Tesla Autonomy Day: What We Learned
I’m writing this as I watch the stream, and as usual, the start of the stream is late. From having covered a number of these events, that seems like a pretty standard Tesla thing. And since I’m watching on YouTube right now and the numbers have only gone up since the time the stream was supposed to have started until now, maybe they’re on to something.
The pause gives me a bit of time for me to share before we get started that I’m a skeptic of this stuff. Don’t get me wrong, it’s not the electric car side of things that I question — not just have we gone fully electric in the past two years, with a Leaf and Model 3, but I know a bunch of other people who did so after seeing our cars.
Note that this intro section is before watching the Tesla Autonomy Day presentation. For my reactions afterward, read the Summary section.
No, it’s the autonomy stuff that I have gotten a bit more skeptical on. Tesla made a huge Autopilot update right after I got my Model 3 last year, and as an ex-programmer, I felt — and still do feel — that I have a pretty good grip on what Tesla is doing in the way of autonomous driving. From what it appears to me, Tesla is essentially breaking each part of autonomy into its own thing and then solving for that issue, and when they find edge cases, they break those out further, and solve those.
By having the system in cars with people watching them actively, Tesla then can look at the disengagements and try to determine what caused that disengagement. As they improve the system, they will get fewer and fewer disengagements, and once the disengagements are almost non-existent (disengagements due to user error would still remain), the system will theoretically be ready for prime time — once regulators agree.
So, anyway, what I’m looking for with today’s presentation is how Tesla believes it will be able to speed up both the regular deployments of new features, and how the system will move so quickly to learn. The Autopilot updates we’ve had before this point have been impressive, but they aren’t perfect.
Anyway, theoretically, the presentation will be starting soon, so here’s hoping I get some answers to my questions. As I type this, we’re 39 minutes late. That’s longer than usual for Tesla. …
… One minute later, it starts!
Why Does Autonomy Day Exist?
As the presentation starts, I’m first struck by how plain the room is. This looks less like a standard Tesla presentation and more like an investor conference, which is interesting. We get a quick explanation of why Tesla decided to do an Autonomy Day — simply, that Tesla feels that everything has been so focused on the Model 3 that the rest of the company’s story has been lost.
This is a fair point to make about the past year, as I will be the first to say that if the company is valued only on the back of the Model 3, as great as I think my car is, it’s not great enough to justify the entire company at current market valuations. Tesla is working on a bunch of other things.
The Chip
Pete Bannon, the designer of the Full Self Driving chip, came to describe the system, what the timeline was for the system, and the design requirements. He went on to describe the redundancy of the system, which is important in any systems like this. Bannon then dove into the exacting way that the chip works in great detail. As someone who has worked with things like this, it’s all good information, but not anything the average person can understand in a simple way.
The biggest takeaways that I have are first that they looked at the problem to determine how to create a chip to serve the problem. Often, computer chips are made and then programmers create software to make the hardware do what they want. What they did instead was look at the problem and work to solve specifically the problem they had for it, which you can only do by designing a new chip, which is something most companies don’t have the luxury to do, as it is extremely expensive.
Tesla made the investment, and this new hardware allows the cars to increase their power by 21 times above the current chip. That’s a stunning improvement and one that could only be done by feeling that you are so certain of your solution that you can dedicate the money to solve it.
For better or worse, Tesla is extremely certain of this being its solution.
Musk actually said this near the end, that the chip they made was specifically designed for their own self-driving software to run their self-driving solution. This chip is not 21 times faster than any other chip at doing something else, it is specifically 21 times faster than any other chip at running Tesla’s self-driving solution.
Finally, we had the answer many people were really curious about — this chip is now shipping in every Tesla being made.
The Neural Network
Andrej Karpathy joins us to cover the neural network. He starts by describing how a neural network works in a nice and simple way using iguanas and examples. He then starts to explain how the data was annotated, before moving into the things that Tesla specifically looks at.
He goes on to explain the limitations of the self-driving simulations, before diving into a huge description of what sorts of “inaccuracies” they can find, and how they can use those inaccuracies to better train the system. He goes on to explain that they can then start having the system identify inaccuracies itself.