Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
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.


screenshot-teslamotorsclub.com-2022.01.26-21_30_17.png


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).

 
Last edited:
Hesitate at a turn: Touch the pedal to encourage the car. Janky deceleration coming to a stop sign: Touch the pedal to smooth out the deceleration.

This doesn't necessarily apply to you, but I feel like some testers refuse to work with the software and just expect perfection, possibly due to paying too much for a beta experience. I have hundreds of hours on 11.3 & 11.6 now and it just needs to be played like a musical instrument.

Yes, it can do crazy things and the lane selection is poor, but I'm pretty sure that's part of BETA testing.

This thread is nearly unreadable because of certain naysayers high-fiving each other for their new beta acronym and latest one-line refutation of the hard work going on behind the scenes, any chance we can discuss what's better without all the FUD? Because things are better, just look at the reaction from the guy on the California road trip 10 pages ago. The tone in here is acidic and it smacks of regret for bad financial decisions.
Naysayers? That’s terrible! We should run them out of here.

Jonque.
What’s the trick for “pull across traffic in front of an oncoming truck then come to a complete stop”?

FSDb really REALLY bad in my car. There seems to be a wide spectrum of experiences, of course, but my car can go maybe a mile or two before it puts itself into a situation that can’t be overcome by nudging the throttle or something.

My litmus is how long I can go before I get “turn that #*%+ing thing off!” from my spouse. It’s been about half a mile since September of 2021.
Same experience here, wonder if you also drive a flagship refresh model S/X.
Thank you for making my points for me:

1. You expect bleeding edge software to work perfectly.

2. You admit to regret from a bad financial decision
Let’s talk about #1. If Elon announced FSD late last year and said it’ll come out January 1st as an alpha/beta, then fast forward to today and people expected it to be perfect - I agree, that would be unreasonable.

Is that the case though? Or has Elon lied to us for many years, with many years of internal beta testing, followed by over a year of public beta testing. The expectation is that 4 years later this should NOT be a “bleeding edge” software that people paid $6,000-$15,000, some multiple times, in multiple vehicles, without ever enjoying the end product.

Elon’s attitude and promises towards it DID NOT help the situation.
 
10.69 was arguably V11.

Elon should keep with the program and release it as 11.69
Anyway, the whole idea of "end-to-end NN" is just another pie-in-the-sky.

Rules of the road can't all be inferred by NN - besides they change somewhat based on jurisdiction.

v12 is reserved for when FSD is end-to-end AI, from images in to steering, brakes & acceleration out.
 
It's likely because FSD still has mulitple neural nets making the final driving control decision. Elon had this back to say about FSD v11.3 on Jan 14, 2023:


Telsa is still in the process of cutting over all its legacy C++ code into neural nets, and then making those NNs play nice with eachother. Expect it to have some wrinkles, they're working on it.

This is a consequence of the ongoing effort to create a "unified vector space" for FSD (see #348,760) IE: with 11.3.x FSD creates two vector models of the world, one for static objects, and another for moving objects. It's easy to imagine how any miscategorization of a moving object into the static space could result in sudden lurchs and abrupt manoeuvers. Elon explains the potential errors that could arise in this discussion:

(41:45) FSD Beta Development "Unified vector space" | TSVOC on Youtube (Jun 22, 2022)


Again, they're working on it. Tesla will add 10x compute power to its training cluster in 2023, and plans a further 10x increase in 2024. They'll get there (did I mention it takes time?) ;)

Bottom line is IT'S STILL BETA, YOU ARE THE DRIVER. and you are responsible, regardless. If however, you think you are NOT able to properly supervise the beta features, please turn off FSD beta and drive manually (see above).

Cheers!
Thanks. Tha explanation makes sense. I don’t have the technical insights. When I am driving and it makes a turn - which it had done impressively correct numerous times - and it suddenly takes the wrong lane I am “WTF’.

On the whole, I am using FSD beta like 80% of the time. Love it.
 
You know what really kills me? FSDb, that's what. Happens once or twice a week. I keep getting reincarnated as a Tesla owner with FSDb. What are the odds of that? The Gods must be crazy...........

And that's precisely why I switched to FSDp.....for my M3P. I recommend it........

FSDp.......perfection.....
2ca05i.jpg
 
Last edited:
  • Funny
Reactions: Ramphex
Anyway, the whole idea of "end-to-end NN" is just another pie-in-the-sky.

Rules of the road can't all be inferred by NN - besides they change somewhat based on jurisdiction.

It's not intuitive to me either.

