FSD V12 Takeaways - A New Era (Elon’s Sneak Peek)
Vložit
- čas přidán 5. 06. 2024
- Consider supporting this channel on Patreon: / aidrivr
Save up to $1000 on a Tesla & get 3 months of Full Self Driving for free by using my referral: ts.la/john27670 - Věda a technologie
You’re consistently the gold standard for FSD beta content.
I really appreciate that, I know there's a lot of good FSD channels out there!
Facts
Agreed - it’s the most informative and also entertaining, thank you
@@AIDRIVR I only watch your video's on FSD because you know what your talking about and also know the technical things about FSD.
Well said.
Your videos are very easy to watch and you breakdown the breakthrough in AI driving so simply and precisely.
Was watching the livestrem and actually thinking it would be better if Elon had AI Drivr in the car to video it.
Thank you! I didn't have time to watch the full livestream. These highlights are just right, and your comments make it even more informative.
One of the 'problems' predicted with FSD was that authorities would lose revenue from issuing speeding tickets, so it's good to know that that 'problem' has also been solved. :-)
Cops need better things to do than bother people going perfectly acceptable speeds anyway. There are no traffic laws in Romania.
In Spain, most intersections use yield signs and roundabouts instead of stop signs, making for smoother, safer traffic flow.
I didn’t get that from the video? Just curious what you mean. Thx
@@mike01mike Same in the Netherlands
@@TheXshotI think same in most of Europe. As an American I'm very jealous of the city design of the Netherlands, seems nearly perfect for me! I wish we took some inspiration off of you guys for our infrastructure.
I believe at 3:40, the car thought it was in a left turn lane. That's why when the left turn light turned green later in the drive, the car tried to go even though the main light was red. I'm unsure we can attribute this move to intersection courtesy but noticed the same thing in the livestream!
Damn didn't expect to see you here
This is what i think. That the reason why the car will Drive and elon make the Intervention
@@Astra2 i mean there are teslas in his game.. gotta stay up with what's going on!
The route planner was planning to go straight, and the lane the Tesla was in can turn left OR go straight. I think it started to accelerate because the left turn green signal was a ROUND light and not a green ARROW, as it is at 99.9% of other locations. I can understand how it got confused there. If it had been a green left turn arrow, that would not have happened.
The classic philisophical question: is being more human better or worse 😂
As it stands, driving with a bunch of other humans, probably better to be more human. That being said it, humans driving with other humans is worse than a bunch of computers driving could be.
I fully expect them to nudge the program towards safer habits while still remaining fairly human-like.
Great video and discussion!! I also caught that subtle scene where the car waited at the intersection with the green light because it was trying to prevent gridlock - totally amazing!!! And as you probably noted, neither Elon nor Ashok said anything about that amazing behavior! And the move to assist the cyclist by powering ahead to clear the gap instead of causing an awkward and robotic move was another superb example ... and yeah, we just did the trade-in of our 2020 Model Y to a new Model Y (did the FSD transfer), so we have the new HW4, too - hope that v12 will run relatively soon on HW4 ... in any case, v12 is going to totally rock!!
A sucker Musk bot born every day.
There is no full self driving. Just endless Musk lies. And lies. And lies. 🎉 Drug dealers gotta park their billions somewhere in troubled times.
All I can say is wtf the way the steering wheel doesn't flick around like it used to just makes it feel so much more advanced and perfected
Great video. You have some insights that no other commentors have mentioned like not blocking the intersection and speeding up to close the gap.
Looks promising. Can't wait for it to get refined further
Same here! Can't wait to try it out for myself and really put it through it's paces
Best review of Elon's live stream I've seen (of 4).
Your commentary has always been great, doesn't matter if you are behind the wheel or not as the point is to see how FSD behaves.
Also, offline navigation with no map: i just realized how useful it can be for the cybertruck.
Why the just CyberTruck and not all Tesla's with FSD?
@@stevesimpson4695I assume because of off-road capabilities? Not sure...?
