This new aspect ratio (2:1) should look better on a phone. Do you like it? Sponsored: Use my special link www.piavpn.com/AIDRIVR for an 83% discount, 4 months free and 30-day money-back guarantee on the best VPN - Private Internet Access
It stopped for the gate In my community to scan even when It was open so It didn’t close on us and moved over a lane on a 4 lane road because of a biker and smoothly changed back for the turn ahead. And the 6 roundabout’s in my area It does way faster with much more confidence and then handled the gate to my community. I can’t make this up
When FSD Beta first came out, I was obsessed with it. I used to watch every video every day. It used to be a thrilling experience to watch, never knowing when it would make a mistake, rooting for it to complete every maneuver. Fast forward a few years and it got very boring since interventions became so rare. RU-vidrs that just post driving videos seemed like a waste of time. But you are the only one that I've never stopped watching! No boring stuff here. You make comparisons, you edit out the fluff, and your commentary is always smart and on point. I can't think of anyone better to be a tester for this technology. Congrats!
Agreed. This morning here in SoCal, I clicked on FSD in my S, put on a podcast and 30 mile laster took over to finish parking at the golf course. It is going to become standard.
I agree. I watched his videos for a while on 12.4. Then when 12.5 came out I forgot his username, so I just randomly searched for FSD 12.5 videos and no one else came close. Then i chanced upon this channel again and am so glad I did!
Exactly! The well written script and smooth narration makes it worth it. Almost as smooth as 12.5 itself :D Btw I'd love if you check out my Autopilot videos from Europe on this channel and give feedback :)
Been using 12.5 for a couple of days now and it's really good now. There really aren't many more key improvements needed to be Robotaxi ready but here is my top 5 list. 1) Respond to hand gestures by police/road crews 2) Respond to emergency vehicles. In my state that requires you pull over regardless of the direction of the emergency vehicles. 3) Pot holes. Probably my larges percentage of disengagements are for pot holes 4) stop for school busses with flashing lights 5) School zones
I agree with AI Driver that this is the best version so far. Just went to lunch 16 miles away 22 minutes driving each way. FSD 12.5 was on from parking to parking with mixture of highway and local traffic, only one accelerator push (personal choice). And it was smooth
@@catbert7I mean yes but if you’ve been following this channel it’s pretty clear that some updates tend to improve certain aspects while making others slightly worse, while this upgrade has seemingly just made the car better in all situations and substantially so
I asked Claude 3.5 Sonnet to count how many times you mention smoothness. "Smooth" appears 11 times. "Smoothness" appears 10 times. So pretty smooth I guess!
FSD 12.5 is truly amazing. But, I am commenting to express my appreciation for the great way that you talk us through the demonstration. I feel like I learn a great deal from your FSD videos, unlike many others that lack the thoughtful commentary that you provide. Plus, you do it in a very friendly fashion. You are very easy to listen to. Thanks.
@@HJeff Some people like to watch humans chase a ball around a field, I like to watch robots drive cars around cities. There's no accounting for taste, I suppose 😅
Basically 5x in Parameters means it can learn more subtle differences in scenarios. More parameters just means more attention to detail. Doesn't mean it gets higher resolution input or something... just means it can learn more things. (This comes at a cost Tho, as it also needs much more training data)
Not really... Musk said they have more than enough data - what is kinda expected. The issue is with computing power to train the model. If you put the same data to 5x smaller model, it'll generalize too much, miss some details. So you're right here. But it doesn't mean we have to feed more data into a bigger model. It makes more sense to feed more data, but it's not required.
It doesn't need more training data. If we have two models one with 1x parameters and the other one with 5x parameters and we use the same training data, the model with 5x parameters will outperform the model with 1x parameters! But the 5x parameter version benefits more from more data.
If HW4 is required for full-scale FSD, Tesla should offer that to HW3 customers (who have bought FSD) to retrofit for free or at least for a very small cost. Should also be included in the purchase of FSD for the people with HW3 who never went for the FSD option before.
