FSD v12.x (end to end AI)

[Daniel in SD]

Some major déjà vu happening here reading everybody's take on the next, yet-to-be-experienced-but-promised-to-be-fire FSD version. Either this is a whole new group of posters or the posters here have short-term memories. What I have learned from driving on the FSD ADAS suite since October of 2018 (when they released NoA) is that Elon's, Omar's, Dirty Tesla's, and all the rest's comments don't mean squat - I'll believe it when I drive it.

Same people. haha.

I'm loving the 11.4 videos. This year might finally be the year we get to 2x safety fsd with no interventions in most of the USA by land area.

 

[AlanSubie4Life]

Some major déjà vu happening here reading everybody's take on the next, yet-to-be-experienced-but-promised-to-be-fire FSD version. Either this is a whole new group of posters or the posters here have short-term memories. What I have learned from driving on the FSD ADAS suite since October of 2018 (when they released NoA) is that Elon's, Omar's, Dirty Tesla's, and all the rest's comments don't mean squat - I'll believe it when I drive it.

Reading between the lines, I think the expectation of most here is that v12 will just be a slight improvement over v11, and people are hoping it will somehow improve faster than v11 did.

Other than the normal select few, who are mostly tongue in cheek, I don’t think anyone expects any ADAS revolution here.

There may also be low morale amongst those hoping for L3. Tantalizingly close, but also so far away.

 

[lzolman]

Some major déjà vu happening here reading everybody's take on the next, yet-to-be-experienced-but-promised-to-be-fire FSD version. Either this is a whole new group of posters or the posters here have short-term memories. What I have learned from driving on the FSD ADAS suite since October of 2018 (when they released NoA) is that Elon's, Omar's, Dirty Tesla's, and all the rest's comments don't mean squat - I'll believe it when I drive it.

IMO there's a qualitative difference now, because of the NN approach. Prior to V12, what you saw is all you got. There was no guarantee it would necessarily get better; the weak points rotated around for the past year but the overall quality of the drive did not improve (for me it went from tolerable to less than that). But if the first public release of V12 is as good as it looks from Omar's vids, at there's hope now, if you believe in machine learning. I mean, what in the world could ever cause it to regress now? If we ever see it regress, the only conclusion would be that the NN is broken (and that may very well still happen, we just don't know...)

But if the NN actually works, it can only get better (more reliable, more confident, less hesitant, safer, yada yada). So that's what I'm waiting to see. I'm not running around saying "Tesla has finally solved FSD" but I'm feeling optimistic for the first time in a year and a half.

 

[kabin]

IMO there's a qualitative difference now, because of the NN approach. Prior to V12, what you saw is all you got. There was no guarantee it would necessarily get better; the weak points rotated around for the past year but the overall quality of the drive did not improve (for me it went from tolerable to less than that). But if the first public release of V12 is as good as it looks from Omar's vids, at there's hope now, if you believe in machine learning. I mean, what in the world could ever cause it to regress now? If we ever see it regress, the only conclusion would be that the NN is broken (and that may very well still happen, we just don't know...)

But if the NN actually works, it can only get better (more reliable, more confident, less hesitant, safer, yada yada). So that's what I'm waiting to see. I'm not running around saying "Tesla has finally solved FSD" but I'm feeling optimistic for the first time in a year and a half.

I suspect training has been FSD's weak link and v12 makes that even more challenging.

So far v12 vehicle controls seem smoother, more human-like, but there's still near misses, stop sign runs, poor planning, poor decision-making. For me it's still too soon to say v12 will be anything more than an early ADAS novelty.

We'll know TSLA made it when a significant number of customers buy it and other auto manufacturers begin to license it. Until then it's still a crowdsourced sucker bet.

 

[archae86]

Reading between the lines, I think the expectation of most here is that v12 will just be a slight improvement over v11, and people are hoping it will somehow improve faster than v11 did.

My guess differs from that. I think routine driving will be noticeably different, with quite a few of my "usual awkward and rough spots" smoothed out noticeably. I also suspect some of my "not just rough but actually a problem" spots will just go away and be good. BUT, I expect new places that were never a problem before will suddenly be a problem on this version, and stay that way until repair occurs. How all this will net out regarding either enjoyment or safety remains to be seen.

But that is just my guess.

 

[AlanSubie4Life]

My guess differs from that. I think routine driving will be noticeably different, with quite a few of my "usual awkward and rough spots" smoothed out noticeably. I also suspect some of my "not just rough but actually a problem" spots will just go away and be good. BUT, I expect new places that were never a problem before will suddenly be a problem on this version, and stay that way until repair occurs. How all this will net out regarding either enjoyment or safety remains to be seen.

But that is just my guess.

Overall that sounds like a slight improvement (though of course depends on the weight of each of the factors).

 

[diplomat33]

I am curious if V12 will fix some of the long standing issues I've been having with V11, specifically:
1) Moving over to a right turn lane at the last minute, cutting off other drivers, without using the turn signal
2) Moving over into the suicide lane before the turn only lane and then moving back into the main lane when it realizes it is not in the turn only lane
3) Stopping at railroad crossings when there is no train.

