FSD Beta v11.x

Mardak · Jan 3, 2022

EVNow said:
I wonder whether in their tests raw image at night is doing better than processed

People using TeslaCam or backup/blind-spot reapeater cam views can already see the better night vision provided by the RCCB sensors even after the processing to YUV for human displays. So similarly, the non-raw neural networks were likely already getting some benefit of the camera's sensitivity in low light. The main prediction accuracy benefit of switching to raw might actually just be primarily from an increase in data.

At CVPR'21 in June last year, Karpathy had a slide showing "8 cameras measuring [1280 x 960 x 3] uint8 @ 36Hz" presumably referring to the 3 Y,U,V-converted channels each supporting up to 256 values for ~17M colors (where some say humans generally perceive ~10M colors). Whereas at AI Day just 2 months later, the slides show "raw 1280x960 12-Bit (HDR) @ 36Hz" where assuming these are 3 channels with 4096 values each, that's now a range of ~69B colors (3 channels * 4 extra bits = 4096x).

Bladerskb · Jan 4, 2022

EVNow said:
See the James Douma interview I linked to in OP.

James is the last person anyone should be listening to when it comes to NN or Self driving cars. He is a full blown fanboy and spreads only misinformation

Mardak · Jan 4, 2022

EVNow said:
I think the CNN / Montecarlo must be better or atleast compable. Otherwise they would not use it.

I'm not sure if that's certain. Given that they presented this with a toy problem, it seemed quite early stage. At a high level, conceptually sure it sounds like the right approach of replacing hardcoded heuristics with neural network predictions, but practically there are real-time constraints for FSD that were not as relevant for MCTS used for AlphaZero.

I suppose one thing going for this approach is that the top-down/birds-eye view is basically a 2D projection perhaps not too dissimilar to the Atari games played by MuZero. Also, perhaps not too surprising that Ashok leads/presented both the planning and simulation portions of AI Day as most likely this MCTS approach will need to be evaluated with the simulator.

EVNow · Jan 4, 2022

Bladerskb said:
James is the last person anyone should be listening to when it comes to NN or Self driving cars. He is a full blown fanboy and spreads only misinformation

Did you listen to the interview ? Doesn't sound like a fanboy at all and definitely sounds a lot more experienced in NN or self driving than you.

EVNow · Jan 4, 2022

Mardak said:
I'm not sure if that's certain. Given that they presented this with a toy problem, it seemed quite early stage. At a high level, conceptually sure it sounds like the right approach of replacing hardcoded heuristics with neural network predictions, but practically there are real-time constraints for FSD that were not as relevant for MCTS used for AlphaZero.

All I'm saying is that is one of the first questions they would have asked.

Bladerskb · Jan 4, 2022

EVNow said:
Did you listen to the interview ? Doesn't sound like a fanboy at all and definitely sounds a lot more experienced in NN or self driving than you.

You mean the guy who said that the best lidar will only give you a-couple points on a pedestrain a few meters away (when the truth is in the thousands) or when he said that Lidar in 10 years will still be expensive and cost 10k (when the reality today is already $100-$500 at scale).

Or when he says companies only use lidar for localization or that he won't talk about Nvidia's Orin chip because it comes out in 2024/2025 (when Orin is already in production and will be on the NIO ET7 in 2 months).

I could go on about NN and vidar and depth map and radar.
I could literally make a 1 hr youtube video on the stupid completely nonsense patently and demonstrately false things he says or advocate.

I haven't found someone more clueless about NN and self driving with a big platform as much as James Douma. He has literally zero professional experience in AI or NN or EE.

You are better off going through a free udacity or udemy course or playlist on YT than watching a video with James on it.
Notice how the TSLA community love promoting people with absolutely zero experience, zip.

Oh but hey lets listen to James promote a anti-mobileye video from another tsla fanboy with absolute nonsense gibberish from start to finish as "mostly right". There's no one more clueless than james.

https://twitter.com/x/status/1434659914526302208

EVNow · Jan 4, 2022

Bladerskb said:
I could literally make a 1 hr youtube video on the stupid completely nonsense patently and demonstrately false things he says or advocate.

