FSD V12 Elon Demo

[fasteddie7]

Say most decelerate ahead is a bit presumptuous. Certainly accelerating ahead is normal behavior, but I would expect regulators to come down hard on routine early acceleration.

I made a video on this a few years back, my buddy was on autopilot and got pulled over and ticketed for slowing down at the sign instead of reaching the limit speed when reaching the sign. He attempted to fight it in court but lost, the officer and judge stated that the vehicle must be going to speed limit when reaching the sign, not slowing after.

 

[AlanSubie4Life]

is able to more or less equal it in most of its mechanics (and far exceed it in some areas) already

Can you please provide some data which led to this conclusion? Big, if true.

Don't get me wrong - I think this is a pretty cool demonstration of end-to-end capabilities. But I think we have to be realistic about actual capabilities and potential limitations.

For example: Someone would have to run statistical analysis on the likelihood of a red light being run in 20 minutes, vs. the likelihood of that happening with V11 in the same period of time. I suspect that seeing that red light issue in a single demonstration suggests that it's likely that performance in that area at least is inferior for V12 vs. V11 at the current time (but we can only say that in a statistical sense - there's no way to know for sure since we know V11 will also run red lights on occasion).

Sure, they can retrain this, but how low can they get the red-light-running rate? How low should it be?

Otherwise, in terms of driving smoothness and other aspects, it's difficult to tell from the video how smooth it was or what the limitations were. We'll have to wait for a better quality video, which may take some time.

Even the "pulling to the curb," while new, did not seem ideal as it seemed to pull over a bit too early (it said 100 feet but might have been as much as 150-200 feet early). There's also not really anything specific to end-to-end that enables this - they could do this with v11 as well, they just haven't done so.

I don't expect this to be released to the public before the end of the year, and I'd be mildly surprised if it took less than six months.

I assume they have extensive simulations (which I am sure they continue to run) showing how much regression there is with this build. It's unfortunate that Elon didn't talk about this aspect of the development process.

I'd be kind of hesitant to get this on my car since it seems much harder to predict with end-to-end when it is going to fail (it's relatively easy to predict failures and disengage with the current build). Fortunately I don't think I'll have to worry about this any time soon!

 

[Dewg]

I'm curious if Chill / Average / Assertive and MLC (Minimize Lane Changes) are part of the 300K lines of code being shed for NN. If so, do we lose those features and have no control over follow distance, lane changes for speed, etc?

 

[JulienW]

I'm curious if Chill / Average / Assertive and MLC (Minimize Lane Changes) are part of the 300K lines of code being shed for NN. If so, do we lose those features and have no control over follow distance, lane changes for speed, etc?

I would bet at least as of now it's not available since that would be "rules" added to the system. Seems it is 100% on just driving however it interprets the way to drive from its training. They may have to train 4 different versions to accommodate Chill/Average/Chill/MLC since he indicated they even have to train for the different Tesla models.

 

[Supcom]

That might be true if every country and every U.S. state had uniform rules of the road including uniform signage, lane markings etc etc. But of course, they don’t. For example, when I lived in California, the car pool lanes had defined entry and exit points and one could not cross a solid line to enter or exit. Right next door in AZ, one does cross solid white lines to enter/exit. And on portions of Phoenix freeways there are double white lines to cross instead of a single white line. Not sure how feeding videos will solve that situation and there must be dozens of other such situations just in this country, much less internationally.

So unless the car has tons of videos from every U.S. state/local jurisdiction and every “state” in every other country — and most importantly, the car knows which ”state” or other local jurisdiction rules to follow, I don’t see how this can work.

Tesla should have 'tons' of videos from every state (at least that's what the storage hardware might weigh), and certainly from every major city where there are likely to be unique traffic laws, so, unless the NN parameters are in the order of gigabytes for each jurisdiction, it very well may be feasible to have a set for each jurisdiction.

Even then, each car doesn't need the entire world's driving parameters. It only needs those necessary where it could possibly drive before it could download new parameters. If the car were imported from the US to Europe, it could easily download a new set of driving parameters. Even within the US, a car in California would not need NYC specific parameters unless it is getting close to New York.

I'm certainly not saying that's how Tesla will implement things. But, it's a possibility if there are conflicting driving rules in different locations as the car can only be in one place at a time.

 

[Supcom]

Can you please provide some data which led to this conclusion? Big, if true.

Don't get me wrong - I think this is a pretty cool demonstration of end-to-end capabilities. But I think we have to be realistic about actual capabilities and potential limitations.

For example: Someone would have to run statistical analysis on the likelihood of a red light being run in 20 minutes, vs. the likelihood of that happening with V11 in the same period of time. I suspect that seeing that red light issue in a single demonstration suggests that it's likely that performance in that area at least is inferior for V12 vs. V11 at the current time (but we can only say that in a statistical sense - there's no way to know for sure since we know V11 will also run red lights on occasion).

Sure, they can retrain this, but how low can they get the red-light-running rate? How low should it be?

Otherwise, in terms of driving smoothness and other aspects, it's difficult to tell from the video how smooth it was or what the limitations were. We'll have to wait for a better quality video, which may take some time.

Even the "pulling to the curb," while new, did not seem ideal as it seemed to pull over a bit too early (it said 100 feet but might have been as much as 150-200 feet early). There's also not really anything specific to end-to-end that enables this - they could do this with v11 as well, they just haven't done so.

