ROCKETS, ELECTRIC CARS, solar panels, batteries—whirlwind industrialist Elon Musk has set about reinventing one after another. Thursday, he added another ambitious project to the list: Future Tesla vehicles will run their self-driving AI software on a chip designed by the automaker itself.
“We are developing customized AI hardware chips,” Musk told a room of AI experts from companies such as Alphabet and Uber on the sidelines of the world’s leading AI conference. Musk claimed that the chips’ processing power would help Tesla’s Autopilot automated-driving function save more lives, more quickly, by hastening the day it can drive at least 10 times more safely than a human. “We get there faster if we have dedicated AI hardware,” he said. He didn’t say how far along Tesla is in developing a chip, or when it will start shipping inside vehicles.
This may not be the ideal time for Musk and Tesla to be juggling a new complex and expensive technical project. Some 400,000 people have plunked down $1,000 to join the waitlist for the company’s new Model 3 sedan, but last month Musk conceded production was months behind schedule.
Musk took the stage Thursday in a historic Spanish revival building in Long Beach, California. Alongside him were Andrej Karpathy, Tesla’s director of AI, and Jim Keller, a veteran chip engineer who became vice president in charge of Autopilot hardware last year. Their audience comprised 200 or so lucky attendees of NIPS, a premier academic machine-learning conference that has become a vital bragging and recruiting venue for leading tech companies.
Musk pitched his party as a kind of group hug with the AI community, parts of which he has sometimes been at odds with. He swore that he and Tesla care deeply about the field, and spoke of the company’s need for AI talent in software and hardware. Musk joked self-deprecatingly about his habit of using public appearances to warn that AI poses an existential threat to humanity. “You’ve all heard me sound the alarm bell—there he goes again,” he said, to friendly laughter from the free-drink swilling crowd. “I also think there are things where AI can really be useful, well before you get to godlike uber intelligence.”
As the evening wore on, Musk spoke of his worries about military uses of AI. And he suggested a regulatory agency of some kind might someday require very advanced AI systems to include ethical foundations. But Tesla’s primary use for AI is making sense of data from the cameras, radar, and other sensors through which its Autopilot system perceives the world.
Tesla owners are instructed to only use the system on highways today. Musk has said that a future software upgrade will permit “full self-driving” using the hardware inside existing vehicles. He repeated that claim Thursday, saying that the new chip in the works would improve the reliability of what was already possible. “If you have an order of magnitude more computing power, at a first order approximation that’s an order of magnitude more reliability,” he said.
Reliability will be crucial for self-driving cars. Software hiccups matter when you’re propelling thousands of pounds of machinery around the streets. A Tesla owner died last year when his Model S steered by Autopilot drove into the side of a tractor trailer pulling across the road ahead. The car’s vision system failed to register the white trailer against the bright sky. Tesla’s AI director Karpathy said Thursday that vision algorithms can be troubled by things like trucks with reflective rear ends, or walls painted to appear like roads. Musk hastened to add that he believes cars will soon be harder to fool than people, noting how they can use multiple sensors such as radar and cameras to verify what they’re seeing.
It might seem unlikely that an auto company could design a chip better than a chip company. But Musk’s chip guru Keller told the audience Thursday that nothing on the market is a good fit for Tesla’s mixture of sensors, or the reliability requirements of an auto. “You can get something a lot better if you really design what you want,” Keller said. He previously worked at Apple, AMD, and storied computing pioneer Digital Equipment.
In designing its own chips for AI, Tesla is following other big tech companies. The technique known as deep learning used by Tesla and others for tasks like interpreting camera data is taxing for conventional computer chips. Google, Microsoft, and Apple, have all created custom chips to power deep learning in the cloud or on mobile devices.
Those projects pose a challenge to established silicon suppliers, and Tesla’s chip effort could too. Tesla announced last year that all its vehicles would be powered by a computer for automated driving from Nvidia, the graphics-chip company that has morphed into the leading supplier of high-powered silicon for machine learning. Nvidia said queries about its relationship with Tesla should be directed to the automaker. Tesla declined to comment.
Musk dodged a question from WIRED Thursday about the nature and exact function of the chip his team is working on. But he did say it had features that address shortcomings of graphics chips that limit their efficiency—perhaps a reference to Nvidia. Musk said Tesla engineers calculate their chip will match the performance of existing products while consuming a tenth of the power, and costing a tenth of the price.
The chip project reflects Musk’s high expectations for Autopilot, and progress in competing self-driving projects. Last month Waymo, Alphabet’s self-driving unit, said it no longer needed safety drivers in the front seat of its prototype automated vehicles in Phoenix.
On Thursday night, Musk predicted that his cars will be able to fully drive themselves better than a human in less than two years, and 10 times better in three years. The assembled AI experts roared with delight and astonishment.
CORRECTION, Dec. 10, 1:05 PM: Elon Musk predicted that Teslas would be able to fully drive themselves 10 times better than a human in three years. An earlier version of this article incorrectly said he predicted the cars would drive 100 times better.