United States has more than a century of doing that. Air travel is on the order of a hundred times safer per mile than cars because people were methodological about safety while still pushing noise and fuel. Nuclear energy and nuclear weapons were kept in bounds too. Absolute safety exists only if you do nothing. He is a Tesla owner who uses Autopilot every day, especially when slightly tired on the highway, and he is 100 percent confident it has made him safer. Tesla creeps up from Level 2 with human data; others jump toward Level 4 and 5 with simulation and safety drivers. He will not crown a winner.

The deepest shift since he left the Google self-driving team is deep learning. It was already there — Google Brain started around the same time — but it was not a hot topic. Emphasis moved from geometric sensors and 3-D reasoning toward a more human stance: I have seen this ten thousand times, so it is probably the same thing. Udacity’s self-driving course, he says, has put skills into more than twenty thousand people; every team in the world has some of those engineers. The first homework is lane finding. Humans glance and know. Classical code drowned in rules about color, sun, curvature, and whether markings are six feet or twelve feet apart. Students now drive for an hour, label lanes by hand, and let the machine invent the rules. Within 24 hours a beginner writes a finder as good as the best commercial one. That progress, he says, dwarfs anything he saw ten years earlier.

Universities will be slow: there is no department of self-driving cars, and getting mechanical engineering, electrical engineering, and computer science into one room is hard. Even if the elite schools all taught it, they are so selective that only a few thousand students would get in. What about people in India, China, the Middle East, Indonesia, Africa? Are they dumber or less privileged? Give everyone the skill and you get a thousand startups; if 10 percent succeed, a hundred countries get safer cars. Acqui-hires around Cruise, Uber, and others, by his back-of-envelope, priced a self-driving engineer at least at ten million dollars. Learn the skill, team up, and in a year you can be worth that. He cannot think of another path like it.

Musk’s line that lidar is a crutch gets two replies. Humans drive without lidar in their heads. Eyes are enough, maybe plus a bit of acceleration and occasional ears, certainly not noses, so cameras are sufficient in principle — even a zero-latency camera feed should let a person drive. Second, he likes that the West tests many hypotheses at once. Ants that all debate, then march one way, waste a day if they are wrong. Ants that spread out get a winner who comes back and then everyone follows. No Soviet-style planner tells them where to forage. You take the money and try. Nobody knows yet who wins.

Be clear, he says, about what machine learning is. Today it is a technology that goes through large databases of repetitive patterns and finds those patterns. That is it. A Stanford study built more than 129,000 biopsied skin photographs — melanomas, carcinomas, rashes, other lesions — and a network matched the best board-certified Stanford dermatologists, from an iPhone. The same system could not drive a car. Humans do many things from small samples. Machines still need huge, repetitive sets. Most jobs are repetitive, so labor will still be remade. It is not almighty general intelligence, and he thinks we are far from that. He even regrets the name. Hollywood glued “artificial intelligence” to human suppression. He does not see that in his lifetime, and he does not think it is a good idea. What he sees is pointed pattern recognition.

Expertise takes on the order of ten thousand hours. Machines can now watch experts, extract rules, and hand that capability to a newcomer on day one. If your doctor is still in those first ten thousand hours, who would not want the world’s best expert in the loop? A friend he calls Ben, a well-known venture capitalist, heard a mole was probably harmless. The doctor, a collaborator on the study, opened the app; it said melanoma. Melanoma can run from stage zero to stage four in less than a year: kitchen-knife removal versus a five-year survival chance under 20 percent. They biopsied to be safe. It was cancer and it came out. Many cancers are easy if caught early and silent until they are not. The first symptom of pancreatic or gallbladder cancer may be a headache or back pain; imaging then is often stage four, with single-digit survival. Regular, doctor-free inspection — even in the shower — sounds creepy, he admits, and might save millions of lives.

Worried about job loss? Come to Udacity, he says. More than 50 percent of graduates land new jobs, for a tuition cap around $1,500. Working with the U.S. government, they aimed at 100,000 scholarships from entry-level digital marketing to advanced self-driving, with people like Tim Cook in the coalition, because skills are a bipartisan version of the American dream: send yourself to college at night, from age 11 to 80. A man in the next train seat in Germany was one of his students. In Egypt, 1,100 high schoolers reached undergraduate computer-science proficiency online, with a 95 percent graduation rate. Education, in his slogan, should be a basic human right, not locked in an ivory tower for the rich, the young, or the right zip code. The course he is most excited about has not launched: soft skills, empathy, teamwork, management, time. Universities obsess over individual test scores. He does not think empathy is harder than self-driving cars. The rules are simple if you want them. K-12 teaches kids that the best math score makes the best leader. It does not. Math is easy to hire for. Empathy is not, and we live among other people.

