The reality is that any competent Autonomous Driving team could successfully create a robotaxi in a year or two if given enough data and computing power required to train the AI to drive. Consequently, the challenge of autonomy ultimately comes down to solving the data problem; how does a company get the billions of miles of real-world driving data required to train an AI? Well, sometimes the obvious answer is the right answer, and the obvious answer to most large data collection challenges is “crowdsource it”. Get information from the largest crowd of contributors possible, so that as each new person joins the crowd the velocity of data gathering increases. When a million+ people anonymously share their driving experience, then averaging 10,000 miles per year per driver results in 10 billion miles of data per year. One obvious but often overlooked benefit to this model is that the resulting data covers all the routes that your contributors actually use, so the resulting AI is already familiar with how contributors handle the risks particular to their area. But that leaves another challenge in that how does one get a million drivers to contribute their driving experiences?
In the case of Tesla, the answer is “sell a million people connected cars equipped with sensors.” At an average vehicle selling price around $52,000 there are about 825,000 cars equipped with all the sensors and computers needed to train the self driving AI, and that fleet should reach one million by late summer. Which is why the race is won, but not over. Tesla already has over three billion miles of data, whereas their nearest competitor, WAYMO, has twenty million. So whether it takes 20 billion miles to train and AI or 50 billion, doesn’t really matter, because Tesla will get there first. And to the dismay of their competitors, Tesla is expected to announce their own battery cell production plans later this month, so they can accelerate the rate at which they scale production. Up to now, their growth has been limited to about 50% per year because that’s as many batteries their suppliers (Panasonic, LG Chem and Samsung SDI), could provide. Later this month Tesla is expected to hold “battery day” for investors, where they’ll announce cell production plans so they can “take control of our own destiny’, in the words of CEO Elon Musk.
Want to hear a story about training AI? Consider the Google owned AI effort called AlphaGo trained to play the most difficult board game in the world, the ancient Chinese board game Go. The number of possible positions in the game is so astronomically large (10170) that even most computer scientists believed no AI could strategize effectively and in a timely manner. (As a point of reference, there are only about 1034 atoms in the universe.) If it could be fast enough it wouldn’t be strategically coherent. If it was strategic enough, it would be too slow. But DeepMind Technologies spent ten years training an AI to play Go by simplifying their algorithms, optimizing task mapping and maximizing their computational efficiency. Google bought DeepMind and provided the computational firepower needed to realize their goals and in 2016 AlphaGo won four out of five games in a televised match against world class Go champion Lee Sedol. While the results shocked the world, there were still doubters.
The next year AlphaGo played a three game match against Ke Jie, the highest rated Go champion in the world, and won all three. The Chinese government viewed the loss by a world Go champion to an American company to be such an insult to it’s national pride that it banned broadcast of the second and third matches. Then it declared AI a national priority and vowed to surpass the US by 2025.
Here’s the thing. Google took all that they learned from the AlphaGo project and applied it to a next iteration AI called AlphaGo Zero while removing all human knowledge aside from the rules of the game, (called training from first principles). After three days of self-learning, Google played the new Zero version against the real-world wining version (AlphaGo), and beat it 100 games to zero. The next iteration after that, Alpha Zero, (it drops the Go, because this version was trained in three games; Go, Chess and Shogi), surpassed AlphaGo after just eight hours of training. It also beat the 2016 world champion chess AI called Stockfish after just four hours of training, and beat the state of the art AI for Shogi (called Elmo) after just two hours of training. This is important because it illustrates how fast learning can take place once the machine learning system is perfected for the task. That’s all the nerd-stuff behind the AI that allows mountains of data to be processed efficiently in pursuit of mastering tasks. In the realm of robotaxis, Tesla has trained it’s self-driving engine how to do one thousand tasks so far, and now needs the data from the long tail of oddball circumstances that affect those tasks to occur in sufficient numbers for the learning to be polished.
VW’s CEO Herbert Diess wrote about Tesla’s Autopilot program in an internal communication that was leaked to a german magazine:
What worries me the most is the capabilities in the assistance systems. 500,000 Teslas function as a neural network that continuously collects data and provides the customer a new driving experience every 14 days with improved properties. No other automobile manufacturer can do that today.
Diess isn’t being explicit about what no other automobile manufacturer can do today, so I’ll do it for him. No other automobile manufacturer can:
- Build a car with one centralized brain. Legacy manufacturers source their systems from suppliers. Each supplier builds their own control units, meaning every car has software from numerous suppliers rather than one centralized brain written by the manufacturer.
- Update cars using Over the Air (OTA) updates. Until there’s a centralized brain, the utility of OTA is limited...only some systems can be updated, so centralizing the brain and proving OTAs should be done simultaneously.
- Collecting data through a fleet that is over 500,000 vehicles strong. Let’s unpack this:
- Collecting data requires sensors, Teslas have 8 cameras, 1 radar and 12 ultrasonic sensors
500,000 >825,000 strong fleet with the sensors already installed — all Teslas have included FSD sensors and computers since October of 2016.
- Updates every two week. The pace of iteration, validation and distribution must be extremely fast, not bureaucratic, tied to model years or distributed through car dealers.
- Function as a neural net. I’ll quote Cathie Wood from Ark Invest, here; “We think Tesla may have the best AI team in the world. Not just in automobile manufacturing, the best in the world.”
The race for robotaxis isn’t over, but the winner is absolutely clear. They have a world class team of AI experts, a custom designed computer chip and neural net, and they’re collecting over 20 million miles of real-world driving data every day.
My personal projection is that by the end of this year YouTube will be full of videos of Teslas driving from suburban driveways to suburban office parking spaces without driver intervention. And by late next year, at least one jurisdiction will have approved the use of Tesla vehicles to operate as driverless robotaxis either along bus routes, or across all routes it has mapped. At that point, robotaxi legalization will begin sweeping the country, and the race will truly be over.
This diary ends a series of diaries I’ve written about what Tony Seba calls the Clean Disruption. The convergence of cost curves for computing power, for battery production, and for solar panels is creating virtuous cycles which feed itself into exponential growth curves. As mentioned earlier in this diary, Tesla has been growing at about 50% per year over the past decade, but they’ve been constrained by battery cell production. That limiting factor will be lifted soon, which will allow production to scale more quickly, and for EV prices to drop below their gas equivalent, (known as price parity.) You’re all familiar with supply and demand curves, and what effect reducing the price of a product below the cost of its substitutes has on demand. The shift to EV’s has been a long haul to this point, but things are going to be very exciting over the next three years. And by then, robotaxis will be operating where you live.