6 ms·

"Our Autopilot team has been focused on a fundamental architectural rewrite of our neural networks and control algorithms." Does anyone else find this worryin

"Our Autopilot team has been focused on a fundamental architectural rewrite of our neural networks and control algorithms." Does anyone else find this worrying?

mushufasa 6y ago
no. it's not even surprising. George Hotz at CommaAI seems to be doing the same thing or similar. Apparently some papers came out late last year that have laid the foundation for much better approaches.
- tjchear 6y ago
  Any pointers for said papers?
- person_of_color 6y ago
  Got a link?
  - leesec 6y ago
    (1) Skip to 1:03:00. George Hotz comments on the paper (2) Model-Predictive Policy Learning with Uncertainty Regularization for Driving in Dense Traffic: 1. https://www.youtube.com/watch?v=yhzU43ewcEA https://www.youtube.com/watch?v=yhzU43ewcEA 2. https://arxiv.org/abs/1901.02705 https://arxiv.org/abs/1901.02705
jconley 6y ago
It's probably a good thing as long as they can reuse the training/annotation data. Startups often do full rewrites of underlying technology multiple times as they scale. With the pace of advancement of AI/ML right now I would assume some better algorithms have come along in recent years...
mft_ 6y ago
It’s worrying... that... they’re working very hard to make it better? (I honestly don’t get your point?)
nickik 6y ago
What are you worried about?
SEJeff 6y ago
The current system is a separate neural network running against every frame of every one of the 8 cameras. The new one is what elon calls the "4D" image processing. It uses all cameras to create a composite 3d representation where the 4th dimension is time. One of the things they use this for is to predict where an object will be in the future.
zhaphod 6y ago
I have an eithical dilemma for you. As of now on an average 33000 ppl are killed in road accidents in USA. Lets say Tesla's or other ICE makers' autopilot reduces this to 20000 deaths. But of those 20000 deaths, say 2000 were purely because of autopilot alogrithm which wouldn't have happened if there was no autopilot. As a regulator what will you do. Approve autopilot because it saved 13000 people or ban it because 2000 were killed just because of it? I know how this will play in media. Unfortunately autopilot will be demonized and most likely will get banned.
- ksec 6y ago
  I have had the same question as well. I dont think it will get banned, I just think the society as a whole will take a long time to adjust those perception. And we will likely have zones where all cars are AV. So as long as human sticks to rules and not popping out on a street for no reason. I still think we could have a future where AV becomes the default.

"Our Autopilot team has been focused on a fundamental architectural rewrite of our neural networks and control algorithms." Does anyone else find this worrying?

mushufasa 6y ago

no. it's not even surprising. George Hotz at CommaAI seems to be doing the same thing or similar. Apparently some papers came out late last year that have laid the foundation for much better approaches.

tjchear 6y ago
Any pointers for said papers?
person_of_color 6y ago
Got a link?
- leesec 6y ago
  (1) Skip to 1:03:00. George Hotz comments on the paper (2) Model-Predictive Policy Learning with Uncertainty Regularization for Driving in Dense Traffic: 1. https://www.youtube.com/watch?v=yhzU43ewcEA https://www.youtube.com/watch?v=yhzU43ewcEA 2. https://arxiv.org/abs/1901.02705 https://arxiv.org/abs/1901.02705

jconley 6y ago

It's probably a good thing as long as they can reuse the training/annotation data. Startups often do full rewrites of underlying technology multiple times as they scale. With the pace of advancement of AI/ML right now I would assume some better algorithms have come along in recent years...

mft_ 6y ago

It’s worrying... that... they’re working very hard to make it better? (I honestly don’t get your point?)

nickik 6y ago

What are you worried about?

SEJeff 6y ago

The current system is a separate neural network running against every frame of every one of the 8 cameras. The new one is what elon calls the "4D" image processing. It uses all cameras to create a composite 3d representation where the 4th dimension is time. One of the things they use this for is to predict where an object will be in the future.

zhaphod 6y ago

I have an eithical dilemma for you. As of now on an average 33000 ppl are killed in road accidents in USA. Lets say Tesla's or other ICE makers' autopilot reduces this to 20000 deaths. But of those 20000 deaths, say 2000 were purely because of autopilot alogrithm which wouldn't have happened if there was no autopilot. As a regulator what will you do. Approve autopilot because it saved 13000 people or ban it because 2000 were killed just because of it? I know how this will play in media. Unfortunately autopilot will be demonized and most likely will get banned.

ksec 6y ago
I have had the same question as well. I dont think it will get banned, I just think the society as a whole will take a long time to adjust those perception. And we will likely have zones where all cars are AV. So as long as human sticks to rules and not popping out on a street for no reason. I still think we could have a future where AV becomes the default.