The First Person to Hack the iPhone Built a Self-Driving Car. In His Garage

Share this…

A few days before Thanksgiving, George Hotz, a 26-year-old hacker, invites me to his house in San Francisco to check out a project he’s been working on. He says it’s a self-driving car that he had built in about a month. The claim seems absurd. But when I turn up that morning, in his garage there’s a white 2016 Acura ILX outfitted with a laser-based radar (lidar) system on the roof and a camera mounted near the rearview mirror. A tangle of electronics is attached to a wooden board where the glove compartment used to be, a joystick protrudes where you’d usually find a gearshift, and a 21.5-inch screen is attached to the center of the dash. “Tesla only has a 17-inch screen,” Hotz says.

He’s been keeping the project to himself and is dying to show it off. We pace around the car going over the technology. Hotz fires up the vehicle’s computer, which runs a version of the Linux operating system, and strings of numbers fill the screen. When he turns the wheel or puts the blinker on, a few numbers change, demonstrating that he’s tapped into the Acura’s internal controls.

Inside George Hotz’s Acura ILX

After about 20 minutes of this, and sensing my skepticism, Hotz decides there’s really only one way to show what his creation can do. “Screw it,” he says, turning on the engine. “Let’s go.”

As a scrawny 17-year-old known online as “geohot,” Hotz was the first person to hack Apple’s iPhone, allowing anyone—well, anyone with a soldering iron and some software smarts—to use the phone on networks other than AT&T’s. He later became the first person to run through a gantlet of hard-core defense systems in the Sony PlayStation 3 and crack that open, too. Over the past couple years, Hotz had been on a walkabout, trying to decide what he wanted to do next, before hitting on the self-driving car idea as perhaps his most audacious hack yet.

“Hold this,” he says, dumping a wireless keyboard in my lap before backing out of the garage. “But don’t touch any buttons, or we’ll die.” Hotz explains that his self-driving setup, like the autopilot feature on a Tesla, is meant for highways, not chaotic city streets. He drives through San Francisco’s Potrero Hill neighborhood and then onto Interstate 280.

With Hotz still holding the wheel, the Acura’s lidar paints a pixelated image on the dash screen of everything around us, including the freeway walls and other cars. A blue line charts the path the car is taking, and a green line shows the path the self-driving software recommends. The two match up pretty well, which means the technology is working. After a couple miles, Hotz lets go of the wheel and pulls the trigger on the joystick, kicking the car into self-driving mode. He does this as we head into an S curve at 65 miles per hour. I say a silent prayer. Hotz shouts, “You got this, car! You got this!”

The car does, more or less, have it. It stays true around the first bend. Near the end of the second, the Acura suddenly veers near an SUV to the right; I think of my soon-to-be-fatherless children; the car corrects itself. Amazed, I ask Hotz what it felt like the first time he got the car to work.

“Dude,” he says, “the first time it worked was this morning.”

Breakthrough work on self-driving cars began about a decade ago. Darpa, the research arm of the Department of Defense, sponsored the Grand Challenge, a contest to see how far autonomous vehicles could travel. On a course through the desert in the inaugural 2004 event, the top vehicle completed just 7 of 150 miles. In subsequent years, the vehicles became quite good, completing both desert and city courses.

It took a great deal of sophisticated, expensive technology to make those early cars work. Some of the Grand Challenge contestants lugged the equivalent of small data centers in their vehicles. Exteriors were usually covered with an array of sensors typically found in research labs. Today, Google, which hired many of the entrants, has dozens of cars in its fleet that use similar technology, although dramatic advances in computing power, sensors, and the autonomous software have lowered the overall cost.

Artificial-intelligence software and consumer-grade cameras, Hotz contends, have become good enough to allow a clever tinkerer to create a low-cost self-driving system for just about any car. The technology he’s building represents an end run on much more expensive systems being designed by Google, Uber, the major automakers, and, if persistent rumors and numerous news reports are true, Apple. More short term, he thinks he can challenge Mobileye, the Israeli company that supplies Tesla Motors, BMW, Ford Motor, General Motors, and others with their current driver-assist technology. “It’s absurd,” Hotz says of Mobileye. “They’re a company that’s behind the times, and they have not caught up.”