But what video in, brake / accel / steering out entails is that the system is considering all manner of photon information and making planning decisions based on that, so even nuances based on jurisdiction can be trained (because the car can understand the jurisdiction based on the differences in photon information from street signs / lane markings / etc.).

What Tesla is doing is creating unit-testing system comprised of 100s of thousands or perhaps millions of planning examples, the training stack is fed video, along with the proper planning decisions, and then the final output is an end-to-end FSD system, which is then unit-tested again on previously untrained video, and the planner is supposed to navigate through the video in a predictable manner.

I saw a hint of this during the FSD evaluation talk (bicycle example):

 
But what video in, brake / accel / steering out entails is that the system is considering all manner of photon information and making planning decisions based on that, so even nuances based on jurisdiction can be trained (because the car can understand the jurisdiction based on the differences in photon information from street signs / lane markings / etc.).
The end to end NN is something a lot of companies (NVidea has an example from a few years ago) have talked about and examined. But they all came to the conclusion, it doesn't work.

May be Musk has seen how difficult planning & control are to get right over the last 2 years and now has put his faith in end-to-end NN.
 
  • Like
Reactions: Ramphex
Elon’s attitude and promises towards it DID NOT help the situation.
Exactly. I'm actually happy with the latest progress, though lacking in many ways still of course. Elon over-promises and under-delivers consistently. That's the main source of frustration. Were he as realistic as possible and would give even minimal explanations for delays I think we'd all be much more forgiving.
 
But they all came to the conclusion, it doesn't work.

Perhaps it didn't / doesn't work for them because they don't have a large fleet providing millions of planning examples in all sorts of situations, with raw video and all.

Intuitively, it should work: if the car can perceive the world accurately then it should also be able to plan its way around it.
 
Throttle control doesn't always solve incorrect brake or indecisiveness. Sometimes FSD needs to brake and no amount of throttle changes it. In fact sometimes after removing throttle input FSD brakes harder. it's much smoother and safer to disengage when the FSD underwhelms. This isn't a game. And each driver makes their own decision based on comfort, safety, and performance.

It's incorrect to assume 'testers' as you put it should just drive and shut the hell up. We are customers and our input is important for system improvement. On the other hand, it feels like if Elon and the TSLA team had their way they'd have stuck a fork in FSD a long ago with the crap they have been putting out.

After 9 years or so of FSD BS let me be the first to welcome you to reality.
Throttle control (both accelerating and decelerating) solves at least half of my issues on the fly.

If you are on FSDb, then you are a tester. If you want to simply be a customer, go back to legacy stack or yes, just drive. It's not my fault you listened to a huckster.

Newcomers, as someone called it, have a fresh perspective and aren't burdened by their PTSD and regret. You've been in the trenches a long time and I salute you, but if this is just a complaint thread, then some people around here obviously need a break from testing.
 
  • Like
Reactions: JHCCAZ and PACEMD
Naysayers? That’s terrible! We should run them out of here.

Jonque.

Same experience here, wonder if you also drive a flagship refresh model S/X.

Let’s talk about #1. If Elon announced FSD late last year and said it’ll come out January 1st as an alpha/beta, then fast forward to today and people expected it to be perfect - I agree, that would be unreasonable.

Is that the case though? Or has Elon lied to us for many years, with many years of internal beta testing, followed by over a year of public beta testing. The expectation is that 4 years later this should NOT be a “bleeding edge” software that people paid $6,000-$15,000, some multiple times, in multiple vehicles, without ever enjoying the end product.

Elon’s attitude and promises towards it DID NOT help the situation.
Careful, your PTSD is really showing right now. Fool me once, fool me twice, heck I'm a fool! Again, it's not my fault you listened to a huckster.
 
  • Funny
Reactions: Ramphex
Exactly. I'm actually happy with the latest progress, though lacking in many ways still of course. Elon over-promises and under-delivers consistently. That's the main source of frustration. Were he as realistic as possible and would give even minimal explanations for delays I think we'd all be much more forgiving.
I think everyone except Omar is tired of Elon’s ‘overoptimism.’ What surprises me is how some people here continue to believe him and post their disappointment. Honestly, if you’re that dim witted and gullible that you still believe him why are you posting it for everyone to see?
 
I think everyone except Omar is tired of Elon’s ‘overoptimism.’ What surprises me is how some people here continue to believe him and post their disappointment. Honestly, if you’re that dim witted and gullible that you still believe him why are you posting it for everyone to see?
What do you mean? I'm not going to be able to use my car as a robotaxi by the end of the year?

But I overextended my budget to get this car banking on using it as a robotaxi.... What am I going to do now?!?!?!?!?