Tesla's FSD 12.4.3 First Impressions & Major Challenge in Chicago
ฝัง
- เผยแพร่เมื่อ 6 ก.ค. 2024
- This was filmed on July 5th, 2024 - Tesla's latest version of Full Self-Driving v12.4.3 was used in this video. This is v2024.15.15 of the software, released on Monday, July 5th, 2024 to original group testers.
Please visit my Patreon site to support me. I provide insider tips, advance notice of videos to be released, and additional behind-the-scenes content.
/ techgeektesla
I first received FSD (Beta) 10.3 back on October 24th, 2021 as a '99' safety score driver.
My Build & Info:
-Location: Chicago area - Northwest Suburbs
-Vehicle: 2019 Tesla Model 3 Long Range, Rear Wheel Drive, HW3 Intel
-S/W Status: Full Self-Driving Supervised enabled, Premium Connectivity
Follow me on X @jonbbc.
Check out TLYARD for high-tech accessories for your Tesla Model 3 or Model Y! During the month of July 2024, you can use my Discount Code: TechGeek
for 23% off your purchases:
Whole website:
shrsl.com/4l1kw
T3 Pro Display:
shrsl.com/4l1l1
01:46 - FSD Won't Turn On
03:36 - Turning Left from the Wrong Lane
05:27 - Excellent Merge onto Highway
06:00 - Poor Lane Changes
07:48 - Scary Reaction to Merging Cars
09:23 - Lower Wacker Dr
11:10 - Hesitating for Right Turn
11:35 - GPS Connectivity Issues
13:19 - Sub Lower Wacker Dr
14:00 - Blocked Road
14:25 - Encounter w/ Chicago Police
16:20 - Misses Green Light Opportunity
17:22 - Mixed Feelings about FSD
18:20 - Huge But Very Important Edge Case
20:32 - Horrible Human Driver
21:15 - FSD Moves Proactively Before the Light Turns Green
Are you located in the Chicago area? Would you like to test FSD Supervised? Come take a ride with me!
You can book a ride here: www.tezmee.com/profile/376
Please subscribe to my channel and smash the Like button if you enjoy watching my videos. Thank you so much for your support.
Here's the gear that I'm using to make my videos:
Camera Kit - GoPro Hero 7 Black:
amzn.to/3sVfsBV
High-speed Micro SD Card:
amzn.to/3szEQQ7
Camera mount (outside car):
amzn.to/3KMDHeA
Camera mount (in car):
amzn.to/3wOKKNa
Camera Rod Extension:
amzn.to/3t2a4wV
Phone:
amzn.to/3t0N69o
Phone mount:
amzn.to/3LRYElX
Latest (2024) Microphone:
amzn.to/496n8mQ
Alternative Microphone (for iPhone - simple, but works great):
amzn.to/3MU21dr
#TeslaFSD #FSDSupervised #robotaxi
Music by Hotham
Stream: linktr.ee/hothammusic
Free Download: drive.google.com/drive/folder...
Music by Atch
Stream: linktr.ee/atchmusic
Free Download: drive.google.com/drive/folder... - บันเทิง
I have 4.3 as well. Most TH-camrs say it's perfect. Even with my simple driving and destinations I find it uncomfortable. So thanks for bringing reality to FSD. Yes, so much progress, but for the everyday driver who is not a test pilot, it's still early. Thx!
Yes, FSD is still years away from Full Autonomy. And the Robotaxi is 4 years behind schedule too.
I think it depends on the location. Some locations have more training data so it can drive better. I definitely think this end to end AI approach is the correct approach but I think more cameras and compute power is necessary to solve for true FSD. Hopefully I am wrong and HW3 is sufficient.
FSD V12 is so much better than any earlier version. The rate of improvement is phenomenal!
Love your videos!
I’d say in this Chicago area yes, but if its somewhere that has good enough map data then it performs pretty well. What general area are you using it?