I understand how you feel. I was an early adopter with my first Tesla in 2012. Many iterations had occurred after that and then I bought another one in 2017. Many iterations occurred after that and then I bought another one in 2023. It’s not reasonable to expect the company to upgrade hardware just like it’s not reasonable for them to upgrade seats or chassis and stuff like that. Nonetheless, I know exactly how you feel!
The issue is HW4 means different cameras, computer takes more space, and I think some cameras are in different places. So pretty much no way they are going to do that. But... after some optimization v13 on HW3 should be better than 12.5 on HW4. At least I think so... So even though versions for HW3 will be worse, they might be good enough.
@@rvboondocker2559 My main concern honestly is the huge hit to Tesla's credibility and the legal hole they'd be in if they've promised fulls FSD capabilities for HW3 cars and then having to walk back on that promise. Tesla is aware of this, of course, and I'm sure they'll do all they can to make sure it doesn't have to be that way.
A few months ago they said that HW3 will be capable of full scale FSD. The difference between HW3 and HW4 in the end product will be safety margin. HW3 might achieve 99.9999% safety while HW4 could reach 99.99999% safety (extra 9, meaning 10x safer). But all versions will be safer than humans at minimum.
13:20 To make it even smoother when braking it should have a range not a fixed gap. You can safely give up some following distance to keep your momentum.
no you shouldnt unless you kept extra distance before. if you do that and then the car in front of you suddenly starts braking all in, thats how many accidents happen
@@sirati9770as you both slow down, the braking-distance-to-stop also gets shorter, so yes, it's okay to give up some distqnce, as long as margin for emergency is maintained. The condition that there should be enough at the start with some leftover is a given in any case. What the video showed was frankly a bit too cautious, even if the reaction time was, itself, an impressive demonstration. It didn't need to match the truck mph-for-mph, in good traction conditions and good visibility. That said, I'm all for somewhat defensive driving, so I don't consider what it did wrong either. Just not as smooth as it could have been.
Feeling relaxed and safe is key! Not just being safe! This is how my dad taught me to drive, we had a closed container of water in the smooth floored estate boot (trunk). If it slid about I wasn't driving smoothly enough! I am now teaching my kids the same way! Safe and smooth makes passengers confident in your driving abilities. Great stuff Tesla!
Huge problem with 12.5. Twice, within 2days, 12.5 turned in the opposite direction from what the nav path directed! Crazy. Once, right, the other time, left. Both turns ignored nav and wandered off. I let the first one run, and it wandered around a parking garage lot and finally found its way out to continue the route. On the 2nd one, I disengaged because there was traffic. Both times, I was not prompted for a report. Terrible. I’ve been driving 12.3.6 for a while, 11.3 before that, and do on. Never saw this once before.
Yes, that was embarrassing in this video as well, and could turn dangerous. Hopefully someone at tesla is watching these videos and taking notes. They are short and well put together, and provide good feedback, the developers should use that and not rely only on automated reporting
@aidriver Man your graphics are top notch! I really appreciate the time you take to make them. Your commentary and video content is the best! I got 12.5 last night so far only 3 drives but smooth as silk and no interventions. Amazing. I believe this will truly be better than a human driver in most scenarios in the future
Still with Tesla stock, investors can never be sure what will happen next, bearish periods ultimately establish a new set of stocks to buy and watch while setting the stage for a robust new uptrend. I have been reading articles of people that grossed profits up to $250k during this crash, what are the best stocks to buy now or put on a watchlist?
Investing without proper guidance can lead to mistakes and losses. I've learned this from my own experience.If you're new to investing or don't have much time, it's best to get advice from an expert.
Even with the right technique and assets some investors would still make more than others, as an investor, you should’ve known that by now, nothing beats experience and that’s final, personally I had to reach out to a market analyst for guidance which is how I was able to grow my account close to a million, withdraw my profit right before the correction and now I’m buying again.