Why don't you ? Infact, why don't you make a one hour video about Tesla FSD, commentary about AI day or version 11 (which BTW, is the topic here). Then, we can judge how much you know ...

I don't car what he says about MobilEye or radar. Those are purely OT here. Tell us if he has said something fundamentally wrong about Tesla ... (you may have noticed this is a Tesla forum, for Tesla owners and enthusiasts).

kellyds · Jan 4, 2022

EVNow said:
Elon says 11 was planned for this year but was *surprise* delayed. It is in ”alpha” now.

Question is what is alpha … rough with a lot of bugs ? But that is what Beta looks like !

To me alpha sounds like a POC - proof of concept. Significant dev work remains, apart from fixing bugs. Hope we get 11 by this time next year.

More appropriate might be MVP (minimum viable product) and MVCR (minimum viable capability release) While everything may not be there, those components that ARE there are functioning as expected by the development team and stakeholders. This pattern matches rapid release cycles that are observed from Tesla. I strongly suspect the release cycle internally is much more frequent than what we see from the outside.

Alpha-Beta mentality really equates to the old waterfall methodology and that is certainly not what we are seeing from Tesla (71 updates over 30 months - that is FAST for this type of development)

israndy · Jan 4, 2022

I can only imagine what you are replying to there, as I have set Bladerskb to be ignored, it's blissful

Mardak · Jan 5, 2022

EVNow said:
No they use a hybrid approach now

When you say "now," do you mean in current FSD Betas or are planning to release in an upcoming version maybe 11? They showed this cost-minimized planned path at AI Day:

But that doesn't seem to be reflected here in 10.8 even though the visualization of the road was quite stable while the predicted path seems to be closer to a low confidence neural network output wanting to encroach on the outer left turn lane:

Mardak · Jan 5, 2022

kellyds said:
Alpha-Beta mentality really equates to the old waterfall methodology and that is certainly not what we are seeing from Tesla (71 updates over 30 months - that is FAST for this type of development)

At least from a developer perspective, it seems obvious Tesla firmwares are released from multiple branches of parallel development that can merge in changes when ready, so it wouldn't be surprising if Autopilot software and neural networks are developed similarly. "Alpha" could just be some development branch that hasn't made it to the main release branch, which happens to be called "FSD Beta," and internal builds can be made from these in-progress branches for special deployment to employee vehicles.

With multiple feature branches, even if something was originally targeting a FSD Beta 11 release, potentially a feature could be merged in to the 10.x branch with relatively minimal additional work. Tesla has been working on "Beta 11" even before "Beta 10" was released, and we've seen continued improvements with 10.x releases very likely including changes that first landed as part of 11.

Given the special aura around the "Holiday Release," there could have been an extra push to do more than a minimal amount of work porting over changes to 10.7/10.8 leading to Elon Musk's "should arguably be called 11" comment probably more of a praise to the team in getting so many changes into this "minor" 10.x version (but 11 name is "reserved" for single stack).

Mardak · Jan 8, 2022

EVNow said:
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

I wonder if the explicit mention of city + highway means parking lots won't be significantly improved with the initial Beta 11.x releases?

https://twitter.com/x/status/1479590174828998663

Smart Summon does seem to use some pretty specialized control different from "regular Autopilot" so with the many changes to the FSD Beta perception outputs, merging the stack here might be too much for now.

JulienW · Jan 8, 2022

Mardak said:
I wonder if the explicit mention of city + highway means parking lots won't be significantly improved with the initial Beta 11.x releases?

Smart Summon does seem to use some pretty specialized control different from "regular Autopilot" so with the many changes to the FSD Beta perception outputs, merging the stack here might be too much for now.

For $12K it should be.

