Copying my post from investor thread:
Thanks! Super interested in these videos, watched all of AKs videos! =)
Guess the new information was the EKF-SLAM style occupancy grid which I will try to analyze more, the sonar and camera road edge visualization was cute, but felt a bit dated compared with the rest. The neural network deconvoluted map was really cool, felt very modern and interesting! But my main takeaway is that they use the multitask CNN to do a lot of object detection etc and use the high level features of this neural network as input to train and to run a RNN. Which feels a bit similar to OpenAI-five:
Note the LTSM 1024 units to the left, which would be the equivalent of this RNN and the input to this network would be high level features of the AKNET_V9 style multitask network.
This was new information. Pretty cool, but not super happy that Tesla are telling the rest of the industry the correct answer. Imo it seems like a really good solution.
Also liked the go on holiday and come back to improved performance quote. Imo that is something awesome with Tesla’s software 2.0 solution, once it is up and running it can only be improved and there are so many ways to work to improve it and the effects will ripple out to all parts of their software 2.0 stack.
Project Dojo seems to be a custom system for training Tesla’s neural networks. I wonder how much own hardware they have developed for this, I assume that Elon Musk gave Pete Bannon two task:
1. Develop HW3
2. Develop hardware for Project Dojo
We know how well he did 1 from the Autonomy day. I would guess he did 2 equally well.
I wonder how much these cost, would not be surprised if Elon gave a green light to split the budget 50/50 for these and that Pete Bannon has built some pretty massive amount of very specific hardware to train Tesla’s neural network. Don’t think that these will compete with Alphabet’s TPU 3.0, but maybe they will perform better for Tesla’s specific needs. If so Tesla’s moat might just have gotten wider. Maybe they could compete, if so maybe Amazon/Microsoft/Google could be customers/competition. Would be pretty crazy if Tesla decides to enter the cloud compute market at some point. Hyperchange should make a video about this! =)
To be followed!
Thanks! Super interested in these videos, watched all of AKs videos! =)
Guess the new information was the EKF-SLAM style occupancy grid which I will try to analyze more, the sonar and camera road edge visualization was cute, but felt a bit dated compared with the rest. The neural network deconvoluted map was really cool, felt very modern and interesting! But my main takeaway is that they use the multitask CNN to do a lot of object detection etc and use the high level features of this neural network as input to train and to run a RNN. Which feels a bit similar to OpenAI-five:
Note the LTSM 1024 units to the left, which would be the equivalent of this RNN and the input to this network would be high level features of the AKNET_V9 style multitask network.
This was new information. Pretty cool, but not super happy that Tesla are telling the rest of the industry the correct answer. Imo it seems like a really good solution.
Also liked the go on holiday and come back to improved performance quote. Imo that is something awesome with Tesla’s software 2.0 solution, once it is up and running it can only be improved and there are so many ways to work to improve it and the effects will ripple out to all parts of their software 2.0 stack.
Project Dojo seems to be a custom system for training Tesla’s neural networks. I wonder how much own hardware they have developed for this, I assume that Elon Musk gave Pete Bannon two task:
1. Develop HW3
2. Develop hardware for Project Dojo
We know how well he did 1 from the Autonomy day. I would guess he did 2 equally well.
I wonder how much these cost, would not be surprised if Elon gave a green light to split the budget 50/50 for these and that Pete Bannon has built some pretty massive amount of very specific hardware to train Tesla’s neural network. Don’t think that these will compete with Alphabet’s TPU 3.0, but maybe they will perform better for Tesla’s specific needs. If so Tesla’s moat might just have gotten wider. Maybe they could compete, if so maybe Amazon/Microsoft/Google could be customers/competition. Would be pretty crazy if Tesla decides to enter the cloud compute market at some point. Hyperchange should make a video about this! =)
To be followed!