FSD Beta v11.x

The next big milestone for FSD is 11. It is a significant upgrade and fundamental changes to several parts of the FSD stack including totally new way to train the perception NN.

From AI day and Lex Fridman interview we have a good sense of what might be included.

- Object permanence both temporal and spatial
- Moving from “bag of points” to objects in NN
- Creating a 3D vector representation of the environment all in NN
- Planner optimization using NN / Monte Carlo Tree Search (MCTS)
- Change from processed images to “photon count” / raw image
- Change from single image perception to surround video
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

Lex Fridman Interview of Elon. Starting with FSD related topics.

Here is a detailed explanation of Beta 11 in "layman's language" by James Douma, interview done after Lex Podcast.

Here is the AI Day explanation by in 4 parts.

Deep Understanding Tesla FSD Part 1: HydraNet

From Theory to Reality, Analyze the Evolution of Tesla Full Self-Driving

saneryee-studio.medium.com

Here is a useful blog post asking a few questions to Tesla about AI day. The useful part comes in comparison of Tesla's methods with Waymo and others (detailed papers linked).

Twelve Questions about Tesla Autopilot Technologies Exposed at 2021 AI Day

On Aug. 19 2021, Telsa held an AI day , exposing more technologies about Autopilot , from perception, planning, data labeling, simulation…

yuhuang-63908.medium.com

 

[Mardak]

Tesla has been wanting to merge FSD Beta into the main production vehicle software, so potentially 2023.12.10 with FSD Beta 11.3.6 is the first release doing so, and it makes some sense to use an "old" FSD Beta version to reduce risk of trying to incorporate 11.4.x changes that are still changing. Similar to how say the media player functionality doesn't need to change on every vehicle update, FSD Beta functionality doesn't need to change either when going to say 2023.16.x.

Going forwards, Tesla can still maintain "early access" releases to test out newer versions of FSD Beta probably to the pre-Safety-Score group that the rest of us wait for it to get to a newer production version. In some sense, basically all of us will be in very similar situation of those waiting to get FSD Beta for the first time where "too new" software prevents updating to a compatible FSD Beta (e.g., until yesterday, those on 2023.12.5 could not update to 2022.45.15 / FSD Beta 11.3.6 nor 2023.7.5 / FSD Beta 11.4.1) except there will be some baseline version of FSD Beta available starting with 11.3.6.

To be clear, Tesla can still segment who receives updates by hardware, software, region, etc. such as MCU1 vehicles vs with FSD Capability vs in North America. So another next step for FSD Beta is being available on all vehicles even without FSD Capability so at minimum, people newly subscribing or purchasing the software can start using it right away without waiting for a software update. Another next step is to make use of FSD Beta neural networks even without FSD Capability such as replacing legacy stack for basic Autopilot.

 

[YNoVA]

wow and you've never been a beta tester before? mine is greyed out

Mine was also greyed out. I did try FSD Beta back in December but didn't keep the subscription.

 

[swedge]

Big oops there. Updating to 2023.12.x means you have no chance to get 11.4.x, because 11.4.x is based on 2023.7. They don't go backwards in time. 2023.7 means it is based on the code as of the 7th week of 2023, while 2023.12.x is based on th 12th week of 2023. Elon's time machine only goes forward in the, not backwards.

Speaking of time, when Elon says something like 11.4.2 is going to be rolled out "this weekend", what he really means is that it will not be rolled out before this weekend. He lives in a time bubble which only looks to the future, a time diode, so to speak. Like a >, a "greater than" symbol. He sounds like he means "this weekend", but he really means "not before" this weekend. He has never meant "Equal To", always "More Than", actual semantics not withstanding. History shows us that he speaks of the future, not the present. Always.

;-)

The lengths Elon will go just to prove me wrong!

Just after I started writing the above, Tesla started rolling out 2023.12.10 which contains FSD beta 11.3.6 . This probably means that what I said above was untrue by the time I posted it. Sorry.

As of this morning, 5/23, there are only 77 pending plus installed copies, so it may still be a few days before everyone gets it. I hope they keep putting FSD beta on current releases going forward. On the other hand, today is in week 21 of 2023, so even 2023.12.x is many weeks behind reality...

I also hope we get 11.4.2 soon, like Elon ... uh...suggested.... we would.

I expect this has already been discussed, but I wanted to apologize ASAP.

