FSD Beta v11.x

AlanSubie4Life · Aug 12, 2023

Twiglett said:
Isn’t this “almost complete rewrite” number four or is it five? I lose track of how many times they’ve realized the path they’re on won’t get them to real FSD.

It’s so transparent now. Obviously v12 won’t get them there either! Will it be better? Sure! They can quantify that and they won’t release it unless on average it is better.

But we are orders of magnitude away from good enough, so got a lot of “better” to go. I’m looking forward to 90% on ULTs.

Think I might be waiting a while!

When will we get to the first 9??? Then the march can begin!!!

Yelobird · Aug 12, 2023

AlanSubie4Life said:
It’s so transparent now. Obviously v12 won’t get them there either! Will it be better? Sure! They can quantify that and they won’t release it unless on average it is better.

But we are orders of magnitude away from good enough, so got a lot of “better” to go. I’m looking forward to 90% on ULTs. Think I might be waiting a while!

When will we get to the first 9??? Then the march can begin!!!

The fact that many Pre-Assume V12 won’t work or get them there based on no fact or information defines the mind set. Will it get there, have no clue personally butAssuming it won’t seems like a foolish guess.

kabin · Aug 12, 2023

Yelobird said:
The fact that many Pre-Assume V12 won’t work or get them there based on no fact or information defines the mind set. Will it get there, have no clue personally butAssuming it won’t seems like a foolish guess.

Given FSD's track record and Tesla's goal of cost over performance it's been a pretty accurate guess. Optimism falls into the foolish category.

sleepydoc · Aug 12, 2023

kabin said:
Given FSD's track record and Tesla's goal of cost over performance it's been a pretty accurate guess. Optimism falls into the foolish category.

FSD has a track record of steady, incremental improvement. I’m not making any excuses for Elon’s habitual…err… ‘overoptimism,’ nor am I saying I don’t wish it would improve faster, simply pointing out the fact that since it’s release FSD has improved dramatically.

sleepydoc · Aug 12, 2023

Twiglett said:
You just used the word “normally“ in the context of Tesla and FSD rollout

I was more referring to how software releases in general are numbered.

AlanSubie4Life · Aug 12, 2023

Yelobird said:
based on no fact or information

* All prior promises of mind blowing have been broken.
* There is no evidence that LLMs or similar will work for driving, due to their significant limitations. No one knows how to make FSD work (there is no one currently who can do it, or even get close).
* There seem to be hardware limitations for HW3 (which have been known for years, but are more topical, since people can now compare to HW4 (not necessary to compare of course)).
* Every major release leads to incremental, fairly significant improvements. (Obviously not enough for what is required.)
* Intervention rates are currently orders of magnitude higher than they need to be.

There’s really not a pre-assumption or mindset. I’ll be the first to say I was wrong if there is some step change in performance. It’s just about setting reasonable expectations, so I know what I’m likely to get with v12. Helpful for planning purposes!

Yelobird · Aug 12, 2023

AlanSubie4Life said:
* All prior promises of mind blowing have been broken.
* There is no evidence that LLMs or similar will work for driving, due to their significant limitations. No one knows how to make FSD work (there is no one currently who can do it, or even get close).
* There seem to be hardware limitations for HW3 (which have been known for years, but are more topical, since people can now compare to HW4 (not necessary to compare of course)).
* Every major release leads to incremental, fairly significant improvements. (Obviously not enough for what is required.)
* Intervention rates are currently orders of magnitude higher than they need to be.

There’s really not a pre-assumption or mindset. I’ll be the first to say I was wrong if there is some step change in performance. It’s just about setting reasonable expectations, so I know what I’m likely to get with v12. Helpful for planning purposes!

Fair enough all options count granted nothing makes them fact.
* Mind blowing is one persons Opinion of his teams efforts. Could say they suck and it’s kinda better but that’s not really a motivation or positive drive to improve. His opinion of his team would be his fair opinion. Don’t see how thats wrong.
* No evidence, agreed. Granted also no evidence that this isn’t the team that might solve this Never before solved program.
* HW3 limitations at this point have not been validated as certain, they have been Assumed for years not known. The only thing we know now in comparison to HW4 is that the cameras look clearer which honestly means nothing. All cameras and CPU’s improve over time year that doesn’t make it certain that the old ones won’t work. See iPhone……
*Improvements take time. As the base simple 90% foundation is laid, the last Hard 10% will always create unknown issues and complications. 1 action often creates many reactions. It will be slower than all expected.
* Interventions Will increase as the complexity, changes, and expectations of the product develop. Early on most Every step was an improvement, now every step complicates All previous changes and expectations.
Would not in Any way say I wouldn’t want a finished product sooner or that this one is perfect. (Far from it). As long as they continue making an effort I am willing to stay for the ride. The day they say ”we give up” or “XX product is legacy out” then I will jump on board with the “J” clown posse. Lol

jebinc · Aug 12, 2023

Twiglett said:
Isn’t this “almost complete rewrite” number four or is it five? I lose track of how many times they’ve realized the path they’re on won’t get them to real FSD.