Mobileye spokesman Yonah Lloyd denies that the company’s technology is outdated. “Our code is based on the latest and modern AI techniques using end-to-end deep network algorithms for sensing and control,” he says. Last quarter, Mobileye reported revenue of $71 million, up 104 percent from the period a year earlier. It relies on a custom chip and well-known software techniques to guide cars along freeways. The technology has been around for a while, although carmakers have just started bragging about it. Tesla, in particular, has done a remarkable job remarketing the Mobileye technology by claiming its cars now ship with “Autopilot” features. Tesla’s fans have peppered the Internet with videos of its all-electric Model S sedans driving themselves on freeways and even changing lanes on their own. (In an e-mailed statement, Tesla spokesman Ricardo Reyes writes: “Mobileye is a valued partner, but supplies just one of a dozen internally and externally developed component technologies that collectively constitute Tesla Autopilot, which include radar, ultrasonics, GPS/nav, cameras and real-time connectivity to Tesla servers for fleet learning.“)

Hotz camera kit

Hotz plans to best the Mobileye technology with off-the-shelf electronics. He’s building a kit consisting of six cameras—similar to the $13 ones found in smartphones—that would be placed around the car. Two would go inside near the rearview mirror, one in the back, two on the sides to cover blind spots, and a fisheye camera up top. He then trains the control software for the cameras using what’s known as a neural net—a type of self-teaching artificial-intelligence mechanism that grabs data from drivers and learns from their choices. The goal is to sell the camera and software package for $1,000 a pop either to automakers or, if need be, directly to consumers who would buy customized vehicles at a showroom run by Hotz. “I have 10 friends who already want to buy one,” he says.

The timing for all of this is vague. Hotz says he’ll release a YouTube video a few months from now in which his Acura beats a Tesla Model S on Interstate 405 in Los Angeles. The point of the exercise is twofold. First, it will—he hopes—prove the technology works and is ready to go on sale. Second, it will help Hotz win a bet with Elon Musk, chief executive officer of Tesla.

Hotz lives in the Crypto Castle. It’s a white, Spanish-tiled house, which, other than the “Bitcoin preferred here” sticker on the front door, looks like any other in Potrero Hill. The inside is filled with a changing cast of 5 to 10 geeks. The bottom floor largely belongs to Hotz. His room is a 15-by-5-foot closet with a wedged-in mattress. The space is lined with shelves packed with boxes, car parts, towels, and a case of women’s clothes left behind by a former resident. There’s a living room in the back with couches and a television. “I hate living alone,” Hotz says. “I was playing Grand Theft Auto with my roommates last night. It was super fun.”

Just a couple feet from his closet is the garage where Hotz works. His two-monitor computer sits on a desk next to a water heater. On a wooden table, there’s a drill, a half-dozen screwdrivers, a tape measure, some black duct tape, a can of Red Bull, and a stack of unopened mail. Most of the garage is taken up by the white Acura. Hotz has decorated its hood with a large, black comma, and the back bumper reads “”—the name of his new company—in big, black letters. “A comma is better than a period,” he says.

George Hotz in his garage

Hotz grew up in Glen Rock, N.J. His father oversees technology for a Catholic high school, and his mother is a therapist. “Like, Freud talking and stuff,” Hotz says. At 14, he was a finalist in the prestigious Intel International Science & Engineering Fair for building a robot that could scan a room and figure out its dimensions. A couple years later he built another robot called Neuropilot that could be controlled by thoughts. “It could detect different-frequency brain waves and go forward or left based on how hard you were focusing,” he says. The next year, 2007, he won one of the contest’s most prestigious awards, a trip to attend the Nobel prize ceremony in Stockholm, by designing a type of holographic display. “I did terrible in high school until I found these science fairs,” he says. “They were the best thing for me. I could build things, and there was the salesmanship, too, that I loved.”

He hacked the iPhone in 2007 while still in high school and became an international celebrity, appearing on TV news shows. Three years later, he hacked the PlayStation 3 and released the software so others could use it. Sony responded by suing him, and the two parties settled their feud shortly after, with Hotz agreeing never to meddle with Sony products again. These achievements were enough to earn him a profilein the New Yorker when he was 22. “I live by morals, I don’t live by laws,” Hotz declared in the story. “Laws are something made by assholes.”