@@nimasahabi9421 _I definitely think this end to end AI approach_
Clearly, after 100 days of training the Neural Nets, FSD still needs human interventions on any challenging drive. End to end AI is not getting the job done.
this video gets a cult status how you let fsd "explore" the underworld of Chicago.
Hey, the car is just curious! 😁 It wants to see the world.
And we have seen something, we have never seen before 🤩
😂😂
bird flap sound effect is phenomenal.
😂 🤣
100% that whole sequence was amazing
also the police siren 😂
Chicago.. 😱 this city is the right place for harsh training of FSD.😅😅😅 Thanks for your patient driving!
Thanks Jon, that was fun. I can’t believe you let the car do its thing for so long when it gets confused. I have a limit for how poorly it can behave for other drivers. You’re way more tolerant than I am.
And who knew there was a lower lower wackier!
It has achieved visiting tourist driving level.
You're generous. And I live in a small mountain city with a lot of tourists.
😄
This is THE best FSD video Eva! Thank you for your bravery 😂
LOL dude, the editing was hilarious. great stuff, subbed
such a fun video haha, can't believe I didn't find your channel sooner
Best edited FSD video I’ve ever seen. So easy to follow what was going on. First time I’ve seen your videos but not the last.
Monsters and trolls and rats… 😂
Best edited FSD video I’ve ever seen. So easy to follow what was going on. First time I’ve seen your videos but not the last.
Best fsd commentary so far that Ive seen. Subscribed.
You just earned a subscriber! I believe in the tech and feel it is just a matter of time before robotaxis exist. However your videos give a realistic view where FSD shines and where it fails miserably. Keep up the good work and I’ll continue to watch!
Thanks. Nice work! Looks like a very stressful FSD drive, and you're braver than I am. You're right about it being schizophrenic. Some days it's perfect, other days it just seems to lose its mind. I'm still on 12.3.6.
100%
In a busy big city, the beta FSD software is never perfect. Full Autonomy is years away.
@@DerekDavis213In my tiny mountain city, FSD (Supervised) is schizophrenic too. Not just signaling right to U turn, but changing to the right lane to U turn left (on several tests). And so much more derangement. Maybe decades away. Or never. Performing well on a few repeated influencer routes isn't progress. I suspect they train it for where Musk drives, plus a few top influencers, most of whom do the same route over and over.
@@bearcubdaycare I totally agree with you. Despite all the hype and false promises, real Level 5 Autonomy is a long ways off, IMO.
I love your video, a FSD user myself I love watching others experience with it, especially in a major metro area. New subscriber today!
You are my new favorite FSD TH-camr 🎉
Many great training scenarios! FSD seems so close yet so far away.
So true, isn’t it?!
Hopefully we're not saying the same years from now ;p
Amazing review. Please keep doing this
Love the unbiased review. Long time FSD owner too. I went all in on Tesla when Musk gave his first demo drive of FSD 12.0, my reasoning was notwithstanding cars, deep-learning will leapfrog from cars to Optimus, because a lot of factory jobs are simple and idiot proofable .
nice video!
I think your video exposes the flaws of the system. This proves that alot of work is still needed to solve for true FSD. Hopefully these issues can be solved soon.
I guess many humans do the same kind of mistakes when new there. But we eventually learn the area we usually drive in.
FSD drives like it has never been there before!
If FSD can drive perfectly without knowing that specific place, it would be better than human drivers are.
Thank you, very nice video. FSD is not yet ready for Chicago.
FSD is also not ready for Toronto Canada, New York City, or Los Angeles. Busy big cities are beyond the grasp of the beta FSD software.
Nor small mountain cities like mine.
Basically FSD hasn't really evolved since I last watched a video 2 years ago. Not even close to being a useful commercial application.
@@rioriggs3568 Agreed, FSD is a novelty, a toy, but not something you would trust with your wife and child.
Great video/commentary.
i love this video, good job, you let tsla try.
Excellent video! The intrigue; the suspense! Very informative; thank you for all your work!