I watched the live event and have watched at least five others comment on this v12 while all expressed a similar amazement this video by far does the best at highlighting some of the more impressive features of v12. Thanks
Hey! Been subscribed for around 1 year, I just wanted to say that I appreciate your videos, did enjoy the new format, and as a rule I would say to all the CZcamsrs that I follow : don't hesitate to diversify if you feel like it! Keeps things fresh and natural.
Take care!
I'm impressed, I will be floored when it all works in snow and blizzards as well.
AI Driver, thank you for always taking the time to consider, plan, curate and narrate your videos so professionally. From my experience, you produce one of the best, if not the best, Tesla FSD coverage on the internet. Keep it up.
Regards Stephen,
Sydney Australia
A collaboration between you & Elon/Tesla for future videos like this would be so amazing.
.. and MUCH better video quality & videography!!! :)
Please keep making videos like this. Your style and commentary are top-notch. I love watching your videos. You are great at what you do. Thank You!!!
Long term tesla owners always understood FSD needs to drive a little worse / take a little more risk to get a smoother humanlike ride. Shit has been so jerky and jittery and risk averse for too long. Finally a worthwhile update
Thank you for the highlights, because seeing the video at 144p wasn't an option for me :D
Version 12 seems like a big step forward and once DOJO gets online I think it will be amazing in every way.
Speechless. This is such a huge jump. I just have nothing to really knock it for.
Mind officially blown. I missed ALL of this watching the original stream. Your best video yet (and that says a lot
imagine seeing Elon musk live streaming in his tesla that's in full self driving
I was hoping we’d see something like that during the live stream lol
Why imagine?
@@madhououinkyoma are you saying I should track Elon down and physically watch his car drive him somewhere by itself?
@@Totallynormalvideos oh you mean imagine seeing him in real life. The livestream is us seeing him so…
@@madhououinkyoma Nah I'm talking about you walking down the street and you just see Elon in a Tesla that's driving his ass to idk like rich people Walmart
I briefly watched the live stream, but got bored with it. This is a perfect summary of the highlights. Thanks for putting it together!
Huge fan of the channel, and a huge fan of this video. Maybe do like 80-90% fsd driving videos and then 10-20% “fsd news etc” videos like this? This video rocked.
Yeah I second this! I would love a weekly, or even monthly, rundown on the state of Autonimous driving in general, not just FSD.
Thumbs down. This video felt like it was written by Elon's PR team. Makes me wonder about objectivity, or if money is changing hands. Much prefer critical analysis of real world driving. Will unsub if vids like this keep showing up.
@@rustyshackle917 Okay Karen
not the first to put a video out, but he took his time to make the BEST video out !
Artificial general intelligence is coming!!! Much quicker than anyone had imagined or predicted
Jeff Dean forecasted 2027 or 2029, can't remember exactly, a few years back.
@@FlorentHenry like two years before that a "generous" estimate was 2120 and now we're like "yeah it's going to be in like a year"
@@aonodensetsu I do not think it will be next year.
By what methods do you guys think this will happen? Aren't existing AI neural networks already running into bottlenecks of scale? Both the amount of good quality input data and the mass-human-quality control of lets say Chatgpt are already at it's limit and chatgpt is still far away from general intelligence
@@leonox7313 there is still a ton of non-text data that chatGPT hasn't tapped into. Google is training a new model that will use image and video data as well
Very useful! Short and straight to the point! Good job!
Definitely should be Yield signs in most cases you have Stop signs., In Europe stop signs are usually reserved for extra risky intersections with low visibility or a strong precedence for the crossing traffic.
the problem with stop signs is that most people on the road are idiots who want to have it their way like little children, thats why stop signs are useless
A lot of people use the stop sign as a yield already...
Thank you for the high quality content you always deliver ❤
What improvements could this neural network change bring to Autopilot behavior moving forward? It's going to be interesting to see this become much more enticing to future buyers as well
I think there should be a way to give FSD rules -- like a list of regulations to follow. Maybe including videos of human drivers following the rules, and then giving those videos a greater weight in the neural net.