But Hotz wasn’t a so-called black-hat hacker, trying to break into commercial systems for financial gain. He was more of a puzzle addict who liked to prove he could bend complex technology to his will.

From 2007 on, Hotz became a coding vagabond. He briefly attended Rochester Institute of Technology, did a couple five-month internships at Google, worked at SpaceX for four months, then at Facebook for eight. The jobs left him unsatisfied and depressed. At Google, he found very smart developers who were often assigned mundane tasks like fixing bugs in a Web browser; at Facebook, brainy coders toiled away trying to figure out how to make users click on ads. “It scares me what Facebook is doing with AI,” Hotz says. “They’re using machine-learning techniques to coax people into spending more time on Facebook.”

On the side, Hotz produced an application called towelroot, which gave Android users complete control over their smartphones. The software is free to download and has been used 50 million times. He kept himself entertained (and solvent) by entering contests to find security holes in popular software and hardware. In one competition, Pwnium, he broke into a Chromebook laptop and took home $150,000. He scored another $50,000 at Pwn2Own by discovering a Firefox browser bug in just one day. At a contest in Korea designed for teams of four, Hotz entered solo, placed first, and won $30,000.

By the fall of 2012 he was bored with the contests and decided to dive into a new field—AI. He enrolled at Carnegie Mellon University with the hope of attaining a Ph.D. When not attending class, he consumed every major AI research paper and still had time for some fun. At one point, the virtual-reality company Oculus Rift failed to man its booth at a job fair, and Hotz took it over, posing as a recruiter and collecting résumés from his fellow students. None of this was enough to keep him interested. “I did two semesters and got a 4.0 in their hardest classes,” he says. “I met master’s students who were miserable and grinding away so that they might one day earn a bit more at Google. I was shocked at what I saw and what colleges have become. The smartest people I knew were in high school, and I was so let down by the people in college.”

Although Hotz makes his university experience sound depressing, it left him brimming with confidence and eager to return to Silicon Valley. He’d devoured the cutting-edge AI research and decided the technology wasn’t that hard to master. Hotz took a job at Vicarious, a highflying AI startup, in January to get a firsthand look at the top work in the field, and this confirmed his suspicions. “I understand the state-of-the-art papers,” he says. “The math is simple. For the first time in my life, I’m like, ‘I know everything there is to know.’ ”

He quit Vicarious in July and decided to put his conviction to the test. A friend introduced him to Musk, and they met at Tesla’s factory in Fremont, Calif., talking at length about the pros and perils of AI technology. Soon enough, the two men started figuring out a deal in which Hotz would help develop Tesla’s self-driving technology. There was a proposal that if Hotz could do better than Mobileye’s technology in a test, then Musk would reward him with a lucrative contract. Hotz, though, broke off the talks when he felt that Musk kept changing the terms. “Frankly, I think you should just work at Tesla,” Musk wrote to Hotz in an e-mail. “I’m happy to work out a multimillion-dollar bonus with a longer time horizon that pays out as soon as we discontinue Mobileye.”

“I appreciate the offer,” Hotz replied, “but like I’ve said, I’m not looking for a job. I’ll ping you when I crush Mobileye.”

Musk simply answered, “OK.”

“For the first time in my life, I’m like, ‘I know everything there is to know’ ”

Hotz has filled out since his days as a scrawny teenage hacker, although he dresses the same. Most often, he wears jeans and a hoodie and shuffles around the garage in socks. He has a beard of sorts, and some long, stray whiskers spring out from his Adam’s apple. His demeanor doesn’t match the slacker get-up. Hotz’s enthusiasm is infectious, and he explains just about everything with flailing hands and the wide eyes of someone in a permanent state of surprise.

It’s easy enough to draw a connection between Hotz and Steve Wozniak. Like Hotz, Wozniak began his hacking days on the fringes of the law—in the early 1970s, before he and his pal Steve Jobs founded Apple. Woz was making small devices that let people place free long-distance phone calls. Even in Silicon Valley, few people are equally adept at hardware and software. Woz was, and so is Hotz.

