I discuss with James Douma about Tesla's approach to depth mapping and why Tesla's approach is likely a better overall solution compared to Lidar.

James Douma on Twitter: https://twitter.com/jamesdouma​

James Douma Playlist, https://www.youtube.com/watch?v=iMtujONU_0I&list=PLfibpgBinf9R7KIedEU3y-YjrA63LSKHX

Andrej Karpathy talk at ScaledML 2020: https://www.youtube.com/watch?v=hx7BXih7zx8

Timestamps

00:00 - Introduction

1:30 - Elon saying next FSD update will be a step-up change

4:04 - Timeline of FSD

9:00 - How much is Tesla's ML applicable to other fields?

10:55 - Pseudo-lidar

17:00 - How do NNs determine depth?

22:58 - Unsupervised vs self-supervised

25:00 - Overlapping camera views and time

26:35 - Challenging NNs to learn depth improving accuracy

30:20 - As NNs recognize depth, will birds eye view get more accurate?

31:55 - Strength and weakness of liar vs vision based approach

35:38 - Benefits of lidar?

37:43 - Can lidar read street signs?

42:09 - If dependent on lidar, dependent on HD maps?

43:18 - When will folks using lidar realize Tesla's approach is better?

46:18 - Matterport camera

47:55 - Lidar in the iPhone 12

49:59 - NNs are more significant to world than Wright brothers' flight

51:50 - Converging tech enabling NNs

52:45 - Moore's law and machine learning

55:20 - Conclusion

Social