 

[swedge]

The next 11.4 will likely be 12.X, they obviously are having issues as this 2023.12.10 went out today to get all the new cars added.

Just because 11.4.1 was 2023.7.5 doesn't mean the next version will be in the .7, plus 11.4.1 was a minor upgrade from 11.3.6. I think new testers would rather have something than nothing.

You've proven correct already! Well done!

 

[SSDTester#1]

Someone said stacks…. Made me think of this

 

[JulienW]

.....I also hope we get 11.4.2 soon, like Elon ... uh...suggested.... we would......

Elon said 11.4.2 would (not just suggested) start rollout LAST weekend. Clearly that didn't happen so "soon" is defiantly not going to the masses this week or even next next. Clearly 11.4.x has much more serious or at least harder to correct problems that Elon led us to believe. 23.12.10 has to be a stop gap rush out since 11.4.x is not near ready.

 

[aronth5]

You are getting confused by FSD 12.X and 2023.12.X

No, I don't think so but why do you say that? My post was only related to FSD 12.x and 11.4.x.

 

[uscbucsfan]

No, I don't think so but why do you say that? My post was only related to FSD 12.x and 11.4.x.

Because no one else is talking about FSD 12.X, yet twice you've brought it up.

 

[aronth5]

Elon said 11.4.2 would (not just suggested) start rollout LAST weekend. Clearly that didn't happen so "soon" is defiantly not going to the masses this week or even next next. Clearly 11.4.x has much more serious or at least harder to correct problems that Elon led us to believe. 23.12.10 has to be a stop gap rush out since 11.4.x is not near ready.

Or another guess. Maybe Tesla decided it was more important to drive FSD revenue?

 

[aronth5]

Why do you think this is likely?

11.4 was 2023.6.15, 11.4.1 is 2023.7.5. I think it's more likely we see 11.4.2/3 at 2023.12.X.

This isn't the first time revisions have jumped up on base software. 10.69 did the same.

Because no one else is talking about FSD 12.X, yet twice you've brought it up.

I only brought it up when you referenced FSD V12 in your earlier post. See above. Did I misunderstand your post?

 

[uscbucsfan]

I only brought it up when you referenced FSD V12 in your earlier post. See above. Did I misunderstand your post?

You did, it states 2023.12.X which is software version, not the FSD version. 2023.12.10 is the newest Beta push going out with 11.3.6.

There was no mention of V12 FSD.

 

[JulienW]

Or another guess. Maybe Tesla decided it was more important to drive FSD revenue?

That is NOT another guess but a part of it. There is fiduciary pressure to deliver FSD Beta after the recall. But Tesla expected 11.4.x to be that vehicle and obviously not ready.

 

[D Good]

My taste - We won’t see FSD 12.x until 2025…

 

[uscbucsfan]

My taste - We won’t see FSD 12.x until 2025…

I think Holiday 2024, but close enough.

 

[aronth5]

You did, it states 2023.12.X which is software version, not the FSD version. 2023.12.10 is the newest Beta push going out with 11.3.6.

There was no mention of V12 FSD.

Thanks for clarifying. I just misunderstood how you used the build number.

 

[Mardak]

Looks like Tesla has more than doubled the rollout of 2023.12.10 / FSD Beta 11.3.6 still going to new additions instead of those already on 2022.45.15 / FSD Beta 11.3.6.

A few more previous versions detected with the wider rollout including some factory versions: 2023.2.10, 2023.2.12, 2023.2.201, 2023.6.8, 2023.6.9, 2023.6.11, 2023.6.100, 2023.12.1, 2023.12.1.1, 2023.12.5, 2023.12.9

2023.12.10 Official Tesla Software Update - Teslascope

Release notes and statistics for 2023.12.10 with Full Self-Driving (Beta) v11.3.6 and more.

teslascope.com

TeslaFi.com Firmware Tracker

www.teslafi.com

Seems like it might indicate it's a good time to add FSD Capability subscription -- except for those with HW4 probably waiting for 11.4.x.

 

[jebinc]

My taste - We won’t see FSD 12.x until 2025…

At the earliest.

 

[outdoors]

My taste - We won’t see FSD 12.x until 2025…

Which makes me laugh when I think about Tesla wanting me to trade in my S with FSDb and Unlimited Free Supercharging. Pay again for the privilege I guess.

 

[Tronguy]

Looks like Tesla has more than doubled the rollout of 2023.12.10 / FSD Beta 11.3.6 still going to new additions instead of those already on 2022.45.15 / FSD Beta 11.3.6.