Yep, The Great Charlatan is also hard at work!

JHCCAZ · Aug 12, 2023

Among the usual range of personalities at work, we had one particular talented but highly cynical engineer. He was a friend but a very persistent Danny Downer .

We necessarily would have ongoing discussions about challenges and possible solutions, he would reliably cut in and predict failure, delay and disappointment. Most of the time I would take it in stride, but on occasion I felt it was interfering with the progress of the discussion. Then I would point out to him that predictions of trouble in engineering, and particularly predictions of delay, are really the only technically safe forecasts. You'll be right 90+% of the time, but is that helping us advance towards the goal?

Don't get me wrong, a good engineer needs to focus on knowing what he doesn't know, needs to have a healthy paranoia and an element of failure contingency planning. Furthermore, a degree of humor and self-deprecating observations keeps the team honest with themselves, actually sustaining morale through difficult periods. But these are all characteristics of healthy skepticism and within the team, contribute towards the eventual solution.

I love following and participating in this FSD project. In my situation I can see that I will actually need it someday. There is definitely, up to now at least, nothing like this available for any other car I can buy.

I also enjoy the banter on the forum here. But tlately, it seems that the ratio of bitterness to enthusiasm has gone somewhat out of whack, and as usual people are forming up teams. It's hard to make a bemused comment about FSD problems for delays without triggering an avalanche of worn-out "we've all been robbed" bitterness. Completely exaggerated accounts of the death machine mostly go without challenge, while positive comments are quickly laughed at and derided, if not attacked.

Supposedly it's different over in the investor threads, but I'm not really looking for a safe space to hang out; I'd just like to be able to discuss and theorize without being forced to swear loyalty to either the Defenders or the Destroyers.

I do agree that Elon could mitigate this and set intermediate expectations more accurately, but I can't really say that it would improve the attitude of the committed cynics.

Anyway, time for lunch.

Yelobird · Aug 12, 2023

JHCCAZ said:
Among the usual range of personalities at work, we had one particular talented but highly cynical engineer. He was a friend but a very persistent Danny Downer .

We necessarily would have ongoing discussions about challenges and possible solutions, he would reliably cut in and predict failure, delay and disappointment. Most of the time I would take it in stride, but on occasion I felt it was interfering with the progress of the discussion. Then I would point out to him that predictions of trouble in engineering, and particularly predictions of delay, are really the only technically safe forecasts. You'll be right 90+% of the time, but is that helping us advance towards the goal?

Don't get me wrong, a good engineer needs to focus on knowing what he doesn't know, needs to have a healthy paranoia and an element of failure contingency planning. Furthermore, a degree of humor and self-deprecating observations keeps the team honest with themselves, actually sustaining morale through difficult periods. But these are all characteristics of healthy skepticism and within the team, contribute towards the eventual solution.

I love following and participating in this FSD project. In my situation I can see that I will actually need it someday. There is definitely, up to now at least, nothing like this available for any other car I can buy.

I also enjoy the banter on the forum here. But tlately, it seems that the ratio of bitterness to enthusiasm has gone somewhat out of whack, and as usual people are forming up teams. It's hard to make a bemused comment about FSD problems for delays without triggering an avalanche of worn-out "we've all been robbed" bitterness. Completely exaggerated accounts of the death machine mostly go without challenge, while positive comments are quickly laughed at and derided, if not attacked.

Supposedly it's different over in the investor threads, but I'm not really looking for a safe space to hang out; I'd just like to be able to discuss and theorize without being forced to swear loyalty to either the Defenders or the Destroyers.

I do agree that Elon could mitigate this and set intermediate expectations more accurately, but I can't really say that it would improve the attitude of the committed cynics.

Anyway, time for lunch.

Agreed and well said. If Elon shares his opinions or belief it quickly becomes a lie or missed promise. IF at some point he simply Stops all sharing about progress and simply says “wait for the next release notes“ there would be a reverse mutiny for not telling us what we want to know as if it was owed. Can’t win either way so sadly there will always be defenders and destroyers forever. It’s the entitlement way.

swedge · Aug 12, 2023

AlanSubie4Life said:
* All prior promises of mind blowing have been broken.
* There is no evidence that LLMs or similar will work for driving, due to their significant limitations. No one knows how to make FSD work (there is no one currently who can do it, or even get close).
* There seem to be hardware limitations for HW3 (which have been known for years, but are more topical, since people can now compare to HW4 (not necessary to compare of course)).
* Every major release leads to incremental, fairly significant improvements. (Obviously not enough for what is required.)
* Intervention rates are currently orders of magnitude higher than they need to be.