They said in the stream that they only use clips of good drivers, and that they can selectively train it on specific clips of problem areas. So if it isn’t doing what it should/what they want they have a solution in place
You've completely lost your mind. But, I still appreciate your production value.
I just love your content. Keep making that good stuff for us.
Great content. Keep it up, it's always interesting and informative.
Probably one of the most useful videos in your channel, as it offers a new perspective. Thanks!
Much needed video! Absolutely loved the breakdown
Love this kind of video! Can't wait for more updates on self-driving technology. Thanks for sharing. Subscribed! 😊
As many have said, your videos set the standard for FSD vids. Thanks so much!
As feedback, I love hearing your reflections on the evolution of our laws and society's views and beliefs about self-driving cars. We are in the middle of a profound change on the planet, and most don't see it coming. Please keep shouting from the rooftops so others begin to anticipate the world in 2027 and 2030.
Great recap, couldn't see any of this stuff from watching it live, a huge step forward!
Far better and simpler explanation of the tech behind it than any of the other comment vids I have seen about this. Which was not surprising. Keep up the excellent work! BTW I am in the EU and like Musk himself, I have a facelift Model S with FSD on HW3. Can't wait for the beta (or who knows release software!) to become available here, hopefully some time next year
Loved the video! Thank you :)
Very impressive self-driving, can’t wait to see updates
Thanks for making this video. Great to hear your perspective and see the bullet point summary at the end. Keep up the great work!
Stop signs are a good example of a traffic control designed to ensure people don't rashly make decisions about going through intersections without properly looking and to give a defined order for who goes when. Full self driving cars don't need that, they are fully attentive 100% of the time. Consequently stop signs should be considered as information to FSD cars and not a revenue generating mechanism for law enforcement. But you can be sure FSD cars will be beaten up for it and fine for it too.
I sincerely hope they're training models unique for each country. Here in Australia, the speed limit is very strict and if you speed more than about 5kph over you're at high risk of being booked.
First time on the channel. Great job and great analysis. Your explanation of the pros and cons of AI based FSD logic was succinct and helpful.
So helpful to see your breakdown.
Hope there’s clips of slamming to a full stop when there’s a cop watching the intersection.
The off the cuff video seemed great. Grreat info and summary. Made sense the whole time. Please keep it up!
There has to be vehicle model specific training model because of the small differences in video capture, which is part of the reason they have issues with HW4
I did notice the keeping the intersection clear, though after it may just been confused on the light.
I am very happy to see more dynamic speed and follow distance. Hopefully speed is also remembered on how fast adjacent lanes are going on highway when changing lanes.
I hope there are more multi-lane roundabouts it can do. Seems most are single lane but multi-lane are more common. It's almost like a virus how they are installed today.
V11 was already very smooth on acceleration and braking. Clearly some of which was already on nets.
The most simple solution to over use of stop signs is basically to convert every stop sign to a yield sign, and reserve some stop signs for the more complex of exits/entrances.
Excellent video. It was really great to see version 12 in action.
A side note on the stop signs. I'm a 38 year old swede living in Sweden and i think ive seen more stop signs in the about 10 weeks of vacation ive had total in the US (multiple trips) than ive seen in the whole rest of my life.
Here, if you ever see one, you definitely stop, but they are so rare that they are almost endangered.
Love your last two videos, especially this.
Saying people who do the speed limit are not only a nusance but a danger and then following it by showing the beta almost running a red light; pure poetry
This is fantastic! Thanks for covering it
Well summarized. Thank you!
I was watching the live stream and getting bored and then Elon said the "this is probably better sped up and cut" thing and I had a hunch you'd be the guy that'd do it and do it well. ya never disappoint dude, great video.
Thanks AIDRVR for posting your video of Elon's FSD Beta 12 test drive. This video is a must watch supplement to Elon's video. Good job in summarizing the capabilities of V12.
