FSD Beta v11.x

Supcom · Feb 4, 2023

powertoold said:
There are a lot of major challenges IMO.

A big part of V11 is replacing the USS (from the notes about 4x spatial transformer resolution), and they haven't released that yet (already 4 months since they removed USS in 10/22). Until they can achieve USS replacement, I think there's little hope in whatever they're doing with V11.

While the USS replacement software is a requirement for autopark and a replacement smart summon capability, it is not needed for driving on streets.

Obviously, the USS replacement is needed for non-FSD cars as well, so it should have some priority. But, it may be a different team of programmers from those working on V11 issues.

powertoold · Feb 4, 2023

Supcom said:
While the USS replacement software is a requirement for autopark and a replacement smart summon capability, it is not needed for driving on streets.

Obviously, the USS replacement is needed for non-FSD cars as well, so it should have some priority. But, it may be a different team of programmers from those working on V11 issues.

That's true, but if you read all of V11's release notes, the idea is that Tesla is trying to achieve a generalized object detection NN for both low (parking / summon) and high speed (highway obstacle avoidance).

Basically, they're not looking for a NN for obstacles at low speed and a different NN for obstacles at high speed.

It doesn't seem like they're going to release a replacement for USS until they've achieved this.

Scott Chapman · Feb 5, 2023

Reset that clock!

Mardak · Feb 5, 2023

Supcom said:
the USS replacement is needed for non-FSD cars as well, so it should have some priority. But, it may be a different team of programmers from those working on V11 issues.

That actually might make it higher priority than FSD Beta highway driving as it affects the usability of new/recent vehicles vs. the incremental benefit over existing highway Autopilot (which is a major technical milestone but maybe not as immediately user noticeable). The October sensor removal was most likely based on the Occupancy network progress presented at 2023 AI Day in September, so that code is probably very much coupled with FSD Beta and not worth trying to separate especially if it seemed like single stack was "two weeks" away.

If FSD Beta 11 single stack was delayed by highway-specific optimizations and improvements, it's a bit unfortunate that people had to deal with the lack of park assist and other features for even longer.

However, if this is true, then it sounds like single stack with Occupancy network will be going to at least all new vehicles in some form to restore features on non-FSD vehicles. This would probably allow for even wider fleet shadow mode data collection for FSD Beta.

sleepydoc · Feb 5, 2023

Has anyone else noticed how many times Tesla has had to ‘rearchitecture‘ the NNs or otherwise change their algorithm paradigm? It seems like they end up practically starting from scratch about every 9-12 months. It’s no wonder we’re not any further but it also doesn’t bode well.

My other fear is that we’re all going to be subjected to a major regression in the functionality and reliability of Autopilot. Since I bought my car it’s been nearly flawless and is one of the best features.

Zilla91 · Feb 5, 2023

Scott Chapman said:
View attachment 903549
Reset that clock!

So it should be shipping today?

powertoold · Feb 5, 2023

sleepydoc said:
Has anyone else noticed how many times Tesla has had to ‘rearchitecture‘ the NNs or otherwise change their algorithm paradigm? It seems like they end up practically starting from scratch about every 9-12 months. It’s no wonder we’re not any further but it also doesn’t bode well.

My other fear is that we’re all going to be subjected to a major regression in the functionality and reliability of Autopilot. Since I bought my car it’s been nearly flawless and is one of the best features.

This is actually an advantage of Tesla's approach. They are very nimble in applying the latest NN techniques, and they have a huge testing infrastructure to reduce regressions.

At the same time, you're right. It still seems like they're trying to find the right architecture.

I've always wondered whether Tesla's lack of progress lately is an auto-labeling/training problem or an architecture problem. More and more, it seems like an architecture problem (or perhaps a HW3 limitation).

KLam · Feb 5, 2023

sleepydoc said:
Has anyone else noticed how many times Tesla has had to ‘rearchitecture‘ the NNs or otherwise change their algorithm paradigm? It seems like they end up practically starting from scratch about every 9-12 months. It’s no wonder we’re not any further but it also doesn’t bode well.

My other fear is that we’re all going to be subjected to a major regression in the functionality and reliability of Autopilot. Since I bought my car it’s been nearly flawless and is one of the best features.

What do you expect if they keep operating on the brain and feeding it with new drugs!

dckiwi · Feb 5, 2023

powertoold said:
This is actually an advantage of Tesla's approach. They are very nimble in applying the latest NN techniques, and they have a huge testing infrastructure to reduce regressions.

