Tesla AI Day 2022
Tesla2 HN recommenders · 2022–2026
2.3M views · talk · dated · Uploaded 2022-10-01
Tesla's full AI Day 2022 presentation, covering the Tesla Bot hardware and software, Full Self-Driving planning and perception, the occupancy network, data labeling, simulation and Dojo training infrastructure.
HN uses it as primary-source evidence on how Tesla's vision-based driving stack and training pipeline work.
2:25:52The spelled-out intro to neural networks and backpropagation: building micrograd
Andrej Karpathy
18:40But what is a neural network? | Deep learning chapter 1
3Blue1Brown
54:37How Self-Driving Cars will Destroy Cities (and what to do about it)
Not Just Bikes
15:01What Happened to Horses Is Happening to Us
CGP Grey
1:30:28AlphaGo - The Movie | Full award-winning documentary
Google DeepMind
27:14Transformers, the tech behind LLMs | Deep Learning Chapter 5
3Blue1Brown
51:22USENIX Security '18-Q: Why Do Keynote Speakers Keep Suggesting That Improving Security Is Possible?
USENIX
1:08:40Lecture 1 | Machine Learning (Stanford)
Stanford
8 comments from Hacker News
No, and they should be criticized for that. However, this is the transparency that makes up for it: [VIDEO]
As stated by Tesla just a few days ago, the occupancy network is not currently shown in the customer-facing visualisation. (Some unrecognised objects are represented by generic placeholders. This is derived from, but not the output of, the occupancy network.) This video links to the relevant citation, and if you continue watching for around ten seconds you can see examples of the occupancy network visualised: [VIDEO]
The Tesla AI day videos [1] go into some detail about this. They use multiple networks that are dedicated to specific tasks. [1] [LINK] (2021), [VIDEO] (2022)
It is pretty cool: [VIDEO] They transitioned from 2D to 3D a couple years ago, major transition but it does seem like a critical step. We live in a 3d rather than 2d world.
The replies to your comment don't seem to understand you at all. in the video link here [VIDEO] he clearly states 16x dynamic range as a result of direct photon processing.
Exactly. There's no better facsimilie for the sheer variety of public roads than, well, public roads. Tesla even has rigorous simulation in place for hostile situations that should never occur on public roads: A good general overview from two minute papers: [LINK] Tesla's AI Day 2022 video, timestamped around labeling and simulation [VIDEO]
> If the data from supervised driving can he used for unsupervised driving So you're unaware of the AI day presentations about their driving training pipeline. Learning about their engineering in that regard really changed my perception of the likelihood that they could solve the problem. Give it a watch: [VIDEO] > failure to keep up with competitors like Waymo Looks like each of them lead in different ways to me. Waymo leads within certain specific city limits, Tesla leads on the open road. These seem to be direct consequences of their engineering decisions. Question is which will converge on working everywhere all the time.
>I've never really thought of Waymo as a robot in the same way as e.g. a Boston Dynamics humanoid, but of course it is a robot of sorts. So for the record, with this realization you're 3+ years behind Tesla. [VIDEO]