Tesla
3 HN recommenders · recommended in 15 months, 2019–2024
3:52:20 · 6.5M views · Apr 22, 2019
Tesla's roughly four-hour Autonomy Day investor event covers its in-house self-driving chip, vision-only approach, fleet data engine and robotaxi plans, with talks by Andrej Karpathy and others.
Cited both as a primary source on Tesla's technical autonomy approach and, later, for Musk's unmet robotaxi predictions.
20 comments from Hacker News
Watch the presentation from 6 months ago, where they explain the decision to build their own hardware for inferring : this 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: this 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.
After the recent "Tesla Autonomy Day" video this 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.
Thanks for the question and the papers. Like some of those companies, we are believers in the data centric approach to ML. But labeling is not our focus (unlike Snorkel, HumanLoop, Prodigy or Cord.Tech). We focus on model diagnosis and mining the best data to improve them. So there are more similarities with Aquarium or Lightly. There is a great talk from Tesla [1] on what they call the "Data Engine" (that probably inspired some of us :)). One of the things we took from it was that in order to truly close the loop on the ML data flywheel, we needed to turn production into a reliable datasource. It had to become accessible, understandable and minable. To achieve this we took the approach of combining ML observability with Active Learning mining frameworks. Combining both is important in our view because Observability tells how the model behaves in the real world and Active Learning finds the right samples to fix / improve the model on real world data. They go hand in hand. Technically, it means that we integrate with serving and labeling platforms. We ingest data both in streaming and batch. We can mine on production streams including on device (for iot use-cases where accessing data is a challenge). We have an extensive set of metrics to understand model behavior in the wild and solve use cases like data drift (detecting it, triggering mining and sending the data for labeling/retraining). And we are geared toward automation. Regarding which data sampling works well and which
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