Hotz began working in earnest on his self-driving technology in late October. He applied online to become an authorized Honda service center and was accepted. This allowed him to download manuals and schematics for his Acura. Soon enough, he’d packed the glove compartment space with electronics, including an Intel NUC minicomputer, a couple GPS units, and a communications switch. Hotz connected all this gear with the car’s main computers and used duct tape to secure the cables running to the lidar on the roof.

There are two breakthroughs that make Hotz’s system possible. The first comes from the rise in computing power since the days of the Grand Challenge. He uses graphics chips that normally power video game consoles to process images pulled in by the car’s camera and speedy Intel chips to run his AI calculations. Where the Grand Challenge teams spent millions on their hardware and sensors, Hotz, using his winnings from hacking contests, spent a total of $50,000—the bulk of which ($30,000) was for the car itself.

The second advance is deep learning, an AI technology that has taken off over the past few years. It allows researchers to assign a task to computers and then sit back as the machines in essence teach themselves how to accomplish and finally master the job. In the past, for example, it was thought that the only way for a computer to identify a chair in a photo would be to create a really precise definition of a chair—you would tell the computer to look for something with four legs, a flat seat, and so on. In recent years, though, computers have become much more powerful, while memory has become cheap and plentiful. This has paved the way for more of a brute-force technique, in which researchers can bombard computers with a flood of information and let the systems make sense of the data. “You show a computer 1 million images with chairs and 1 million without them,” Hotz says. “Eventually, the computer is able to describe a chair in a way so much better than a human ever could.”

The theory behind this type of AI software has been around for decades. It’s embedded in products consumers take for granted. With the help of Google, for example, you can search for “pictures of the beach,” and AI software will comb through your photo collection to turn up just that. Some of the biggest breakthroughs have come in voice recognition, where smart assistants such as Apple’s Siri and Microsoft’s Cortana can pick up a person’s voice even in noisy situations. The same goes for instantaneous translation applications, which have largely been taught new languages via deep-learning algorithms that pore over huge volumes of text. With his car, Hotz wants to extend the same principles to the field of computer vision.


In the month before our first drive on I-280, Hotz spent most of his time outfitting the sedan with the sensors, computing equipment, and electronics. Once all the systems were up and running, he drove the vehicle for two and a half hours and simply let the computer observe him. Back in his garage, he downloaded the data from the drive and set algorithms to work analyzing how he handled various situations. The car learned that Hotz tends to stay in the middle of a lane and maintain a safe distance from the car in front of him. Once the analysis was complete, the software could predict the safest path for the vehicle. By the time he and I hit the road, the car behaved much like a teenager who’d spent only a couple of hours behind the wheel.

Two weeks later, we went on a second drive. He’d taken the car out for a few more hours of training, and the difference was impressive. It could now drive itself for long stretches while remaining within lanes. The lines on the dash screen—where one showed the car’s actual path, and the other where the computer wanted to go—were overlapping almost perfectly. Sometimes the Acura seemed to lock on to the car in front of it, or take cues around a curve from a neighboring car. Hotz hadn’t programmed any of these behaviors into the vehicle. He can’t really explain all the reasons it does what it does. It’s started making decisions on its own.

In early December, Hotz took me on a third ride. By then, he’d automated not only the steering but also the gas and brake pedals. Remarkably, the car now stayed in the center of the lane perfectly for miles and miles. When a vehicle in front of us slowed down, so did the Acura. I took a turn “driving” and felt an adrenaline rush—not because the car was all over the place, but because it worked so well.

Hotz’s approach isn’t simply a low-cost knockoff of existing autonomous vehicle technology. He says he’s come up with discoveries—most of which he refuses to disclose in detail—that improve how the AI software interprets data coming in from the cameras. “We’ve figured out how to phrase the driving problem in ways compatible with deep learning,” Hotz says. Instead of the hundreds of thousands of lines of code found in other self-driving vehicles, Hotz’s software is based on about 2,000 lines.