First time watcher. Didn’t have high hopes in terms of quality. VERY pleasantly surprised. Well done, my friend.
Well done! Itemized summary was excellent!
This is the kind of different perspective/approach I was talking about in your last video. Glad to see Tesla took a completely different approach of implementing FSD.
Great analysis! I love my new HW4 but miss the self driving!
Great vid, keep ‘em up!
Where I live in the country road speed limits are followed by most and not following it is the exception I think the car shouldn’t always speed on an interstate but make sure it’s flowing with traffic and what is safe. And what the passengers in the vehicle prefer
Wow! That is amazing. That makes me so excited for the future of AI driving.
"Skynet begins to learn rapidly and eventually becomes self-aware at 2:14 a.m., EDT, on August 29, 1997."
Excellent review, thanks!
Thanks Buddy, perfect timing for your video.
V12 Incredible! And such a great summary of a very hard video to watch.
Thanks
Brilliant video. Loved hearing your insights, as always.
Great video, love this channel.
My God, you are so right about stop signs.
Loved the analysis on this! Keep it up bro
Perfect! This was exactly what I needed to get an understanding about what was said. Well done!
Amazing editing, pretty good topics you noticed!
Appreciate it, editing the insanely low quality twitter stream into a video hurt my soul a little
Really excellent summary, worth every minute of your time, thank you!
Keep doing your videos! I am sure many more than I appreciate your work. Thank You!
Great explanation of how it uses only video clips to learn. So all it now needs is enough video of perfect driving. But, in the real world there's always edge cases, so we may never get above level 3. But true level 3, hands off, able to sleep in the car with significantly less accidents would be a game changer.
Can't wait to see your V12 video in the next few days 🙏🏼
I'm excited about the implications of this approach for general AI as well.
Really good video, seriously. Appreciate it.
Best explanation of FSD 12 ever!
Great summary! Thank you!
On the red stop discussion - if you playback the video the car did went all the way down to 4mph and it only continue because the intersection was clear - if Tesla can provide NHTSA with sufficient data the car is able to do this maneuver safely I think they will approve it
You have more faith in the bureaucracy than I. NHTSA has tons of data from human driven cars already.
Nice summation of the drive. Agree this is the real deal ...
More of this content is always appreciated
Great recap! I wonder how local laws (like no right-on-red red in NYC) will work.
Seems like extra data like rough gps coords needs to be passed into the network's input along with video.. how do you solve this?
Thanks for the great summary!
I'm surprised it took them this long to get to/try end-to-end AI.
I guess because they figured they didn't have the compute power both offline and online. Still surprised though, I would have tried it as soon as possible. But it's easy saying that sitting on the sidelines, there's a lot of other stuff like the financial side, hardware supply/capabilities, humans being humans etc. behind the scenes.
There’s also been a lot of progress on AI in general, and information learned throughout the process of developing the older versions of FSD. Although the founder of Comma AI has said for years that end to end AI is the way to go, so have to give props to him for trying even with Comma’s significantly smaller resources
I thought that same thing about waiting for the traffic at the intersection, that's crazy
yeah totally. I've made the mistake of risking going forward in that situation once or twice. Sucks when it doesn't move forward and the light turns red lol. People get pissed haha, and the dummies in front of you have no regard for people behind them, and leave huge gaps between them and the car in front of them. Like, squeeze in people lol.
Awesome stuff!
I was fortunate to grab a 2021 model s with FSD! And it’s really cool
Don't know what you were doing before but this video was just the summary I needed, thank you. Practical catch-up.
Excellently put together.
Great video, definitely recommend doing more of these as well
Great video. Getting excited again about TSLA
This the video I was searching for. Nice job!
Fascinating video, thank you for sharing! I wonder how similar this is to MIT's VISTA that uses 19 neuron-like modules with Reinforcement Learning (RL) to train a self driving car on a video simulator.