At the same time, you're right. It still seems like they're trying to find the right architecture.

I've always wondered whether Tesla's lack of progress lately is an auto-labeling/training problem or an architecture problem. More and more, it seems like an architecture problem (or perhaps a HW3 limitation).

In the most recent Lex Fridman podcast, Tim Dodd talks about Elon's approach to engineering and innovation. He is not afraid to throw out approaches that aren't working, build something new, and 'fail fast' again. Obviously you want to see incremental progress over time (and we could debate whether FSD Beta has progressed sufficiently over the past few years), but Elon's 'fast fail' engineering approach is often the best way to tackle problems that no one has ever solved before.

sleepydoc · Feb 5, 2023

dckiwi said:
In the most recent Lex Fridman podcast, Tim Dodd talks about Elon's approach to engineering and innovation. He is not afraid to throw out approaches that aren't working, build something new, and 'fail fast' again. Obviously you want to see incremental progress over time (and we could debate whether FSD Beta has progressed sufficiently over the past few years), but Elon's 'fast fail' engineering approach is often the best way to tackle problems that no one has ever solved before.

The problem is one can’t program an entire FSD algorithm quickly, so the ‘fail often and fast’ approach doesn’t work so well.

powertoold · Feb 5, 2023

sleepydoc said:
The problem is one can’t program an entire FSD algorithm quickly, so the ‘fail often and fast’ approach doesn’t work so well.

With NNs taking over C++ code, it's becoming less of a programming problem and more of a training / NN architecture and unit-testing problem. Tesla has been taking the steps towards this for years now (increasing cluster size, developing robust testing infrastructure, using SOTA NNs approaches: transformers, NERFs, autogressive models, etc.).

EVNow · Feb 5, 2023

JulienW said:
After a FULL year of no show I'm starting to believe V11/highway Stack is still in Alpha. So AP would be an Alpha Beta......or is it a Beta Alpha??????

From my Dec-21 post ... I've always been somewhat skeptical of single stack because of regression possibilities ...

EVNow said:
Hope we get 11 by this time next year.

Goose66 · Feb 5, 2023

Elon can say the next version is coming in X weeks all he wants - for months and months in some cases - and people still post his tweets as if they have anything to do with the real release date of the next version. He just keeps saying it until that next version is released, and then starts saying the same thing about the next, next version. It's all a bunch of handwaving to keep the stock price up. It has nothing to do with where the product is in development, how testing is going, when a broader release will occur, etc. At least I don't understand how a rational person could conclude his comments have anything to do with actual release dates given his track record over the last 7 or 8 years in this regard.

JulienW · Feb 5, 2023

Goose66 said:
Elon can say the next version is coming in X weeks all he wants - for months and months in some cases - and people still post his tweets as if they have anything to do with the real release date of the next version...

yosiel · Feb 6, 2023

Hi. When do I get scored for fsd Beta

Thp3 · Feb 6, 2023

Scoring is over. You just agree to the Beta requirement and request FSD.

yosiel · Feb 6, 2023

It says Tesla is reviewing my driving

Scott Chapman · Feb 6, 2023

yosiel said:
It says Tesla is reviewing my driving

We all are.

RenMan68 · Feb 7, 2023

I’m confused about the version numbering convention and what version I actually have. My app says I have 2022.44.30.10. In the car, when I installed this version, it said ”Software Version 11.0”. So I thought I have version 11 but it also looks like this is 10.69.25.2, which seems to me to be a “version 10” variant. Can anybody help me understand this mess?

JulienW · Feb 7, 2023

RenMan68 said:
I’m confused about the version numbering convention and what version I actually have. My app says I have 2022.44.30.10. In the car, when I installed this version, it said ”Software Version 11.0”. So I thought I have version 11 but it also looks like this is 10.69.25.2, which seems to me to be a “version 10” variant. Can anybody help me understand this mess?

All Tesla's have UI Version 11. Just coincidental numbering. There are 3 softwares:

In your case:

UI (11)
Car's firmware (22.44.30.10)
FSD Beta (10.69.25.2)

FSD Beta v11.x

Active Member

Active Member

Member

Active Member

Well-Known Member

Active Member

Active Member

Member

Member

Well-Known Member

Active Member

Well-Known Member

Member

Well-Known Member

Member

Member

Member

Member

Member

Well-Known Member

Similar threads