Likely it is all hands on deck to get highway "perfected" and Stack integrated with little engineering time for Smart Summon. Smart and Summon don't seem to see and take into account lanes and see small objects. It seems to be mostly Ultrasonic based. Needs to be ported over to Vision based. But I bet that is going to be a while.

JohnnyEnglish · Jan 8, 2022

EVNow said:
''''
But I've to say ... the FSD team is probably interested in working on 11 rather than making small improvements to 10.x. Afterall they keep making improvements to 10.x and a lot of us think there isn't much of an improvement or there are a lot of regressions.

My guess is that both of these statements are linked:

The FSD team will not be putting much (if any) effort into those areas that are replaced in 11. The improvements in 11 seem to be at the foundational levels - perception etc
Improvements to 10.x will be in those areas that will not be replaced by 11. These seem to be more in the areas which utilise the results of the foundational levels - planning etc. Hence these improvements may well be more effective than they are currently when utilised with the better foundations provided by 11.

Edit: Oops, missed a page of the thread where I now see this has been discussed. Not sure what the outcome was though

. On the 1 team or 2 teams question I get the feeling that whilst there will be some people currently working predominantly on 11 and some working predominantly on 10.x they will all be working as one team with the necessary interactions being managed through the agile 'reviewing' processes.

EVNow · Jan 8, 2022

Mardak said:
Tesla has been working on "Beta 11" even before "Beta 10" was released, and we've seen continued improvements with 10.x releases very likely including changes that first landed as part of 11.

I think some of the work done was POC level and could even have been done for AI day. It seems to me most of what they showcased in AI day is part of 11 which they planned to release last year, may be as part of Holiday release. Some of those might have been integrated already, but difficult to tell. The 3D environment related improvements are the ones most likely in prod already.

planetary · Jan 8, 2022

I just hope they let in a lot of us folks with qualifying scores for the 10.8.1 or 10.9 releases. I can't imagine having to maintain a 98+ score until next month, or whenever v11 ships. Sweet merciful Zeus.

EVNow · Jan 8, 2022

planetary said:
I just hope they let in a lot of us folks with qualifying scores for the 10.8.1 or 10.9 releases. I can't imagine having to maintain a 98+ score until next month, or whenever v11 ships. Sweet merciful Zeus.

Why not reset the score, get to hundred ? You could have FSD when they rollout to new testers next week …

planetary · Jan 8, 2022

EVNow said:
Why not reset the score, get to hundred ? You could have FSD when they rollout to new testers next week …

I'm at 99 now with several hundred miles. Is resetting to get to 100 really going to increase my chances?

EVNow · Jan 8, 2022

planetary said:
I'm at 99 now with several hundred miles. Is resetting to get to 100 really going to increase my chances?

Oh yes, you should be able to get it with 99. No reason to reset … not sure about 98.

israndy · Jan 8, 2022

I was at 99, following all those YouTube hints to keep my score high and trying to get it back to 100, seems just resetting alone took me to 99 each time. I was doing donuts in the nearby parking lot, driving behind semis down the highway, religiously avoiding the brake pedal, resetting the car when it warned me about a front end collision, watching EACH metric and trying to get the individual number down. But then I did a road trip to Vegas and was just in too many situations where I lost points to keep my score high, it dropped to 98, and then 97, I was in agony, but then the next morning, xMas day, I got the update to the Beta.

I think it's gotta be a few days before a major release (like the 10.8.1 that's been seen now) that they scan for cars that match the profile for adding to the Beta. It seems like once it starts rolling out, people who get their score up to 99 or 100 don't magically join then, until the next round.

FSD Beta v11.x

Active Member

Senior Software Engineer

Active Member

Well-Known Member

Well-Known Member

Senior Software Engineer

Well-Known Member

New Member

Supercharger Hunter

Active Member

Active Member

Active Member

Well-Known Member

Member

Well-Known Member

EndlessCheese spreader

Well-Known Member

EndlessCheese spreader

Well-Known Member

Supercharger Hunter

Similar threads