There is one particular problem with training end-to-end NN to "to what a [good] human driver would do" -- from _only_ that, it cannot learn what to _not_ do. If good human drivers drive slower in fog or rain, for example, it can mimic that slower driving, but it does not know _why._ It does not have the underlying "fear" or "uncertainty" that its visibility is more limited. So if it gets into a novel situation of limited visibility not due to any cause it has been trained on, it could conceivable just continue on at full speed without caution. The car also doesn't learn, to put it simply, that "crashing into things _hurts."_ Now, I think (hope) Tesla is dealing with these issues in a variety of ways. They can also train it in the driving simulator with lots of near and actual accidents, to teach it, in effect, that all forms of accident "hurt" and it should never allow an accident. A little more difficult to get it to generalize on "if there is important stuff I cannot see or cannot see clearly, be more cautious" ("Raining therefore slow down" is not the same thing).
Another area of difficult training is anything involving understanding of what other drivers & pedestrians are thinking/experiencing, like pulling further forward to make room the the car behind you. If it only learns these behaviors by example, it is a case by case basis. It needs a model of "what does that other driver want to to" or "what will annoy the people around me" in order to truly generalize. I think there are things they can do in the simulator to also teach it these things. But are they doing that yet? Not sure. Maybe the could put human drivers in the simulator in lots of situations _around_ FSD beta, and then they could give feedback indicating how annoyed or uncomfortable they are with what the FSD car(s) are doing in the simulation. They should then not only get NN to do control outputs, but to also an output that predicts the "annoyance" of other drivers & pedestrians. This would force it to generalize better in these situations.
I'm sure they are showing clips where unexpected things happen with little time to react. Not quite "fear" but as close as a computer brain can get?
It's shown that by also training the network on negative examples it is capable of building what is bad and avoid it
Like the "negative prompt" in image generation models
@@AIDRIVR The trouble is real world clips of close calls are going to be rare. Yeah, they can get quite a few from the fleet, but like 99.9999% of fleet activity is "normal bland driving" and 0.0001% is "close call", or something like that. Also, humans likely were not making the best choices leading up to the close calls and may not react to the close call in an optimal way, so learning "what a human would do" from that I think is going to be troublesome. Simulator is best, where they can manufacture close calls. And for generalization, it would be even better if they simulated lots of weird even impossible situations. Also for vision to get deepest generalization, they should use simulator and image processing to screw with the visuals. Make it learn drive as correctly/safely as possible when the video input is just a line drawing, or Van Gogh style, lots of random static in the image, and so on and so on. This would force limited numbers of neurons to generalize in the most relevant and flexible ways.
Honestly the way I would train and end-to-end NN would be to start in the simulator and like an infant, let it run around an learn to crashing into things "hurts", learn that driving of the road (on lawns, etc.) is very "uncomfortable", driving wrong way in a lane is very uncomfortable, and so on. The other thing is a foundational model that just predicts the future of video. With that deep foundation of prediction and "avoid pain", then start the serious "road training". The biggest difficulty here is understanding some of the bigger "whys" of human driving, which revolve around human comfort, non-immediate goals, and annoyance of other people around you. March of 9's -- the rarest edge cases are hardest to "train" in brute force interpolative NN model. The deeper it can generalize the better.
@@AIDRIVR Oh: very short answer though. There's no reason the computer brain can't have "fear" programmed in. It would just be a term in the loss function you want to minimize. Likewise pain. It should be able to project things like "if I keep going on this path & speed, I will crash with that oncoming car, which , avoid at all costs (except even bigger pain)." It should know that the "discomfort" of driving on someone's lawn is much better than colliding with a car or running over a child. Pain and discomfort are just terms in the loss function. But it cannot learn what is painful and uncomfortable unless it is trained in a simulator and can make those mistakes. (We don't want it learning those things in the real world, lol!!!)
thank you for not giving us clickbait and fill it with useless filler content like so many tesla CZcamsrs, got so tired of watching their time wasting videos I just skipped all Tesla recommendations