How can I participate in this? I sincerely aspire to establish a secure financial future and am eager to participate. Who is the driving force behind your success?
My CFA ’ANGELA LYNN SCHILLING’ , a renowned figure in her line of work. I recommend researching her credentials further. She has many years of experience and is a valuable resource for anyone looking to navigate the financial market.
Thank you for this tip. it was easy to find your coach. Did my due diligence on her before scheduling a phone call with her. She seems proficient considering her résumé.
Completely agree this is a giant leap. I have a square in the middle of my small town that each time I would have to intervene for every drive. Now it is completely fixed and smooth and flows with traffic effortlessly. Magical
I want the FSD to be perfected before my family decides to put me in the nursing home. That's another 25 years from now. The day before the nursing home schedule, I'm going to jump into my Tesla and scad doodle into the sunset (Vegas strip).😂
At this rate, it's 2025 or 2026 before we see robotaxi-level usability. I think it'll be somewhere around March/April 2026, but it won't need anywhere near 25 years, even 10..
@@shannon6876 Cruise was operating robotaxis with an intervention every few minutes. Tesla is way beyond robotaxi level already, in many parts of the U.S. They just need to set up remote drivers to get it unstuck.
Great video. Very interesting I have 12.4.3, and my car stopped for the flashing red light and continued on after a car crossed on the right side. It treated it like a stop sign. It did this in Palos Verdes california.
Good job on finding interesting scenarios and especially conveying the smoothness of this release. 🙏 I would suggest doing some really heavy stress tests with 12.5 and I think we would all like to see it pushed to its absolute limits! Like this comment if you agree 👍
Amazingly well done video! One thing seems clear, HW3 isn't going to be powerful enough, but HW4 might actually be. For the first time, it seems like the goal of a functional Robi Taxi is actually within sight for Tesla.
It is technically true, but the reason why there is a difference is due to the number of pixels between the HW3, and 4 version. HW4 have more pixels, so more power is needed to be processed. They have to be trained separately due to that difference. Elon said that HW3 processing power IS STILL NOT MAXED OuT, and they can still optimize it more. Until it is maxed out in both processing power, and optimize the parameters, there is STILL room for improvements. Sooo, for now, YOU ARE WRONG in that HW3 still can be improved and is powerful enough up to now and in the NEAR FUTURE.
@@1flash3571 I’m a software developer that works with neural networks. When working with inference models there is a very predictable pattern of increased intelligence the more parameters you have. That’s why a 400B parameter LLM is much more accurate than 4 billion parameter model. For comparison, the human brain is estimated to be about a 10 trillion parameter equivalent. Past a certain point, it’s no longer about pixel FPS processing, but the number of connections the neural network has to draw from, which is a matter of having a vast amount of memory in addition to processing power. Keep in mind also that even this latest 12.5 version still doesn’t have the ability to understand language, which will be needed for a universal robotaxi. That’s around 7 billion parameters at minimum to accurately understand various types of unexpected signage you can run into, as well as potentially understanding first responders who will need to communicate to the vehicle in an emergency. At minimum, that will require the power of a state of the art smartphone all by itself. Keep in mind that Tesla is currently working on AI5, indicating that they already understand that even hardware 4 probably won’t be powerful enough for a Robo Taxi. At this point, I think there’s a chance at HW 4 might be able to do it, but I’m confident that AI 5 will be able to achieve full autonomy, and therefore I’m also confident that even with further optimization HW3 will not be able to make it past the current supervised model.
How should I feel about a self-driving car being pulled over doing something illegal? 14:11 The car illegally goes around the bus, doesn't put on its blinker to indicate what its intentions are and changes lanes when it shouldn't in an intersection. Granted, this is exactly what I'd do in that bus scenario as well, however, a police officer could pull you over for something like this. Just curious how this whole thing would/should play out some day when it inevitably happens.
yeah we are in the future. when v12 released mind was already blown and i didn’t even think it was possible to be blown away again but here we are. i can’t even imagine where this technology will be in just the next few months, we are in for a treat.