There’s really not a pre-assumption or mindset. I’ll be the first to say I was wrong if there is some step change in performance. It’s just about setting reasonable expectations, so I know what I’m likely to get with v12. Helpful for planning purposes!

I think FSD is mind blowing, and has been so all along. Even the standard AutoPilot functions are amazing to me. A computer driving my car? This is f'n amazing. Not ready for prime time yet, but mind blowing none the less, in my opinion.

It might help me better understand what you mean if you restate the goals you are comparing against.

Ather question, how would a Large Language Model take in video and drive a car anyway? Or is this just a straw-man argument?

A while back I compared the HW3 cameras' angular resolutions with those of human foveal vision, and was surprised to find it comparable. So it is not clear to me, at least, that HW4 is necessary, on that account anyway. The mind blowing bit is taking all those pixels and driving around at all. So it seems to me that compute speed and bandwidth could be limiting factors, but then we won't know that until the code is optimized for the existing HW, which as I understand it was designed for NN's rather than 300k lines of C code just to choose a lane.

Most of my interventions are for my comfort or that of other nearby drivers. Most of the things I see it do wrong are not really safety related. I expect safety is the first goal to achieve, and after that comfort may be fairly easy.

Still, I am not yet able to relax while FSD is driving. My wife doesn't use it at all, and prefers that I don't. Heck, I'm not even clear on what the use-case is for FSD, but I do expect I will understand that better when they take away my license due to age-related decrepitude.

I recall a study done a while back which found that many Americans are not good at distinguishing between statements of opinion and statements of fact. Perhaps, to avoid confusion, we could be more careful in our posts to clearly state opinions as such. (Note the words "recall" and "perhaps" in this paragraph. I'm at least trying.)

I'm trying to be patient while waiting for a new release. Considering dropping out of Beta, though, because 11.3.6 was a bit better for me, and I'm anxious to see if Charge on Solar works for me.

jebinc · Aug 12, 2023

swedge said:
I think FSD is mind blowing, and has been so all along. Even the standard AutoPilot functions are amazing to me. A computer driving my car? This is f'n amazing. Not ready for prime time yet, but mind blowing none the less, in my opinion.

It might help me better understand what you mean if you restate the goals you are comparing against.

Ather question, how would a Large Language Model take in video and drive a car anyway? Or is this just a straw-man argument?

A while back I compared the HW3 cameras' angular resolutions with those of human foveal vision, and was surprised to find it comparable. So it is not clear to me, at least, that HW4 is necessary, on that account anyway. The mind blowing bit is taking all those pixels and driving around at all. So it seems to me that compute speed and bandwidth could be limiting factors, but then we won't know that until the code is optimized for the existing HW, which as I understand it was designed for NN's rather than 300k lines of C code just to choose a lane.

Most of my interventions are for my comfort or that of other nearby drivers. Most of the things I see it do wrong are not really safety related. I expect safety is the first goal to achieve, and after that comfort may be fairly easy.

Still, I am not yet able to relax while FSD is driving. My wife doesn't use it at all, and prefers that I don't. Heck, I'm not even clear on what the use-case is for FSD, but I do expect I will understand that better when they take away my license due to age-related decrepitude.

I recall a study done a while back which found that many Americans are not good at distinguishing between statements of opinion and statements of fact. Perhaps, to avoid confusion, we could be more careful in our posts to clearly state opinions as such. (Note the words "recall" and "perhaps" in this paragraph. I'm at least trying.)

I'm trying to be patient while waiting for a new release. Considering dropping out of Beta, though, because 11.3.6 was a bit better for me, and I'm anxious to see if Charge on Solar works for me.

Let me know if you find a way, “out”. I have not and tried everything, short of bricking the car. Good luck with that, should you choose to pursue.

Knightshade · Aug 12, 2023

Yelobird said:
* HW3 limitations at this point have not been validated as certain, they have been Assumed for years not known. The only thing we know now in comparison to HW4 is that the cameras look clearer which honestly means nothing.

This is just flat out false.

We've known, for years now, that Tesla was forced to use extended compute on node 2 of HW3 to get even the L2, missing things needed for higher, system that have today working as well as it does.

