On the road with George Hotz’s $1,000 self-driving car kit

Has George Hotz settled his feud with Elon Musk?


Sitting in an Aria Hotel penthouse suite with views of the Las Vegas Strip and the vast Mojave Desert beyond, George Hotz was ready to bury the hatchet with Elon Musk. “I’m an Elon Musk fan,” he said. “He jerked me around, I jerked him around, we’re even.”

In a splashy Bloomberg Business feature published late last year, Hotz was less conciliatory. Otherwise known as geohot, the programmer first earned worldwide fame by cracking the iPhone and the PlayStation 3 as a teen. Now, he is the CEO and founder of Comma.ai, a company that aims to get self-driving cars on the road by the end of the year. In the Bloomberg profile, he called Mobileye — the outfit behind the driver-assist tech at Tesla and a number of other major car manufacturers — "absurd" and "a company that’s behind the times." The piece recounts how Musk allegedly promised Hotz a lucrative contract if he delivered a superior product, but kept moving the goal line until the deal fell apart. No matter, Hotz said. Within a few months, he claimed he would post YouTube videos of his car trouncing Tesla’s Autopilot system over the Golden Gate Bridge and on LA’s I-405, near Musk’s home. After the article’s publication, Tesla downplayed Hotz’s accomplishment in a tersely worded "correction":

"We think it is extremely unlikely that a single person or even a small company that lacks extensive engineering validation capability will be able to produce an autonomous driving system that can be deployed to production vehicles," it read. "It may work as a limited demo on a known stretch of road — Tesla had such a system two years ago — but then requires enormous resources to debug over millions of miles of widely differing roads."

Comm.ai team

It was late May when I travelled to Las Vegas to meet Hotz and take a ride in Comma.ai’s highly modified 2016 Acura ILX. It was also more than a few months since Hotz had promised to deliver proof of his victory over Tesla. I asked him where the videos were.

Hotz laughed. "It became less relevant," he said. "Both systems have become quite good. It could still happen."  Sitting on the couch in his tech-nerd-turned-CEO uniform — an untucked button-down shirt, sports jacket, and tattered baggy jeans — Hotz went on: "What I started to realize over the last couple of months is that Tesla’s not our competition — our competition is almost nonexistent. Tesla’s never going to sell aftermarket self-driving systems for Honda Civics. That’s what we’re doing."

Unlike Tesla, Hotz says, Comma.ai is committed to building a product for the general public — or as he likes to put it, deliver "ghost riding for the masses." For less than $1,000 Hotz promises to sell a kit that will let customers transform dumb cars into smart, self-driving ones. Installation will be a cinch, he says — no more difficult or time-intensive than building a piece of Ikea furniture. Though he’s vague about which cars the final product will work on — "if your car has electronic stability control and electronic power steering, there’s a chance we could make it work; it won’t work on your ‘72 Chevy" — he’s confident it’ll ship by the end of the year.

What makes Comma.ai’s autonomous driving system unique is that instead of prioritizing a series of hard rules for the car to obey — "always take turn x at speed y" — Hotz takes a grassroots approach to teaching vehicles the rules of the road: the system tracks driver habits, aggregates them, and then instructs the car to emulate them. If most drivers slow down at a particular rate, Comma.ai cars will too; if most drivers would cut into a corner at this speed, Comma.ai will do the same. "Instead of sitting in a room with a bunch of engineers, talking about how driving works and what we think is a good driver, we actually look, get data from good drivers and even bad drivers so we know what not to do, put that into our machine learning algorithms and get a driving agent," says Hotz.

All of which makes gathering massive troves of data so important to developing the company’s technology — by the end of 2018, Comma.ai hopes to have gathered 1 billion miles of information. Unlike Google or Tesla, Comma.ai doesn’t have a fleet on the road collecting data, so the company is trying something different: on the day of my visit, Hotz announced the Chffr (pronounced "chauffeur") app, which runs on a phone mounted to your dash, tracks your driving habits, and feeds them back into the company’s database. The app and its accompanying site will launch this month, available on the Nexus 6P, Galaxy S6, and Galaxy S7 phones with a wider rollout planned down the line. The more miles you log, the more "Comma points" you earn. Those at the top of the publicly published leaderboard, Hotz says, will be rewarded. ("Comma points are really incredible and if you earn Comma points you’ll feel really great. I promise great things will come to the people at the top of the comma point leaderboard," Hotz told me with a grin. My guess? High scorers will be rewarded with the company’s self-driving kits when they launch.)

Comma.ai’s approach certainly has its upsides. It creates a natural player on the road, giving the impression of a human driver rather than an automated pilot. It will also be dynamically adaptable: instead of spending millions of dollars to map every road, with enough driver data a Comma.ai vehicle will naturally pick up local driving habits. For instance, the company had done most of its car training in the Bay Area, where lanes are demarcated by white lines. But the roads around Las Vegas used Botts’ dots. That posed a problem, but one which the car would, after a certain time of driver observation, adapt to. In places where roads aren’t demarcated at all, Hotz says, the car will learn from driver data to, for instance, follow the traffic ahead.