The major advance he will discuss is the edge that deep-learning techniques provide in autonomous technology. He says the usual practice has been to manually code rules that handle specific situations. There’s code that helps cars follow other vehicles on the highway, and more code to deal with a deer that leaps into the road. Hotz’s car has no such built-in rules. It learns what drivers typically do in various situations and then tries to mimic and perfect that behavior. If his Acura cruises by a bicyclist, for example, it gives the biker some extra room, because it’s seen Hotz do that in the past. His system has a more general-purpose kind of intelligence than a long series of if/then rules. As Hotz puts it in developer parlance, “ ‘If’ statements kill.” They’re unreliable and imprecise in a real world full of vagaries and nuance. It’s better to teach the computer to be like a human, who constantly processes all kinds of visual clues and uses experience, to deal with the unexpected rather than teach it a hard-and-fast policy.

In the coming weeks, Hotz intends to start driving for Uber so he can rack up a lot of training miles for the car. He aims to have a world-class autonomous vehicle in five months, something he can show off for Musk. He’s heard that Teslas struggle when going across the Golden Gate Bridge because of the poor lane markings. So he plans to film a video of the Acura outperforming a Tesla across the bridge, and then follow that up by passing the final test on I-405 in Los Angeles where Musk lives. Hotz’s YouTube videos get millions of views, and he fully expects Musk will get the message. “I’m a big Elon fan, but I wish he didn’t jerk me around for three months,” he says. “He can buy the technology for double.” (Says Tesla spokesman Ricardo Reyes: “We wish him well.”)

There’s really no telling how effective Hotz’s software and self-learning technology ultimately will be. His self-funded experiment could end with Hotz humbly going back to knock on Google’s door for a job. “Yeah, of course there will be skepticism,” he says. “This is part of a great adventure. All I can say is, ‘Watch.’ ”

George Hotz in his Acura ILX

Sitting cross-legged on a dirty, formerly cream-colored couch in his garage, Hotz philosophizes about AI and the advancement of humanity. “Slavery did not end because everyone became moral,” he says. “The reason slavery ended is because we had an industrial revolution that made man’s muscles obsolete. For the last 150 years, the economy has been based on man’s mind. Capitalism, it turns out, works better when people are chasing a carrot rather than being hit with a stick. We’re on the brink of another industrial revolution now. The entire Internet at the moment has about 10 brains’ worth of computing power, but that won’t always be the case.

“The truth is that work as we know it in its modern form has not been around that long, and I kind of want to use AI to abolish it. I want to take everyone’s jobs. Most people would be happy with that, especially the ones who don’t like their jobs. Let’s free them of mental tedium and push that to machines. In the next 10 years, you’ll see a big segment of the human labor force fall away. In 25 years, AI will be able to do almost everything a human can do. The last people with jobs will be AI programmers.”

Hotz’s vision for the future isn’t quite as bleak as The Matrix, where robots mine our bodies for fuel. He thinks machines will take care of much of the work tied to producing food and other necessities. Humans will then be free to plug into their computers and get lost in virtual reality. “It’s already happening today,” he says. “People drive to work, sit in front of their computer all day, and then sit in front of their computer at home.” In 20 years, the sitting in front of the computer part will be a lot more fun, according to Hotz, with virtual worlds that far exceed anything we’ve managed to build on earth. “Stop worrying about the journey,” he says. “Enjoy the destination. We will have a better world. We will be able to truly live in a society of the mind.”

Hotz started the autonomous car work because he sees it as Step 1 in the revolution. Transportation is an area where AI can have a massive impact. He hopes to take his technology to retail next, building systems that provide flawless self-checkout at stores. His desire to have AI take over so many jobs stems partly from a near-religious belief in the power and ultimate purpose of technology. “Technology isn’t good or bad,” he says. “There are upsides like nuclear power and downsides like nuclear bombs. Technology is what we make of it. There’s a chance that AI might kill us all, but what we know is that if you’re on the other side of technology, you lose. Betting on technology is always the correct bet.”

All this talk represents an evolution in Hotz’s hacker ethos. He used to rip apart products made by Apple and Sony, because he enjoyed solving hard puzzles and because he reveled in the thought of one person mucking up multibillion-dollar empires. With the car, the retail software, and the plans to roil entire economies, Hotz wants to build a reputation as a maker of the most profound products in the world—things that forever change how people live. “I don’t care about money,” he says. “I want power. Not power over people, but power over nature and the destiny of technology. I just want to know how it all works.”