I don't expect this to be released to the public before the end of the year, and I'd be mildly surprised if it took less than six months.

I assume they have extensive simulations (which I am sure they continue to run) showing how much regression there is with this build. It's unfortunate that Elon didn't talk about this aspect of the development process.

I'd be kind of hesitant to get this on my car since it seems much harder to predict with end-to-end when it is going to fail (it's relatively easy to predict failures and disengage with the current build). Fortunately I don't think I'll have to worry about this any time soon!

I agree that 12.x will likely take longer than anticipated to get to customer cars. All Tesla releases have taken longer than estimated, so why would this be any different?

What we saw was, obviously, a rough work in progress. It was a surprisingly capable work in progress and appeared to be promising, but it was clearly an early effort. Mistakes are to be expected. If the car drove perfectly, we might suspect that someone was driving from the back seat with a video game controller. Hmmmm....

 

[EVNow]

Sure, they can retrain this, but how low can they get the red-light-running rate? How low should it be?

Theoretically yes to the first question - second one probably less than 1 in a million.

Let us take a downtown scenario - about 10 lights per mile. you want no more than 1 error in 10k miles - probably 100k miles with things like running red lights. So, 1 in a million rate sounds about right.

 

[Larry Hutchinson]

Curious to know if FSD will start decelerating/accelerating ahead of a signed speed limit change, like most humans do.

In my experience FSDb (14.4) already does this. Previously, it would change speed exactly as it passes the sign. Now it does, maybe, 50 feet ahead of the sign.

 

[azred]

FSD Beta is now rolling out for HW4 cars! Hopefully it will be a wide release. So the 6 months delay for HW4 cars that Elon referenced apparently just referred to v12 implementation … or something like that

 

[UniResearcher]

What are your expectations on the v12 capabilities? I would expect it still needs close supervision and thus not reaching level 3 automation, except maybe in limited environments but is still a step forward in the development towards autonomous driving.

 

[Yelobird]

What are your expectations on the v12 capabilities? I would expect it still needs close supervision and thus not reaching level 3 automation, except maybe in limited environments but is still a step forward in the development towards autonomous driving.

Kind of early for anyone to predict based on a 30 minute video.

 

[cruiserlarry]

https://twitter.com/x/status/1695238885633314973

Elon is currently in a space with Whole Mars Catalog talking briefly about FSD.
He says he will be placing random pins around Palo Alto to demonstrate how smooth it is, but with the caveat it can go wrong and there's a reason they haven't released it to public yet.

I guess I'm finally emotionally burned out after 7 years of FSD hype, using HW1, HW2, and HW3.....I appreciate the monumental task Tesla has taken on trying to make FSD a reality, given some of the ridiculous personal limitations Elon has placed on the sensor suite it can use, and his endless tweets that never live up to the hype to this point. FSD just won't be "mind-blowing" to me until it is 100% functional WITHOUT any caveats about "things could go wrong". Until then, we have an over-controlling CEO, with a single minded focus that ignores the advice of his hand-picked experts at times, creating expectations and timelines for his loyal PAYING customers that have never wrung true. I love my car, with or without FSD, but I'm tired of the unprofessional way Elon, and by default Tesla, is handling the discussion of FSD development. And let me transfer my long-paid FSD to any Tesla I purchase anytime in the future, as many times as I'm willing to purchase a Tesla, until I am in possession of a version that is 100% feature complete, as Tesla marketed to me since 2015. That will blow my mind - a true appreciation for the guinea pigs we early adopters are, and a policy that demonstrates true respect for our patience. Rant over, flame away

 

[APotatoGod]

Tesla should have 'tons' of videos from every state (at least that's what the storage hardware might weigh), and certainly from every major city where there are likely to be unique traffic laws, so, unless the NN parameters are in the order of gigabytes for each jurisdiction, it very well may be feasible to have a set for each jurisdiction.

Even then, each car doesn't need the entire world's driving parameters. It only needs those necessary where it could possibly drive before it could download new parameters. If the car were imported from the US to Europe, it could easily download a new set of driving parameters. Even within the US, a car in California would not need NYC specific parameters unless it is getting close to New York.

I'm certainly not saying that's how Tesla will implement things. But, it's a possibility if there are conflicting driving rules in different locations as the car can only be in one place at a time.

I've recently been experimenting with the AI large language models (think ChatGPT but running locally on my device), and lately a big focus on those has been on "fine tuning", taking a published large model (often Meta's Llama), and further training it on data related to the desired subject matter (code, or conversation, or trivia, or storytelling, etc), which can drastically change how the model behaves, improving its performance in situations relevant to whatever it was fine tuned on.

I wonder if a similar concept is possible with the FSD AI, where the central model could be trained on broad road behaviors, but then fine-tuned on a smaller dataset demonstrating compliance with region-specific rules? Given that's basically how humans transfer our experience of driving to new situations when traveling, I feel like that could be a very promising option for internationalizing FSD without a bunch of retraining completely from scratch.

 

[Rick81]

Region-specific behavior should be learned by having the location as just another input to the NN. If it makes a difference, it would be taken into account automatically. That should cover road signs between different countries, left/right driving, turn on red/never turn on red, etc. while keeping general behavior such as "don't run over pedestrians" uniform.