 

[Usain]

Some major déjà vu happening here reading everybody's take on the next, yet-to-be-experienced-but-promised-to-be-fire FSD version. Either this is a whole new group of posters or the posters here have short-term memories. What I have learned from driving on the FSD ADAS suite since October of 2018 (when they released NoA) is that Elon's, Omar's, Dirty Tesla's, and all the rest's comments don't mean squat - I'll believe it when I drive it.

I was never one of those posters until now. With V12, things are different.

 

[Mardak]

what in the world could ever cause it to regress now?

There are plenty of ways the neural network behavior can regress with additional training. Similar to how Tesla needed to specially train end-to-end to perform complete stops at the stop line, the "natural" training behavior was to be like the human examples. There are other behaviors that the neural network might happen to do without special training, and sometimes this can be the right or wrong behavior or just low confidence/inconsistent. If it actually was the right behavior, additional general training might actually result in the wrong behavior if it overfits and learns to follow noise instead of desired signals.

 

[Knightshade]

I was never one of those posters until now. With V12, things are different.

Why, specifically?

 

[Supcom]

From what I have seen, I'm expecting V12 to perform maneuvers more smoothly and human-like. I also expect that we will see some things that have been long coming like better road hazard avoidance, not passing school busses, not stopping in railroad crossings, better roundabouts, etc.

But, I also expect that we may see more inexplicable errors like hesitating at stop signs with clear intersections and messing up complex intersections.

Unfortunately, Omar's videos (the only current V12 source I have found) are very heavily weighted for urban streets. Since urban driving is the least driven area for me, most of his videos are not as useful for me as I would like. But, I do see some good improvements with some basic driving mechanics.

 

[gottagofast]

Have they sent v12 out to more people yet or is still the same amount of people that have had it from a few weeks ago?

 

[Supcom]

Have they sent v12 out to more people yet or is still the same amount of people that have had it from a few weeks ago?

No change.

 

[Goose66]

what in the world could ever cause it to regress now

Google "local maxima" and "overfitting." Plus, I assume you've never had a Pandora account.

 

[lzolman]

Google "local maxima" and "overfitting." Plus, I assume you've never had a Pandora account.

OK, so perhaps a singular regression is par for the course, but damn, if it acts like V11, Tesla is screwed.

 

[stopcrazypp]

Why is this topic coming up again?

It’s not a problem that FSD stops at stop signs.

The problem is mostly the going part. And to a lesser extent how it performs the stop itself.

This has been discussed at great length, with video evidence of the issue with the current stop profile.

Just totally a non issue though when it comes to v12. I have not seen any issues arise due to the vehicle stopping at the stop line.

There are lots of problems at stops, of course.

I don't think his point is about the stops, specifically rather the theory is by shoehorning "unnatural" behavior, it may cause issues in other maneuvers.

The huge caveat is this is assuming this is true E2E and not modular (which have not been established), but if it is, there is no isolated hand code that can be modified to force this behavior in a more modular way. That means it's all based on weights and biases that is established via training, and those can affect other behavior unrelated to coming to a complete stop at stop signs or red lights (for unprotected right) in an unpredictable matter.

 

[Usain]

But if the NN actually works, it can only get better (more reliable, more confident, less hesitant, safer, yada yada). So that's what I'm waiting to see. I'm not running around saying "Tesla has finally solved FSD" but I'm feeling optimistic for the first time in a year and a half.

Exactly. There is no known limit to improvement of neural networks. The limiting factor has been the need for massive compute power and quality data. It appears that Tesla now has access to both.

I'm also not saying that "Tesla has finally solved FSD". But I am saying for the first time that there appears to be a clear path to that solution.

 

[AlanSubie4Life]

I don't think his point is about the stops, specifically rather the theory is by shoehorning "unnatural" behavior, it may cause issues in other maneuvers.

The huge caveat is this is assuming this is true E2E and not modular (which have not been established), but if it is, there is no isolated hand code that can be modified to force this behavior in a more modular way. That means it's all based on weights and biases that is established via training, and those can affect other behavior unrelated to coming to a complete stop at stop signs or red lights (for unprotected right) in an unpredictable matter.

I don't really see why they can't train the correct behavior with simulation, since presumably that's what they have to do with stop signs.

Anyway, no reason you can't stop legally at a line and drive perfectly well. It will often result in some subsequent creeping but it's fine as long as you commit to it.

I don't really care what the NN training can or can't do - it's largely irrelevant to me. Kind of a curiosity rather than anything of significance.

 

[Knightshade]

Exactly. There is no known limit to improvement of neural networks.

What's the "known limit" of C++ planning code?

The limiting factor has been the need for massive compute power and quality data. It appears that Tesla now has access to both.

Server-side Teslas stated need for compute power is much higher than their current compute power.

Further, nobody...including Tesla... even knows how much compute is needed car-side. They (Tesla) have guessed incorrectly multiple times at this point, there's no reason to think they know the answer today either.

 

[AlanSubie4Life]

OK, so perhaps a singular regression is par for the course, but damn, if it acts like V11, Tesla is screwed.

Why? There is always v13...and v14, etc.. What's the rush?