6:58 FSD actually plans now paths backwards. In case of the car around the corner, the decision was flickering shortly to the left rear. So FSD wanted to back up to the right side with the back of the car, to make room for the other car. But it's still forbidden for FSD to drive backwards
Consider that hw4 cameras have higher resolution. Hw5 cameras have variable aperture. Even more processing power will be needed than hw4! Luckily Tesla is already onto it and will have HW5/AI5 ready next year.
Tesla can just use the extra pixel count and disable the variable aperture so they don't need to upgrade to HW5, or they can add that function if needed in the future.
@1flash3571 aperture shouldn't require much more compute. Disabling pixel count will seriously affect the full self driving capability. HW3 is dead, accept it. Being stuck with a iPhone 3 would be horrible. More advanced driving will require AI6, which is the reality of technology.
This version is hugely impressive. It really shows how much drivers communicate with each other by subtle acceleration and braking -- and FSD understands that!
In rural areas you really want to fully stop as a deer crosses as 50/50 another is about to come up behind and cause your insurance rates go through the roof…
Ok so ML algorithms are essentially just large matrices that get computed in specific ways. The values in the matrices are the parameters. The parameters are what are tweaked automatically (mathematically) during the training phase. More parameters doesnt always means better because problems can pop up like overfitting.
@@tiloalo FSD an sich aber diese Version (und alles was noch kommt) ganz besonders. Sieht reif für die Straße aus. Ich hab mein M3 seit 2021. mit oder ohne FSD ist es einfach ein tolles Auto. Ich glaube an die Robotaxi-Entwicklung und habe mir daher damals schon die Option dazu bestellt.
It is very unlikely that this will come to europe anytime soon. The EU and german regulators kind of hate Tesla or Elon specifically, and will do anything to try and make them fail.
@@HNedel Well there was the news a few months back that some Swedish EU regulator got an FSD test ride in Germany and was positively surprised. So they’re clearly working on it. I assume however that they will be even more strict than NHTSA in terms of implementation. Regarding Tesla hate, it’s mostly the media honestly. The Brandenburg government for example has always been very helpful with Giga Berlin.
@@HNedel I mean, the Tesla cannot even park correctly or just keep a line on the highway without having some ghost break occuring it seems. This is something all others have figured out and that in the EU too. I don't need full self driving but at least a reliable lane keeping and speed keeping
Wow, thanks for sharing a drive through my childhood neighborhood of Berkeley and Kensington! Timeless memories combined with cutting edge tech. Can't wait for this update myself, although Brooklyn seems to be far more complex a city to navigate.
I wonder how it would deal with Irish roads... grass in the middle, the occasional cow, sheep or deer. Not to mention the reckless young lads in massive tractors on tiny roads! Good luck Tesla!
The hard braking truck was impressive. Does the inverse hold? I notice with 12.4 a lot of times when accelerating from a stop at an intersection, often the car accelerates a bit too fast for the car in front, then it slows down but it slows down too much then falling behind the car in front. Would love to see it just match the acceleration curve of the car in front.
I think the weird lane selection is due to bad map data. I don't think they'll ever solve bad map data as roads are constantly changing, but once they solve the ability to drive smoothly with bad map data, I honestly believe FSD is pretty much solved. It does everything else almost perfectly (am 12.5 driver as well). As for that stop light with the person waving through, I think it went through the red flashing lights because the truck went through it. I think 12.5 is much better at using "context clues" when it encounters a situation it doesn't understand. If other cars a moving, it will use that as "data" and assume it can do the same.
When the car itself report map problems so the next one gets it right, then we will have as perfect maps as possible. I guess that's not a priority right now, but it will be! I guess Robotaxi will learn about congestions as they happens too, and make optimal choices.