Not "assumptions" but "observed fact in the functioning of the code"

Yelobird said:
All cameras and CPU’s improve over time year that doesn’t make it certain that the old ones won’t work. See iPhone……

If you watch the video HW3 cam vs HW4 cam comparisons it's obvious HW3 can't see most signs remotely legibly at any useful distance. Which probably explains why years into this beta Tesla has not added a single bit of sign-reading at all beyond the original stop signs, speed limit signs, and yield signs (and it still fails on those sometimes).

Software isn't sorcery- if the data isn't there it isn't there.

Not sure what the iphone analogy is supposed to mean though-- iphones already do what they're supposed to when released- there's not some imaginary capabilities they hope it can do one day on the existing HW they sell for thousands of dollars up front.

swedge · Aug 12, 2023

jebinc said:
Let me know if you find a way, “out”. I have not and tried everything, short of bricking the car. Good luck with that, should you choose to pursue.

Will do.

I see on TeslaFi that a small number, over 30, of those on current production branch came there from 2023.7.20. Perhaps all FSD subscribers rather than us purchasers. Still, the path exists, if we can just get our cars into the correct exit lane. ;-)

LowlyOilBurner · Aug 12, 2023

Yesterday, 411miles on FSDb(j) 11.4.4. For the most part, pretty decent. Two or the intersections, thought the car was going to just pull right out into traffic, not creeping slowly enough and had to take over out of fear.

Highway should be L3 at this point. Only complaints are:

Not slowing/moving over for merging traffic soon enough.

Not exiting the passing lane soon enough.

Pretty decent for over 400miles in a day.

Still some dumb lane choices on some roads I travel numerous times per week, mainly directly in front of the old Bethlehem Steel Plant in South Buffalo NY (newly CLOSED Republic Steel, I should say). Like, awful lane choices. Have to make a left turn ahead, but FSD is trying pull into the steel plant a half mile before the actual left turn it needs to make. It’s embarrassing and aggravating.

AlanSubie4Life · Aug 12, 2023

swedge said:
how would a Large Language Model take in video and drive a car anyway? Or is this just a straw-man argument?

I said "or similar." I am speaking of the analogous methods required for driving (discussed extensively elsewhere). LLMs and what the AI in self-driving cars do is not that dissimilar and they are picking statistically the best solution, doing error minimization (in a statistical sense) at some level. (As I understand it; I'm not trying to be super precise here other than to say there are similarities in the design methods and the misbehaviors.) Hallucinations and wrong decisions are not allowed! So no, not a straw man; lots of similarities, though presumably the LLMs of ChatGPT etc. are way more complicated than what can reside in the FSD computer.

swedge said:
It might help me better understand what you mean if you restate the goals you are comparing against.

Since we're sort of talking about Elon and his statements, I guess I just mean what Elon has stated, maybe 5x better than humans or whatever - he's said a number of different numbers over the years, but somewhere getting close to an order of magnitude better than a decent human. (I would be pretty amazed at similar to a decent human, though it would not be good enough - this is definitely not good enough for Elon.)

swedge said:
I think FSD is mind blowing, and has been so all along. Even the standard AutoPilot functions are amazing to me. A computer driving my car? This is f'n amazing. Not ready for prime time yet, but mind blowing none the less, in my opinion.

It's pretty interesting and quite good as a driver's aid. It's possible that with an attentive human (a huge ask, as any AP/FSD user knows) it's safer than just driving the car myself (on the freeway only of course) - of course this safety benefit would require me to take over fairly frequently.

swedge said:
I compared the HW3 cameras' angular resolutions with those of human foveal vision, and was surprised to find it comparable.

I guess you can just see if it can read street signs sufficiently well. Seems to me it can't, at least for the desired level of performance, but maybe I'm wrong.

swedge said:
Most of the things I see it do wrong are not really safety related.

I think this is hard to judge.

swedge said:
My wife doesn't use it at all, and prefers that I don't.

Probably she is making some sort of safety judgement at some level and feels that sudden jerks and movements which are unnecessarily harsh, and getting too close to other traffic, or not slowing down for obvious stopped traffic as early as possible, are less safe (she's probably right).

swedge said:
we could be more careful in our posts to clearly state opinions as such.

Everything I've said is my opinion. I'm making predictions about how things will turn out. If I'm still active on the forums (unlikely) I'll try to be the first to admit I was wrong, when that becomes clear.

