George Hotz and Lex Fridman talks about Autopilot

[EvanLin]

It's too deep for me to sort out the key points. I'll leave that to you guys.

Tesla Autopilot big rewrite to 4D perception and annotation

If I ran Tesla, I wouldn't change much

Comma.ai mission is to solve self-driving cars

Tesla Autopilot and openpilot communicating uncertainty

Tesla Dojo, TPU, NVIDIA & hardware optimized for machine learning

Tesla vs comma.ai approach to machine learning
https://www.youtube.com/watch?v=pGF2xK_Hntw

 

[masterxel]

Thanks for posting this, very interesting and relevant!

On Tesla Dojo
George says it's too expensive to train neural networks, your only options are Google or NVIDIA and Google doesn't allow training autonomous driving. He doesn't like the monopoly and wishes Tesla would sell their own chips to provide competition. He thinks anyone could make the same chips though and someone else will do that soon.

On Autopilot rewrite
Comma.ai are also in the middle of a 2d to 3d rewrite, adding Z and unifying things, switching from tensorflow to py torch. When it comes to data training he mentions this karpathy tweet as a very useful thing to do, but he still thinks human annotation is key. https://twitter.com/karpathy/status/1311884485676294151 "When you sort your dataset descending by loss you are guaranteed to find something unexpected, strange and helpful."

On Waymo
George doesn't think they have a worthwhile product. He thinks the product that they're building doesn't make sense because they are slower and more cautious than an Uber so it can't be competitive to consumers. Compares Waymo to scooters as a niche product.