@@1flash3571 It uses map data to "prepare" a turn. Like 0.5 miles out, if the map data says the right lane is ending, it'll move to the left lane. But obviously if that data is wrong, it will move into the wrong lanes.
While Tesla was compute-constrained, there was no reason at all to increase parameter count - because as you mentioned, parameter increases substantially increase compute (training) requirements. But now with a lot more compute online, increasing compute for such tasks is the natural path forward.
FSD 12.5 is a game changer. Multiple complex trips in a row, zero manual control takeovers needed. Its so good you forget your a driver and go into passenger mode, which could be dangerous.
I just got 12.5 on my 2023 MSP, like 2 days ago. Omg been driving it everywhere. Cuts close to curbs on left turns but holy smokes, a huge improvement in every way possible!
Running yellow light: According the to traffic rules in Norway you @AIDRIVR and FSD 12.5 are absolutely right. Here in Norway there are a distinct rule down to the cm (or inch) - if your right or wrong when crossing a light: If the front wheels has crossed the stop line while it is still yellow - you might even preceed over the rest of the cross (getting out of the crossing) on red light.
I learned at my driver licence test never to brake suddenly for yellow lights, as that could cause an accident from a car behind, so the Tesla is driving exactly what the police here in Denmark considers correct
@@skillfulfighter23 Good joke here in Denmark, i rarely see cars keeping the two seconds, and often have my gap filled with cars that think i keep too much distance on highways, without realizing that two seconds is the minumum to be able to stop without a major accident...
Tesla needs to test its FSD in India and Africa, where road edges and sign posts are often non-existent ; the road is randomly occupied by people, objects and animals, and the road is often packed bumper to bumper with erratic drivers, scooters and downright unexpected behaviour, such as scooters or cars driving against traffic, or drivers and scooters crossing multiple lanes abruptly without using blinkers. As crazy as this is, experienced drivers learn to navigate this context generally unscathed, which means that AI can be trained to deal with it. When it's able to do that, it should handle anything in the world, with aplomb. What makes me feel confident this will happen, is that Tesla's lead engineers for FSD are Indian, and therefore they know this better than anyone else.
Will be interesting to see if they can train multiple personalities into the system and have it choose appropriately. There's a lot of behavioural variance needed between Sleepy Small Town, Cop Car in the Rearview, and Rush Hour in New Dehli. Then again, there's also a lot of value for other human drivers in being able to predict how another car will behave. I know I expect rather different types of driving between a Corolla, a Mustang, and a Sprinter. Seeing a Tesla with a robotaxi badge and knowing it's going to be all-limo, all the time is helpful.
Tesla may offer insurance on own vehicles and disrupt the insurance industry. Tesla has the ultimate data that insurance needs too. This could be a sales point for Tesla vehicles, and EV proliferations. Yes, insurance will be way lower. Most accidents are due to human errors, and when human factor is taken out of driving, insurance needs is greatly reduced. Overtime, it does kill insurance as it does to Uber.