But crowdsourcing driver ed also poses questions. By aggregating and copying driver habits, Comma.ai is partially assuming that the average driver is a good driver, or at least a model to be imitated. As someone who has driven on a public road more than once, I’d be hesitant to make that assumption — NHTSA estimates that 94 percent of car accidents are caused by human error. One of the promises of self-driving cars, after all, is to create safer roads, not re-create robot replicas of our faulty selves.

Hotz believes that the bad driver outliers will fade out in the mix. "All good drivers are good in the same way, but all bad drivers are bad in different ways," he says, paraphrasing Tolstoy. "The committee driver will be excellent."

George Hotz Photo by Michael Zelenko / The Verge

So, how does the system perform in the real world? From the Aria suite, we descended to find Comma.ai’s Acura in the resort’s parking structure. Aside from a handful of external sensors and a big black comma emblazoned on the hood, the Acura was inconspicuous. Inside, it was a different story. Minus a joystick, the interior looked much like it did in the Bloomberg profile — large swaths of the dashboard paneling had been removed; a warren of cables snaked out from the gutted glove compartment. A massive 21.5-inch screen (Tesla’s is only 17 inches, Hotz is keen to point out) shows the car’s view of the road ahead but also offers all the amenities of a laptop computer — complete with corresponding keyboard. (The entire final product, Hotz told me, will be the size of a breadbox.) At the time of my visit the car was rigged with forward-facing radar, fastened to the bumper by generous amounts of tape, and a wide-angle camera located just behind the rear-view mirror. The production system would also be linked to a GPS unit, although it wasn't for our demo.

We navigated the two blocks from the Aria to Interstate 15, the main thoroughfare from Las Vegas to Los Angeles. Three lanes in either direction, it’s a long, straight, and predictable expanse of road — a perfect testing ground for Comma.ai’s system. We counted down from five and Hotz poked a button on the flatscreen monitor. With a loud ding (Hotz appropriated the buckle-your-seatbelt ding from commercial planes), autonomous driving was engaged.

Similar to basic cruise control, Hotz set his desired speed. Next, he took his hands and feet off the wheel and pedals, and we were off. (Comma.ai’s system doesn’t read signs and isn’t aware of local speed limits.) With seemingly little difficulty, the Acura stayed in its lane; it slowed down in response to cars ahead, and sped back up when it felt safe. Hotz, meanwhile, made a show of the autonomous abilities, pivoting in his seat to talk to the cameraman in the back seat and gesturing effusively to make a point. At one point, he pulled up Netflix on the monitor and we watched the pilot of Unbreakable Kimmy Schmidt (Hotz was unfamiliar with the show, but seemed genuinely amused by its premise), cruising down the freeway at 70 miles an hour. Maybe the most remarkable aspect of driving on the open highway was just how mundane it felt — aside from a couple dicey moments, I felt safe. Lulled into conversation, we missed our exit and drove off into the desert much farther than either Hotz or I had intended.

That’s not to say the system was by any means perfect, and Hotz was keen to remind me that it is still in beta. We let the car attempt on- and off-ramps and it took both too quickly, cutting into the turn so sharply that Hotz had to retake control. The system likes to follow the car ahead, and on one occasion Hotz had to intervene before the ILX took us off the highway to follow an unwitting driver on their commute.

Comm.ai parking lot

And there are other challenges and shortfalls, too: in the limited city driving I witnessed, the system felt less sure of itself. Hotz says that though the car does a good job of recognizing speed variation in vehicles ahead, it can have a hard time recognizing a car at a standstill. That is, to put it mildly, a problem. And there are some capabilities that the system just doesn’t have: it can’t automatically switch lanes, or read stop signs and red lights. Hotz dismisses these features as cheap gimmicks, though he’s open to incorporating them into the final product. Without a GPS unit, it doesn’t know where it is, and can’t actually direct you to a destination.

But considering the company has been around for all of 10 months, Comma.ai has built a fairly impressive system — a sort of advanced cruise control that I found myself longing for days later when I had to drive a couple hundred dull miles by myself.

And Comma.ai has already attracted significant interest: in April, venture capital firm Andreessen Horowitz offered a modest but notable $3.1 million investment. (Perhaps spurring A16Z to action, Cruise Automation, another self-driving startup that sought to produce add-on kits, had just sold to GM for a reported $1 billion.) If Comma.ai can deliver on its $1,000 kit, it might be a significant player as the industry matures. All of which made me wonder whether, somewhere in the back of his mind, Hotz was still gunning for Musk.

As we drove back to the Aria, I asked Hotz about a recent profile of his company in Forbes. The story featured a photograph from his office of an Elon Musk dartboard, two darts lodged in Musk’s forehead. Now that they’ve reached a detente, I asked Hotz if he’d taken the board down. "There are better targets out there," he smiled, his hands off the wheel. "But it’s still a nice decoration."

Letting your Tesla drive itself is a bad idea