Rockets, electric cars, solar panels, batteries-industrial whirlpool Elon Musk has reinvented itself one after the other. On Thursday, he added another ambitious project to the list: Future Tesla vehicles will run their stand-alone IA software on a chip designed by the automaker itself.
"We are developing customized AI hardware chips," Musk told an AI room experts from companies such as Alphabet and Uber on the sidelines of the world's leading IA conference. Musk said the processing power of the chips would help Tesla's Autopilot automatic driving function, saving more lives, more quickly, accelerating the day when it can drive at least 10 times safer than a human being. "We arrived faster if we have dedicated AI hardware," he said. He did not say how advanced Tesla is in the development of a chip, or when it will begin shipping it inside the vehicles.
Maybe this is not the ideal time for Musk and Tesla to juggle a new and complex technical project. Some 400,000 people have invested $ 1,000 to join the waiting list for the company's new Model 3 sedan, but Musk last month granted production months behind schedule.
Musk took the stage on Thursday at a historic Spanish Renaissance building in Long Beach, California. Next to him were Andrej Karpathy, director of artificial intelligence at Tesla, and Jim Keller, a veteran chip engineer who became vice president in charge of the autopilot hardware last year. His audience consisted of approximately 200 lucky NIPS attendees, an academic machine learning conference that has become a place of pride and vital recruitment for leading technology companies.
Musk launched his party as a kind of group hug with the AI community. parts of which sometimes has been in disagreement. He swore that he and Tesla care deeply about the field, and spoke of the company's need to have artificial intelligence talent in software and hardware. Musk joked with contempt about his habit of using public appearances to warn that AI poses an existential threat to humanity. "Everyone has heard me sound the alarm, there it goes again," he said, for a friendly laugh from the crowd that drank freely. "I also believe that there are things in which artificial intelligence can really be useful, long before we reach super-divine intelligence."
As the night progressed, Musk spoke of his concerns about the military uses of artificial intelligence. And he suggested that a regulatory agency of some kind might one day require very advanced AI systems to include ethical foundations. But Tesla's main use for AI makes sense for the data from cameras, radars and other sensors through which its autopilot system perceives the world.
Tesla owners are instructed to use the system only on roads today. Musk has said that a future software update will allow "total self-management" using the hardware within existing vehicles. He repeated that statement on Thursday, saying that the new chip in the works would improve the reliability of what was already possible. "If it has an order of magnitude greater computing power, in a first-order approximation that is an order of magnitude greater reliability," he said.
Reliability will be crucial for driverless cars. Software problems matter when you are driving thousands of pounds of machinery on the streets. A Tesla owner died last year when his Model S driven by Autopilot drove to the side of a tractor trailer that crossed the road. The vision system of the automobile did not record the white trailer against the bright sky. Tesla's IA director, Karpathy, said Thursday that vision algorithms can be affected by things like trucks with reflective rear ends or walls painted to look like roads. Musk was quick to add that he believes cars will soon be harder to fool than people, and pointed out how they can use multiple sensors like radar and cameras to verify what they are seeing.
It may seem unlikely that an automotive company could design a better chip than a chip company. But Musk's chip guru, Keller, told the audience on Thursday that nothing in the market is a good choice for Tesla's mix of sensors, or the reliability requirements of a car. "You can get something much better if you really design what you want," Keller said. He previously worked at Apple, AMD, and on the legendary computer equipment pioneer Digital Equipment.
When designing its own chips for AI, Tesla follows other large technology companies. The technique known as deep learning used by Tesla and others for tasks such as interpretation of camera data is demanding for conventional computer chips.
Google, Microsoft and Apple created custom chips to enhance deep learning in the cloud or on mobile devices.
Those projects represent a challenge for established silicon suppliers, and the Tesla chip effort could also. Tesla announced last year that all its vehicles would have a computer for automatic driving from Nvidia, the graphics chip company that has become the leading supplier of high-power silicon for machine learning. Nvidia said that inquiries about his relationship with Tesla should be directed to the automaker. Tesla refused to comment.
Musk dodged a WIRED question on Thursday about the exact nature and function of the chip his team is working on. But he said he had features that address the shortcomings of graphics chips that limit their efficiency, perhaps a reference to Nvidia. Musk said that Tesla engineers calculate that their chip will match the performance of existing products while consuming a tenth of the power and cost one tenth of the price.
The chip project reflects Musk's high expectations for autopilot and progress in competitive self-driving projects. Last month, Waymo, the stand-alone unit of Alphabet, said it no longer needed security drivers in the front seat of its prototype automated vehicles in Phoenix.
On Thursday night, Musk predicted that his cars could handle themselves better than a human in less than two years, and 100 times better in three years. The AI experts gathered laughed with delight and amazement.