Tesla Autonomy Day

After the recent "Tesla Autonomy Day" video [VIDEO] (almost 4h long, but worth a watch), I do rather feel like the incumbent services are about to be in a spot of bother. I think there's a bubble in this sector, and it's going to pop.

Watch the presentation from 6 months ago, where they explain the decision to build their own hardware for inferring : [VIDEO] It's not surprising that they also build the hardware for training. Correct me if I'm wrong, but Google use the same TPUs for training and inference, because the underlying operations are the same : multiply then add numbers. Once Tesla built the hardware for inferring, the design of the hardware for training is probably similar. Unlike Google's TPUs, Tesla have a specific use case for the hardware (computer vision for automotive), and maybe than means they can further optimize the computation pipeline with their own specialized hardware.

I heard from a tesla engineer that they do use lidar to train their models, but that they don’t see the value in deploying it to the fleet. It’s worth watching the tesla event from about a year ago if you haven’t seen it for their full position: [VIDEO] The quick notes are that our roads are set up for human vision and a truly capable level 5 system will require human level interpretation of visual data (at which point lidar becomes redundant). He thinks lidar with non human level visual interpretation won’t get us to full autonomy. They see lidar as a short term half measure, a local maximum that’s a distraction from what has to get done anyway.

Based on what Tesla presented in their autonomy day event ( [VIDEO] ) but it's also what they've been saying publicly for a while now... in a nutshell, they ask the cars to send selected summary data back when interventions happen and in certain other cases.

How can they spend all the money on autonomy [1] while refusing to share any more data [2] beyond the couple of sentences in their "safety reports" [3]? This makes no sense. [1] [VIDEO] [2] [LINK] [3] [LINK]

I was worried about the “diffusion of responsibility” issue, so I rented one on Turo for the day and drove around 200 miles on freeways in Northern California. The user interface is really good, you learn pretty quickly the right amount of force to hold the wheel - then you can let it “autosteer”, but taking control is instantaneous, you just make the movement you want and the car disengages autosteer instantly. I probably took over from the car around 10 times that day, mainly in the left-hand lane with the left edge of the freeway marked with 2-foot-high white concrete barriers, it felt like it wanted to go into them. But that might just have been the car trying to drive “properly” in the center of the lane, while I tend to hug the right side in those situations to have a greater margin of error on the left. It definitely reduces the cognitive load, letting you drive better in my opinion. But there isn’t an idiot-sensor in the car, so like every technology, there will be people who abuse it. The interesting question is whether those same idiots would be safer in a regular car, or if, on balance, the “autopilot+idiot” combination is safer - that seems like an empirical question to me. Even at the current level, it’s a great advance on a regular car (I have no idea how it compares to other expensive cars with various assistive technologies, I usually buy super cheap cars...). I also think that if fully autonomous driving is technically possible, then Tesla will be the one to

> They don't want engineers testing code out on real people's lives. When Tesla pushes and runs code in the shadow mode, it is run in parallel to the "production" version of the software. The actual behavior of a car, even with the test code in, still relies on decisions made by the "production" version of the software[1]. The purpose of such testing is to compare decisions made by the updated code vs those made by the old version, and use this info for fine-tuning the next gen algorithms. [1] [VIDEO]

Actually, they did mention radar during Autonomy Investor Day ( [VIDEO] ). Also, [LINK] says: "A forward-facing radar with enhanced processing provides additional data about the world on a redundant wavelength that is able to see through heavy rain, fog, dust and even the car ahead." So it's rather bizarre that you would speculate about how they're not using radar when they explicitly say they do.