@@AdrianMeredith So why go there, because what he just said was serious stuff, we’re talking about saving lives, and what Elon and the folks at Tesla has realized, is something, the other guys don’t seem to understand, and it is the exact comment you just made, that you still don’t get it, as what price would you put on your life, and the lives of your loved ones? How do you equate life with money? So when someone loses their life, to what is a preventable incident, and noticed that I didn’t say an accident, because if you crash your vehicle into someone else, the prosecutors and the plaintiffs, aren’t going to want to hear the word accident, they’re gonna call it what it is, Negligence, because if you fatally harm someone, they gonna call it vehicular homicide, because that’s what it is. So when the software/AI has the ability to do all the things that a car with the capability of a human has to do, we’re going to solve the driving while drunk problem, driving while you’re under the influence of any kind of drug whether it’s legal or illegal, driving while you’re fatigued, and that is a huge deal, because at that point, who in the right mind would purchase a vehicle that is inferior in every way in terms of safety of operation to a Tesla that has full autonomous driving capability? So I think that’s a no-brainer because I’m not gonna put my kids/grandkids in a car that I consider to be inferior/unsafe to be in. So while the so-called competition is out there doing whatever the hell it is they’re doing, Tesla is making the kind of improvements that has got to be scaring the living hell out of them, but they really shouldn’t be in my opinion, because when Tesla has crossed that Rubicon, it’s gonna be great for everyone. So I say keep doing what you’re doing, because the realization that when you, and I drive our Teslas cars, that we’re more statistically likely going to be coming home to our families, then the competition if you can call them that would be when people get into their cars, because future Teslas cars will not even have steering wheels or pedals, because those things are not needed as the AI is driving, and not you and I, and so effectively we will have cars that you, and I can’t crash, because we’re not going to be driving them. So who cares what the insurance companies do or don’t do, because if I can avoid being in a crash, and again I didn’t use the word accident, because there are no accidents, it is human negligence that causes all of these fatalities on our nations highways, and the last time I checked the number was north of 30,000 fatalities on average, every single year, and that is patently unacceptable, and finally we have a company that’s doing something about it, and that company my friends is Tesla, and Full Disclosure, it is the primary reason why I am driving a Tesla, and not a Buick or Toyota or Mercedes or Audi or BMW or Volkswagen, etc., and so I expect future cars, no matter the make or model or brand of automobile, will have Tesla technology inside. Cheers 🥂
@@zhli4238Tesla already does offer their own insurance in certain states in the US. It is cheaper than quotes I was getting from other insurers, at least when I picked up my car.
@@AdrianMeredithInsurance makes money from better safety too, that's why the IIHS has the strictest, most comprehensive safety and crash testing compared to government testing. Injury costs far more than a totaled car, death is even worse, no injuries and no car crashes is even better. When insurance promotes FSD in some way, that's when it'll be fully accepted.
Great video ! Having used the 12.5 version, it still does the same things that the 11.9.x did : still misses to keep straight instead of moving to the right in a Y shape highway split. And it still has problems exiting the highway when the exit lane is short. Oh well, let's hope it will improve.
It still uses the separate 11.x highway stack. It is supposed to switch to the unified v12 stack in the next iterations. Hopefully that should fix all highway quirks.
I really enjoy your commentary, your voice conveys such a positive attitude on FSD. Damn, I wish I used my car more often but almost 82 years old now and my driving is pretty much when I stock up on food supplies every few days. Can't justify 90/mo! Of course, I suppose I could start playing golf again........oh bother, decisions, decisions.
I've been using FSD 12.5 for a while now and have found it to be a big improvement over previous versions. It's much smoother and more confident. However, I do have one complaint: it almost always drives under the speed limit. I have my speed limit set to +2 mph, but 12.5 seems to consistently drive below the posted speed limit. This can be pretty annoying, especially when other drivers are going much faster.
Complex urban and suburban scenarios notwithstanding, FSD 12.5 does need some work on simple two-way secondary roads. I took my 2024 Model S with hardware 4 on a 180 mile round trip yesterday. Most of that drive occurred on a beautiful winding secondary road in the Adirondacks. There are no traffic lights or stop signs on this route, and there was very little traffic. The speed limit varied from 45 to 55 mph. In my autopilot settings I have a +10% speed offset. FSD obeyed my offset and slowed to 50 mph when entering 45 mph zones, but it consistently failed to return to 60 mph when the speed limit increased to 55 mph. Instead, it insisted on staying at 53 mph, even on long straightaways with no one in sight. When I nudged the accelerator to bring the car up to 60 mph, FSD would maintain that speed for a minute or so, and then it would fall back to the low fifties again. This behavior did not occur on Interstate portions of my drive. FSD maintained 72 mph on a 65 mph divided highway without any problem.