I'm honestly surprised that we're not at better than 90% on ULTs at this point. If someone had had me make predictions a year ago for what would be the situation today, I would have been wrong. I would have guessed better than 90% success would be achieved at this point. So I think that means I was wrong about that (even though I scored a lot of beers along the way for being right at interim points).

swedge · Aug 13, 2023

AlanSubie4Life said:
Since we're sort of talking about Elon and his statements, I guess I just mean what Elon has stated, maybe 5x better than humans or whatever - he's said a number of different numbers over the years, but somewhere getting close to an order of magnitude better than a decent human. (I would be pretty amazed at similar to a decent human, though it would not be good enough - this is definitely not good enough for Elon.)

Are you talking about crash statistics with FSD engaged? Or maybe FSD engaged fatalities? How are we doing on those? Do we even have statistics?

I recall some discussions about such stats, including issues of potential selection bias in beta volunteers, freeway vs city streets (FSD had not yet been enabled on freeways if memory serves me), and FSD being human plus FSD, not pure FSD. But the issues were raised to undercut some stats Tesla had published which suggested that driving with FSD engaged was already substantially safer than without.

My take is that market uptake of FSD will be constrained more by the herky-jerky driving and the uncomfortable and unexpected things it does than by actual accidents.

Regarding actual driver-less self driving cars, tonight's 11:00 news in San Francisco's KPIX lead story was a dozen Cruz robotaxies blocking up streets for 15 minutes, apparently due to a nearby large concert which overloaded the local cell phone data service. Oops! Yet another failure mode when "autonomous" needs realtime cloud support. Tesla FSD also needs the cloud at least for route planning, I think. I don't know if the connection must remain intact for the trip duration. My old Prius did not use the cloud at all, but had a dvd drive under the driver's seat with the map data. We would use iPad/google maps to check traffic and deviate from the proposed route if it seemed appropriate.

JulienW · Aug 13, 2023

swedge said:
Will do.

I see on TeslaFi that a small number, over 30, of those on current production branch came there from 2023.7.20. Perhaps all FSD subscribers rather than us purchasers. Still, the path exists, if we can just get our cars into the correct exit lane. ;-)

Old known news that FSD subscribers can simply let their subscription lapse and get the current software/11.3.6 and what is shown on TeslaFi. It is us who bought FSD have no clear path to get the current software. In a nutshell subscribe and it is easy, buy and you are stuck barring some special circumstances.

Twiglett · Aug 13, 2023

LowlyOilBurner said:
Yesterday, 411miles on FSDb(j) 11.4.4. For the most part, pretty decent. Two or the intersections, thought the car was going to just pull right out into traffic, not creeping slowly enough and had to take over out of fear.

Highway should be L3 at this point. Only complaints are:

Not slowing/moving over for merging traffic soon enough.

Not exiting the passing lane soon enough.

Pretty decent for over 400miles in a day.

Still some dumb lane choices on some roads I travel numerous times per week, mainly directly in front of the old Bethlehem Steel Plant in South Buffalo NY (newly CLOSED Republic Steel, I should say). Like, awful lane choices. Have to make a left turn ahead, but FSD is trying pull into the steel plant a half mile before the actual left turn it needs to make. It’s embarrassing and aggravating.

sorry, but there is no way highway is L3 with a "few complaints"
The major "will kill you if this was L3" are things like taking stub right hand turn lanes that should be driven at <40, but trying it at 75. Or coming to a complete stop on a 70mph highway because it it still overrides what its cameras can see and instead uses map data.
This is a known major issue for 11.4.4 with a great many video examples. The incorrect stopping has been an issue for all 11.x and most of 10.x before that
Nope, not L3

Mardak · Aug 14, 2023

sleepydoc said:
Changing the underlying version would require a significant rewrite which wouldn’t normally happen between 11.4.6 as of 11.4.7

The same FSD Beta 11.3.6 version has been on multiple underlying versions: 2022.45.x, 2023.12.x, 2023.20.x, 2023.26.x; so hopefully that indicates there aren't FSD-Beta-incompatible changes requiring significant rewrites for the main production vehicle software. And Tesla has updated the underlying versions in the past while also changing FSD Beta versions, e.g., 2022.36.20 / FSD Beta 10.69.3.1 -> 2022.40.4.5 / FSD Beta 10.69.3.2.

This could still mean Tesla initially releases 11.4.7 on 2023.7.x before merging it with say 2023.26.x that could go to both general population and "early access."

FSD Beta v11.x

Efficiency Obsessed Member

Active Member

Active Member

Well-Known Member

Well-Known Member

Efficiency Obsessed Member

Active Member

Well-Known Member

Electrified Engineer

Active Member

Member

Well-Known Member

Well-Known Member

Member

Active Member

Efficiency Obsessed Member

Member

Well-Known Member

Single pedal driver

Active Member

Similar threads