A few more previous versions detected with the wider rollout including some factory versions: 2023.2.10, 2023.2.12, 2023.2.201, 2023.6.8, 2023.6.9, 2023.6.11, 2023.6.100, 2023.12.1, 2023.12.1.1, 2023.12.5, 2023.12.9

2023.12.10 Official Tesla Software Update - Teslascope

Release notes and statistics for 2023.12.10 with Full Self-Driving (Beta) v11.3.6 and more.

teslascope.com

TeslaFi.com Firmware Tracker

www.teslafi.com

Seems like it might indicate it's a good time to add FSD Capability subscription -- except for those with HW4 probably waiting for 11.4.x.

Suspect that once all or most of the newbies are in, they'll probably let us old farts in so we can get whatever additional bells and whistles are in the 2023.12.X release.

My suspicion is that whenever some SW types are gluing large hunks of software together, Bugs Will Be Found.

Ha. Back in the day when in my Freshman year in Engineering, we were all forced into our very first programming class. In FORTRAN. (Should note: Unix and C were very, very new at the time and not seeing widespread use.) The prof made a point of telling all us newbies the below:

General rule: Write a 300 line program, not including whitespace and comments. 'Way more than half of the lines will have errors in them. Typos, everything else.

Before compiling, run Lint or the equivalent. This knocks the error rate to 25% of the lines. A lot fewer typos, but there'll be plenty of other errors.

Try compiling. The number of errors after fixing all the ABENDS and such in the compiler drops the error rate to 10%.

Start testing the software, varying the inputs and such. If one is lucky, error rate falls to 1%. And, for a very, very long time, that's where things stayed stuck. There might be errors in the code, but the SW will work well enough.

Now, imagine a code base of a million lines. Or two million lines. At 1% error rate, we're talking a 100,000-200,000 bunch of errors in the code.

If you were wondering why Microsoft/Apple/Adobe/etc have monthly updates, well, now you know .

Modern techniques, like running fuzzers, or using languages that put a straight-jacket on one's coding practices, can perhaps knock the error rate down by another factor of ten, but there'll still be a huge number of errors out there.

The prof stated that, once in a while, some miracle worker or other will write a 300 line routine that has no errors, on the first try. People like that are out there to irritate the rest of us and proving that the exception makes the rule. Which only proves that really, really good programmers are worth their weight in gold and become internationally famous, like Moxie Marlinspike and Bruce Schneier.

 

[vamike]

Suspect that once all or most of the newbies are in, they'll probably let us old farts in so we can get whatever additional bells and whistles are in the 2023.12.X release.

My suspicion is that whenever some SW types are gluing large hunks of software together, Bugs Will Be Found.

Ha. Back in the day when in my Freshman year in Engineering, we were all forced into our very first programming class. In FORTRAN. (Should note: Unix and C were very, very new at the time and not seeing widespread use.) The prof made a point of telling all us newbies the below:

General rule: Write a 300 line program, not including whitespace and comments. 'Way more than half of the lines will have errors in them. Typos, everything else.

Before compiling, run Lint or the equivalent. This knocks the error rate to 25% of the lines. A lot fewer typos, but there'll be plenty of other errors.

Try compiling. The number of errors after fixing all the ABENDS and such in the compiler drops the error rate to 10%.

Start testing the software, varying the inputs and such. If one is lucky, error rate falls to 1%. And, for a very, very long time, that's where things stayed stuck. There might be errors in the code, but the SW will work well enough.

Now, imagine a code base of a million lines. Or two million lines. At 1% error rate, we're talking a 100,000-200,000 bunch of errors in the code.

If you were wondering why Microsoft/Apple/Adobe/etc have monthly updates, well, now you know .

Modern techniques, like running fuzzers, or using languages that put a straight-jacket on one's coding practices, can perhaps knock the error rate down by another factor of ten, but there'll still be a huge number of errors out there.

The prof stated that, once in a while, some miracle worker or other will write a 300 line routine that has no errors, on the first try. People like that are out there to irritate the rest of us and proving that the exception makes the rule. Which only proves that really, really good programmers are worth their weight in gold and become internationally famous, like Moxie Marlinspike and Bruce Schneier.

I frankly expected a lot fewer ABENDS in my FSDb due to the improvements in FORTRAN and Linters over the last 30 years