Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Elon: Maybe update first week of March

This site may earn commission on affiliate links.

DanCar

Active Member
Oct 2, 2013
2,909
3,732
SF Bay Area
https://twitter.com/elonmusk/status/1364480679019429888
Quote:
We’re upgrading all NNs to surround video, using subnets on focal areas (vs equal compute on all uncropped pixels) & many other things, so more time needed to write & validate software. Maybe something next week. This is evolving into solving a big part of physical world AI.

Replies from twitter:
TesLatino:
Should we expect next-level improvements on the next release, or will we experience expected steps backs to validate the next deployment? Also, is there a lot of QA done in-house before we (testers) get it?
Elon: 9h
Many steps forward, some steps back. Yes, we do a lot of QA before releasing a beta version.

Eugene Hsue: Will Tesla’s NN AI be able to solve other problems other than FSD?
Elon Musk: 9h : Seems likely

green: 2h:
just like that Andrej K presentation from last year: there's an NN per camera (more than one really, but importantly they are per camera (I guess "focal area" sounds fancier?)) the outputs of which are then fed into a reconciliation NN tying it all together. But more importantly it's likely not coming next week in a usable car to a regular car not in any special programs ;)
 
220px-Fingers_and_thumb_in_circle_downward_motion.jpg
 
https://twitter.com/elonmusk/status/1364480679019429888
Quote:
We’re upgrading all NNs to surround video, using subnets on focal areas (vs equal compute on all uncropped pixels) & many other things, so more time needed to write & validate software. Maybe something next week. This is evolving into solving a big part of physical world AI.

This tweet does not sound like L5 is going to happen this year as Elon previously stated. From this tweet, it sounds like Tesla is in the middle of implementing the 4D rewrite and needs to do lots of other things too. Elon says that there is software that needs to be written and validated. It sounds like they still have a lot to do. It will take time. And there will probably be set-backs which is normal in software development. "Maybe something next week" could turn out to be a month or several months.

So I am happy with the update. It's great to get details on what Tesla is doing. But I think we need to forget any predictions about when Tesla will achieve L5. Just let the team do their work.
 
Awesome L3 and L4 will happen this decade, but extremely unlikely L5. Too many science fiction believers. When general A.I. has been invented, then we will get L5. Elon has been wrong for the past 6 years on this topic: 2015 - 2020. He will continue to be wrong for rest of the decade. Unless you classify L5 as manufacturer intent, which is the SAE definition. Which means functionality doesn't matter, just the intention.
 
Last edited:
Unless you classify L5 as manufacturer intent, which is the SAE definition. Which means functionality doesn't matter, just the intention.

The amount of consternation and confusion that has been caused by people conflating developer intent (i.e. SAE Levels of Driving Automation classification) with actual technical capability (i.e. inchoate idea we still have to devise good tests and metrics for)...
 
  • Like
Reactions: DanCar
This tweet does not sound like L5 is going to happen this year as Elon previously stated. From this tweet, it sounds like Tesla is in the middle of implementing the 4D rewrite and needs to do lots of other things too. Elon says that there is software that needs to be written and validated. It sounds like they still have a lot to do. It will take time. And there will probably be set-backs which is normal in software development. "Maybe something next week" could turn out to be a month or several months.

I just got back from 2031 and this exact same thing was posted on TMC then!
 
Unless you classify L5 as manufacturer intent, which is the SAE definition. Which means functionality doesn't matter, just the intention.

Wrong. Functionality matters. The SAE levels do include functionality. Specifically, L5 must do all DDT and DDT-fallback with no ODD restrictions. So, automakers do need to design their L5 to achieve that functionality. They can't just declare any system to be L5 because they want to. For example, they can't just declare cruise control to be L5. If they have a system that is designed to do all DDT and all DDT-fallback with no ODD restrictions then they can declare the intent to be L5.
 
I wonder if this means they'll use all 8 cameras for all perception tasks instead of Karpathy's presentation seems to suggest 5 cameras (all but main, far, rear) and 3 tasks (missing static objects, road signs, traffic lights, road markings, crosswalks, overhead signs, environment tags, etc.):
more bev.jpg


Perhaps Tesla thought the initial BEV Net with just moving objects, road lines and road edges would be good enough; and indeed FSD beta is already quite capable but turns out not good enough for wider release.

using subnets on focal areas (vs equal compute on all uncropped pixels)
This seems to suggest some hardware limitations on either/both the training and inference side. Autopilot team needed to do clever cropping to fit within HW2/.5 compute limitations, so doing that again at this time is somewhat curious. Karpathy did answer a question at CVPR saying "Also we do worry about the architectures quite a bit because you're right - a lot of information is only at the vanishing line, and 1/3 of the image is clouds and 1/3 is road." So my guess is they're going to run the full camera NNs at a lower frequency than NNs handling full frame-rate cropped input "focal areas" and both outputs are fed into the Temporal module for the BEV Net to make a unified prediction.
 
Eugene Hsue: Will Tesla’s NN AI be able to solve other problems other than FSD?
Elon Musk: 9h : Seems likely

My hope is it will improve side monitoring so it can better track the locations of vehicles around my vehicle.

I also don't believe that FSD owners will have to wait for the FSD beta to see benefits from it. I'm sure they'll get it to the point where the FSD beta stuff is simply through an enable register. As the FSD beta general release is likely quite a ways off.
 
Did he say L5 by the end of this year or is that your interpretation.

From what I remember Musk doesn’t care about L3/L5 etc. When asked a question about levels he assumes some stuff and answers.

He actually said last year. It was at the World Artificial Intelligence Conference back in July 2020. Here is the exact quote that I am referring to:

“I’m extremely confident that level 5 or essentially complete autonomy will happen and I think will happen very quickly,” Musk said in remarks made via a video message at the opening of Shanghai’s annual World Artificial Intelligence Conference (WAIC).

I remain confident that we will have the basic functionality for level 5 autonomy complete this year.

Tesla 'very close' to level 5 autonomous driving technology, Musk says

So, I guess I stand corrected. He did not actually say L5 would happen this year. He said "basic functionality for L5" would happen last year and L5 would happen "very quickly". But it is not unreasonable IMO to interpret that statement as meaning L5 would happen this year.

I guess "basic functionality" does give Elon a nice out since he can probably argue that "basic functionality" just means the building blocks, not actual L5. Elon could probably argue that FSD Beta is "basic functionality for L5" while not actually being L5.
 
Last edited:
using subnets on focal areas (vs equal compute on all uncropped pixels)
This seems to suggest some hardware limitations on either/both the training and inference side. Autopilot team needed to do clever cropping to fit within HW2/.5 compute limitations, so doing that again at this time is somewhat curious. Karpathy did answer a question at CVPR saying "Also we do worry about the architectures quite a bit because you're right - a lot of information is only at the vanishing line, and 1/3 of the image is clouds and 1/3 is road." So my guess is they're going to run the full camera NNs at a lower frequency than NNs handling full frame-rate cropped input "focal areas" and both outputs are fed into the Temporal module for the BEV Net to make a unified prediction.
That's exactly what I got from that part too (that he's talking about doing NNs on cropped portions instead of on all regions). However, it may not necessarily have to do solely (or even at all) with compute limitations. I imagine an NN that looks only at a cropped portion of the image would be quite different than one that looks at the whole camera view. For example, one that only looked for road lines on a crop of the ground would likely be quite different than one that looked for them on the whole image (including the sky).