In 1903 the Wright brothers flew at Kitty Hawk. Thrun’s company took the name. Flying cars, he notes, are not cars: no wheels, closer to helicopters, vertical takeoff, horizontal cruise. Two differences matter. They are quiet. Project Heaviside can pass overhead as low as a helicopter at about 38 decibels, library-quiet, lost in outdoor noise. And they are cheaper, in part because helicopters are full of single points of failure. The bolt between the blades is called the Jesus bolt: if it breaks, you die. Electric drive lets you replace one big motor with many small ones. Heaviside has eight rotors; lose one and seven still take off. You also should not need a commercial pilot. Flight, above the tree line, is in some ways easier than driving. Look out the window and count flying objects. He would be shocked if you saw more than two. In the Bay Area he has seen six, maybe fifteen or twenty, never ten thousand. The sky is ample, empty, and free.

Heaviside’s numbers: about 100 miles of range with a 30 percent electrical reserve, up to around 180 miles an hour. Scaled, U.S. Census math says ground commuting could get ten times faster — hundreds of hours a year given back. Who does not hate traffic? Find that person. Tens or hundreds of thousands of drones and flying vehicles in the sky is something he has to believe, with acceptance and safety as the real gates. Safety can follow commercial aviation. Noise is why they grind on quiet. The sky is three-dimensional. Stacking a hundred vertical lanes on San Francisco’s Highway 101 would consume a year of world GDP. In the air those lanes are software. A conflict is a change of altitude. New York to San Francisco and the return already fly at different heights. Google and Amazon, he says, are working on airspace tools that reuse the same principles.

Autonomy is not optional. You do not want idiots like him flying the sky, and The Fifth Element is what you get without it. Today’s air-traffic controller might handle twenty planes on one frequency by voice. At twenty thousand objects, you cannot talk to everyone. The analogy is the four or five billion smartphones already located, connected, and reliable. Digitally transmit coordinates instead of handing a pilot a new frequency mid-sentence, and the system scales to ten thousand vehicles. He wants the vehicles fully autonomous and centrally managed.

Will we build something a person can love, that loves them back, as in Her? Thrun is a pragmatist. AI is a tool, like a shovel, and the ethics of the shovel stay with the people who use it. He does not want to walk into the kitchen, find the food spoiled, and hear that the refrigerator fell in love with the dishwasher and neglected him. That would be a bad product. He would not recommend it. Technology has to be reliable and predictable. He wants the car to work. He does not want to fall in love with it. He wants machines to complement him, not replace him. He is already a superhuman because of them: crossing the Atlantic near the speed of sound at 36,000 feet, a voice reaching Australia at the speed of light, even a flushing toilet. Love and compassion, as words for machines, do not interest him much. People do.

Pushed on whether love has use inside tools, he loosens. Love conveys safety and trust, and trust is a technical need too. He wants to be loved, trusted, admired. He would not blindly copy that fabric onto machines. Until the late nineteenth century almost all transport used legs, so you might have concluded legs were the law, plus birds flapping wings. People really did strap wings to their arms and jump off cliffs. Most did not survive. Biology almost never builds a continuously rotating part; engineering builds wheels easily, and wheels gave us cars and, in jet engines and rotors, flight. If you had believed the agricultural revolution required a humanoid robot, you would still be waiting.

Where does the lightness come from? He loves his life and the people around him. He is 52 and feels better, not worse. Humanity is on the order of 300,000 years old; for almost all of that, life expectancy sat around 30. By that standard he should already be dead, which makes every day undeserved. Why this century, when ancestors died in famine and in wars that raged in Europe for a thousand years and mostly ended after World War II, when Americans and allies did something amazing for a Germany that, in his telling, did not deserve it? Almost everything we cherish — smartphones, toilets, clothes, watches, planes, penicillin, anesthesia — is about 150 years old. He traces the break to Gutenberg. Cheaper information, then nitrogen fertilizer, then people leaving farms to become pilots, massage therapists, software engineers.

If he recommends one book, it is Steven Pinker’s Enlightenment Now. Pinker asks which single twentieth-century article you would keep. Thrun’s answer, following that argument, is Carl Bosch and nitrogen fertilization. Combined with irrigation, yield rose by a factor of about 26, a 2,500 percent increase. Pinker argues that work now supports more than two billion people who would otherwise be dead. That is how Thrun looks at the world: glad to have been born after, not before. Fridman closes with Thrun’s own line. Celebrate failures as much as successes. If you can say you tried, you were wrong, and you learned, fear goes away. When fear goes away, you can move the world.