Has speed control gotten any better for you? Ever since the very first 12.x series speed control has been a mess for me. It speeds up and slows down for no discernable reason. Often sitting well below the speed limit for periods unless I tap the accelerator at which point it's like "Oh right I should go faster" and then it speeds up and holds a higher speed limit. This happens both in the new "auto" max as well as the traditional % max. It also does not slow down when entering a slower speed zone, again either on auto or %. I have a city government complex on my daily route where the speed drops from 40 to 30 and Ego pays zero attention to it. It marks it on the screen and keeps doing 40-45 depending how it feels. Even stranger if you stop in this zone, when it resumes it will go right back to the 40+ instead of the 30. So not only does it not slow down, it just seems to ignore it all together! Yet it does show the speed limit correctly of 30 in the UI. It has been like this ever since 12 came out. 11.x was perfect on speed control.
12.5 is generally better with speed consistency (not a lot of speeding up and slowing down) but does still occasionally go under or over the speed limit
@@AIDRIVR It's a shame. I don't understand what "broke" and I know that's the wrong term, as I'm a developer myself. I'm not sure I can call it a regression either because this is probably something that went from C++ code to NN, but it's frustrating because it worked so well in 11.x :( And every version of 12 this is the first thing I check and I've seen no improvement in any version. For me at least they are all just as bad as 12.1 in terms of speed control.
@@JohnVanderbeck From the perspective of Tesla owner it must be annoying... But from the developer perspective, it's a huge progress. Slightly wrong speed isn't really something to care about ;p
@@Plajerity Sure it doesn't sound like much, until you're the one in the car being annoyed and embarrassed by its behaviour around other drivers who get mad at you :) I do get your point though and yes it is definitely important to keep perspective. I'm in my late 40s and grew up with among other shows, Knight Rider. If someone had told my younger self that one day my car would be able to drive like KITT I'd have called the looney police on you. The fact that it can even do a small amount of what it does is still mind blowing and it is easy to start just accepting all the magic and only see the faults. At the end of the day though those faults are what prevents me from just kicking back and reading a book on a long drive while it does the world. Or being able to drive myself to the airport, then send my car home on its own knowing it will come pick me up later. That's my hope some day :)
I like at 12:30 for a frame the car wants to slam on the brakes at 10mph then changes its mind. While this isn’t new, older versions you would have felt a small jerk but it seems the car didn’t react at all.
Bruh - that Bus situation was straight up bad driving.. Can't handle it when you just completely ignore that and act like the 'smoothness' was the most important thing..
10:50 just remember that the law is that if the car in the other lane changes into your lane for any reason whatsoever and hits you it’s 100% your fault. I wouldn’t mind it either but I never do this myself. lol
Amazing progress and amazing video! You were so critical in so many videos. It means a lot that you are so excited now. We are getting close to the finish line!
Wow, the improvement truly boggles my mind, as I have seen multiple different channels doing the same testing on previous versions and this looks like its almost ready for mass adoption. Fantastic work Tesla AI team.
Watching this video makes the wait for 12.5 in the Cybertruck much harder. Trying to be patient but waiting for the day I can have it drive me everywhere everyday.
mmm i mean it looks pretty good but yeah it really depends if it has noticable faults occur over longer periods of time, tesla's have a higher crash rate then normal cars pretty substantially because people are using the self driving as a replacement, it actually getting better could make the crash rate increase because people start to put too much trust in the systems and start to zone out, would be interested to see if FSD crashes have decreased over time, & trends you can make from the data to infer when it'll be below avg crash rate (aka when it should be given to as many people as possible to minimise car accidents) get on that AI driver ;)
i think that you need to reduce your expetation, Simon is really doing a good job here in his video, by describing pros and cons of this version. you know it is just a snapshot of right now, it is not like one update now will solve self driving, we have to wait
Sorry, unsubscribing and removing from my feed. Elon's latest disclosure about what should have been a private matter about one of his children -- something that deprived her of her dignity and privacy -- is unforgivable. Elon has shown himself to be a failure as a parent, a human being, and with his support for TFG, a threat to our Democracy. As a father of two, it's just one of those things that I have to take a stand on. I hope that someone finally conquers FSD, but I want no part of anything that Elon creates. The best thing Tesla's board could do is find a new CEO to take the reins.
