Elon Musk took a drive in one of his cars using the latest version of its full driverless software. The current software uses 300,000 lines of code to tell the cars how to turn left, how to respond to pedestrians and other vehicles, and so forth. Instead of code, the latest beta software is based on neural networks, meaning the computer has learned and will learn how to deal with various situations based on past experiences.
“It’s all nets, baby,” he said during the drive, “nothing but nets.” As I understand it, this is the approach George Hotz was taking when he was working on an his autonomous car software. Hotz tried to train his software without going through all of California’s state licensing requirements, and when the state and the National Highway Traffic Safety Administration tried to force him to comply with their regulations, he left the business.
Musk’s autonomous drive wasn’t perfect, but Tesla is willing to work within federal and state regulatory standards and has a few more resources than Hotz. Musk says the company expects to spend $2 billion (with a B) training its autonomous software this year alone and at least that much next year. Perhaps this means that, by 2025, we will be able to buy cars that can drive themselves in almost all situations.
As I’ve noted before, Tesla’s approach is very different from Waymo/Cruise. Where Waymo/Cruise are using shared vehicles that mostly drive themselves but with oversight by people who can remotely take control if necessary, Tesla is incrementally improving its software so that it can eventually handle all driving conditions. Where I once thought the Waymo model was superior, I am increasingly convinced that the Tesla model will win the race. Shared vehicles simply won’t be economically viable if car owners can just get a software upgrade of their own vehicles to making them safely autonomous.
v12 can’t come fast enough. v11 is way too jumpy around pedestrians.