The Invention of a Car's Self-Driving Feature
About ten years ago, the EV automaker Tesla started their Self-Driving project. Their CEO Elon Musk knew that one day it could be a reality. What he didn't know at the time was that AI technology would be needed for self-driving to be way better than a human, and that was still some years away.
From their Version 1 all the way up to Version 11, the software was the traditional kind of software... lines of code created by human programmers. Each line was a "if this, then do that" statement. Version 11 had about 300,000 lines of code. But the self-driving feature still wasn't good enough.
Then, back in January of 2024, there was a breakthrough. AI had come a long way, and it was now good enough to let it do the programming. But how to train it to do this? The brilliant software engineers at Tesla invented a new approach... show videos to the AI, and tell the AI to write software to mimic what it saw, in this case, driving. Obviously you had to show the AI videos of excellent driving, so there was some human curation involved initially. But then they figured out how to automate that laborious process. And that's when the magic started happening... with Version 12.
No more thousands of lines of human created code. Now there were neural networks, one for "perception" and one for "planning". Perception is what the car sees from its 360 degree view of its surroundings thanks to its seven strategically placed cameras (and a microphone that can listen for emergency vehicles). Planning is how to control the steering, brakes, and acceleration (and signaling) based on what the car's computer sees. So the perception network interprets the car’s surroundings, while the planning network decides the best actions – like turning or stopping – based on that data. Then the driving is informed by the Navigation computer that plans the car's route (sort'a like the GPS sat nav devices many car owners have).
But to make an AI training computer required lots of "compute power". Why? Because although the AI can watch videos a billion times faster than we can, once it's watched all the videos to make a new version, it then has to make the version, and that was taking way too long with the computer that Tesla had at the time. If there was an "edge case" that was discovered that needed to be incorporated into the software, you couldn't just stick it into the existing version... you had to add the video that addresses the issue to the existing video data set that the AI gets to "watch", and then compile a new version, and that could take a very long time to compile. Too long.
So Elon musk decided to spend billions of dollars on making the world's biggest and fastest computer. This way, new self-driving versions could be made a lot quicker, and since this translates into potential lives saved, time is of the essence.
From inception to turning it on, the new computer – named Cortex – was made in nine months! And this time-frame shocked all computer experts. Normally, a project of this size and scope would take years. But Musk is good at hiring the best and brightest, and because of how Musk runs his companies, Tesla had the cash to buy the needed hardware without having to go through the lengthy process of borrowing money or doing "capital raises". And today, they're expanding Cortex to more than double its capacity! Why? Because Version 12 worked, and it worked really well. Now all that was needed was faster compiling (the AI processing tons of video and "writing" new neural networks, and then creating the software that is sent to the millions of Tesla cars on the roads or to its Robotaxi ride-hail fleet).
|
What is a "neural network" A neural network is a computer system inspired by how the human brain works. It’s made up of layers of connected "nodes" that process information, such as recognizing patterns in data. For Tesla’s self-driving software, it learns from examples – like images, sensor data, or driving patterns – to make decisions, such as steering or braking, without being explicitly programmed for every situation (those thousands of lines of human created code). Think of it as a brain that gets smarter with experience (the new videos that are shown it). |
And with each new version of the software, it gets better and better. This is called the "March of 9s". It's not enough for self-driving software to be 99% better than the best human drivers; it needs to have a bunch of nines following the 99, as in 99.999% which equates to improved reliability (reducing error rates as close to zero as possible). Technically, it can never be 100%, so the more nines, the better. And that's where we are today with Version 14 that Tesla owners have, and that the Tesla cars being used for their ride-hail service in Austin Texas and other cities have.
And since the state of Texas was convinced of Version 14's safety, it allowed a Tesla Model Y that rolled off of the Tesla factory's assembly line to drive itself straight to the customer's home; a 30 minute drive... with no one in the car (and no issues). Here's a still shot taken while that history-making car was driving itself...
And no, it wasn't
being driven remotely by a human.
Since people can only see in one direction at a time, but a self-driving vehicle can see in all directions simultaneously, and has better night vision than a human, and since a self-driving car's reaction time is faster and more precise than a human's (many deadly highway crashes are “successful avoidance followed by loss of control”), and since the computer can't drive distracted or take its eyes of the road or drive drunk or fall asleep at the wheel, a properly trained self-driving car using neural networks and high quality cameras can drive much safer than we humans can. And if that helps to reduce the 42,000 annual deaths in the U.S. due to auto accidents due to driver error, that's a welcome bit of technology.
REALITY: Today there are hundreds of thousands of consumer-owned Tesla vehicles being driven around by the car's computer. And you can see YouTube videos of these cars preventing accidents on the highway (where most deaths normally occur). Below is a chart with the latest safety stats ("AutoPilot" is the name of Tesla's self-driving highway software). Spend a moment to take in this amazing info...
The difference can
mean avoiding an accident or killing someone
One person died from
an auto accident while you read this article.
The majority of these deaths are preventable if the car is doing the
driving.
It's important to know how the media lies about self-driving cars, especially Tesla's
When a self-driving taxi (Tesla or Waymo) has an "accident", the regulations require that they report it, even if it's a tiny fender-bender that doesn't involve another vehicle. From December-January 2026, Tesla Robotaxis filed five reports; these are publicly available. Here's the details...
Now look at how the media portrayed those incidents (they call them "crashes"). Even though these Tesla self-driving cars are 7 times safer than a human driver (not an opinion), the media uses math shenanigans to make it look like the Teslas are worse than humans (comparing all reported fender-benders made by humans and Tesla Robotaxi isn't apples-to-apples because humans don't tend to report fender-benders such as these, i.e., math shenanigans). Here are four out of dozens of similar pieces...
More info about Tesla's self-driving feature
More enlightening EV info...
Be sure to look at Total Cost of Ownership when considering an EV