For the mind of his own: i think what's happenning is that in this intersection, the IA is much more confident in turning left than going straight, either because he had no information about going straight, or that the IA noticed that going straight is more dangerous, up to a point where the IA will actually overseed the GPS will, and the IA will just do the turn that he think is the safest, instand of the shortest.
As someone who doesn't particularly like driving full self driving can't come soon enough.. I like the "showing intention" part as indicating what you intend to do via "body language" on the road is a thing.
I threw every thing I could give it from downtown Denver and 12.5 handled it like a dream without a single intervention. This is the FSD we all "really" dreamed of.
Oh my God while I was listening to this driving home, my car got to the house and like it’s been doing since I guess 12.3 or something looked for a parking place- it didn’t find one and instead of just turning it off like I usually would I just let it go and it actually drove around the block came back to the house and stopped right in front. I’m hoping that soon it will know to approach the house from the other direction so it’s on the correct side of the street when it looks for a parking spot and maybe after that be able to park in my driveway but wow wow… This is really a game changer. Love your content, well done my friend
Fazard turned me on to your channel. Wow, totally impressed with your analysis, editing, screen display and driving scenarios. In one drive: a stationary cat, a crossing deer, going around a car accident, highway flagman etc. Amazing. 12.5 seems to anticipate rather than react. And thank you for pointing out residual issues such as curb hugging and blinking red lights. Looking forward to your next video! Ready to start using FSD on my ‘21 Model 3 once this version or later is available.
ya, I got 12.5 a few days ago and it now handles situations it failed before. pretty much all situations it couldn't handle before. that was impressive. now literally all I have to do is set destination and start FSD and just sit and wait until the car drive me there. only thing I need to do is find parking spot and start auto parking. so if it can find parking spot itself and do auto parking there will be nothing I need to do except setting destination.
3:59 optimisation for HW3 should be about how to make it run on a less performant HW, optimise memory use, clocking, MIPS usage. They already said this FSD version several times, HW3 is a becoming a slim fit and needs more “training” 😅 to fit
Wish I had 12.5, it is so aware of it's surroundings - airline pilots call it situational awareness - just a major step up. Your ride and Chris' rides (Dirty Tesla) have made me feel like FSD is almost ready for unsupervised although I think a few .x releases into 12.5 will clear up most of the glitches. Looking forward to the Berkeley drives, cheers.
Ashok's quote is interesting. I'm volunteered for longer car trips, often to Orlando either for vacations or to drop someone off at the Airport/Train station and I regularly will get comments from friends/family when the car does something sudden. One of the biggest problems has been lane changes,. It will indicate a lane change then get like halfway between the lanes and then either sit there not finishing, or suddenly abort with a hard jerk back.
That's because of the old highway stack. City FSD is this video you see, but the highway stack is basically stuck in 2019. That halfway change and sudden jerk back is a very 2019 FSD behavior.
That is Ai for ya....It needs to be trained to be more decisive. That is a normal Ai actions without better training. IT has to take into consideration of other vehicles on the Road, and hopefully, they can train the vehicle that way.
I’m loving it too! Every road I been on the last few weeks have all been different as we have traveled 4K miles. 3.6 has been great through most of it but 12.5 is on another level. It was trying to make a u-turn but decided it was too tight. It went ahead and made a left and routed its way to the supercharger. There are so many amazing examples the last two days it’s not possible to list them all. Like you, this is what I’ve known Tesla would be able to do and enjoying it immensely.
