On a stretch of highway outside San Francisco, an Acura ILX sedan cruises along at 55 miles an hour. Jazz from the 1940s softly plays through the car’s stereo. Just underneath the rearview mirror is a rectangular box fitted with a smartphone, projecting lines of neon green onto the image of the nighttime expressway being taken in via its built-in camera. The driver’s hands aren’t on the wheel, because this car is driving itself, but not thanks to technology from the myriad companies all racing to develop self-driving cars. Instead this car owes its autonomous driving ability to George Hotz, a 27-year-old Carnegie Mellon dropout who is turning America’s public roads into his own personal laboratory.
One year ago, Hotz was about to sell the Comma One—a $999 product for that very purpose—when a letter came from the National Highway Traffic Safety Administration threatening daily fines if his startup put untested self-driving tech on the road.
So Hotz scrapped it and, last November, gave his technology away for free, releasing an open-source, self-driving platform called Openpilot. He also released open-source plans for Neo, a smartphone-powered device which plugs into certain compatible Honda and Acura models, and can control the car’s gas, brakes, and steering, and navigate using Openpilot.
“Self-driving cars need nothing but engineers in order to solve it. It does not need manufacturers, regulators, any of these people,” says Hotz. “The best thing they can do is stay out of the way of the engineers.”
Audacious claims are a Hotz trademark, surpassed only by his escapades. He first gained notoriety at age 17, then known by his hacker alias “geohot,” as the first person to unlock the iPhone so it could be used on any wireless carrier. He’s a person whose bold thoughts are simultaneously clear in their aim but are cascading so swiftly they can sound disorganized.
Hotz says 73 drivers are using Openpilot. More than 1,000 people, he says, use Chffr, a dash-cam smartphone app that records drives and uploads them to the data center in the basement of a three-story house in San Francisco that Hotz’s 12-person startup works out of.
Combined, those two tools have collected more than 1 million miles of driving data, and that’s the heart of Comma.ai’s self-driving coup.
His engineers use this data to train behavioral models of driving. Whereas other companies tinkering with self-driving cars are teaching their cars by defining different road conditions and manually labeling driving data—this is a passing lane; this is a stop sign—Comma.ai relies on the patterns and behaviors of everyday drivers to train the models used by Openpilot.
“Most people are trying to take data and label it. He’s not doing that. He’s letting every-thing go into the computer, and if it stands still, it’s a tree; if it moves it’s a pedestrian. He’s counting on his computer to know that on its own,” says Jim McPherson, a lawyer and consultant on self-driving technology.
Hotz says this is the only way to properly define driving: observe how drivers drive, catalog that data, and use it to teach other cars. (He likens driving to a dance, albeit a mechanized one on asphalt with a two-ton object involved in nearly 40,000 deaths in the U.S. in 2016 alone.)
This is where Comma.ai’s approach makes observers of self-driving technology worried. If average drivers are teaching Openpilot how to drive, won’t Openpilot then have a skewed, even unsafe, picture of driving?
“If somebody decides to put one of these on their car and something goes wrong in public, that’s where I start to get scared,” says Venkat Krovi, a professor in Clemson University’s Department of Automotive Engineering. “If a car hasn’t come across a certain case before, it’s hard to know how it’ll respond to that case.”
But Hotz strongly rebuts this, claiming as Comma.ai gathers more data, the system will grow smarter. He’s currently hacking a Toyota Prius to make it Openpilot-compatible, and Comma.ai recently began selling the Panda, an $88 device that plugs into a car’s onboard diagnostic port to collect vehicle metrics on braking force, steering, speed, and more, which he’ll then use to reverse-engineer cars that don’t yet support Openpilot.
The grand plan for Comma.ai, which is still running on $3.1 million of seed money raised from venture capital firm Andreessen Horowitz, is to sell an aftermarket hardware-software package, and then sell subscriptions for its network of driving data.“We’re planning to make this even more open-source than it was, because at the end of the day, I care about winning,” Hotz says. “A lot of people who are trying to compete with us are quickly going to find out that we’re going to open-source their whole business.”
The Big Players
Waymo: The autonomous-vehicle company spun out of Google currently has logged 3 million self-driving miles on public roads. In April, the company started offering rides in its autonomous minivans to people in Phoenix.
Tesla: Every Model S and Model X vehicle sold today comes equipped with Tesla’s Hardware 2 package of -sensors, cameras, and radar. Through future software upgrades, CEO Elon Musk claims Hardware 2 will be capable of full Level 5 autonomy, which would enable drivers to completely disengage from driving and observing the road.
Cruise Automation: Acquired by General Motors in 2016 for more than $1 billion, Cruise Automation is currently testing 30 Chevrolet Bolt cars equipped with self-driving technology in Detroit, Scotts-dale, and Hotz’s backyard of San Francisco.
A version of this article appears in the Sept. 1, 2017 issue of Fortune with the headline “Get Out of His Lane.”