I am 77 and a first time Tesla owner; bought a Model 3 in May. I love FSD (currently have 12.3.6), but it is in no way close to ready for prime time, in my opinion. The improper lane choices are a major concern to me. On a trip to visit relatives on the 4th, several times it threw me (and I mean frightfully aggressively at 60mph) into the left turn bypass lane when there weren’t even any cars turning left, so there was no need to bypass anything. And too, 12.3.6 sometimes makes the wrong lane choice at stop lights, then proceeds as if it didn’t care, even though my car is in the wrong lane (in left turn lane, but going straight). From what I’ve seen, lane indecisiveness (lane dancing) seems to also still be an issue even in the latest release. I love the technology, but there are so many more situations FSD has to learn how to handle before I can truly trust it. But then, hey, that’s why it is called FSD Supervised.
This is a great fsd video
Love the tech but these city drives are probably the truest indication of how far along it has progressed. Some are saying it’s a HW3 issue but I’m curious if that’s actually true. Hopefully more data is gathered in these tough conditions so we get unsupervised FSD soon!
funny video loved the editing good stuff
Love the commentary 😅
You're a lot braver than I am! Can't wait to get this in NYC. Multilevels, routing and GPS inaccuracies are a problem here too.
Wish we could’ve finished the drive and that video was very into it
@@justincrandall3313 need to sleep. 😴 😝
I do not own a TESLA but I am really into the FSD and got my bucks in TSLA stock. Content like this helps investors as well. Thx.
Chicago looks like a cross between night city and megacity 1 in this video.
It’s totally not ready for Robotaxi like this.
Robi taxi will use AI 5, the new hardware.
@@corelsihuom8407 We will see how well this 'new' hardware and software works. Meanwhile, Waymo has a thriving robotaxi business today. With many paying customers, today.
@Ddj2112 wrong. Robotaxi which will contain HW5 and allow Tesla to brute force success for FSD with a really large model (GPT4 much much better than GPT3.5). Then, they can optimize and make FSD on HW3 & 4 as good.
@@danypell2517 you know that queries to GPT are computed in an inference server farm right? Try doing that with instantaneous driving decisions. (You're being lied to about FSD being akin to an LLM inferenced locally)
Don't be silly; he meant 8/8/2088.
lol, dude, you have balls of steel!
i like your style. it's probably not good for your nerves, but this is the type of environment I want to see FSD Supervised tested in.
3:54 this reminds me a whole lot of V11. The hesitancy with any cars it sees at intersections is ridiculous.
🔥🔥🔥
Two things things I did that improved the experience were, setting it to assertive and making sure my max speed is set faster than the flow of traffic
Just got my 22 M3, thanks for the thorough review. Where did you get your pedal covers?
It may not have initiated FSD from the start because you were parked and the line next to you was a solid white line but noticed that as soon as you traversed the white line it went into FSD.
Love your channel. Best camera set up, and most challenging driving conditions. You call it as it is and are extremely fair in your assessments. Kudos!
Thank you! 🙏 🙇♂️
This was the first of your videos I've watched and it was great! You, sir, are being added to my collection of FSD TH-camrs. 🏆
wild video! Thanks
The drama! Best FSD voiceover ever.
Very thorough! This is sort of frustratiion I had too. Elon and fsd team should subscribe your channel!
Such a good video. The struggle at the start is why I don’t get excited for these point releases.
You have nuts! With the red car. But it is quite safe if people le it.
Great video, suggestion at 6:20: put FSD into ‘Aggressive’ mode…7 😢
The most nerve wrecking video ever. The police scene was wild. 😮
I don’t think I’ve ever been able to have full self driving pull away from the curb maybe I’ve just never waited long enough
It seems to me, the more congested and slow traffic is, the better FSD appears to be. But this isn't because it's so great, it's because it's loves being cautious by nature. Slow going is where it likes to be. It's the real-time smooth flowing actions required to keep with traffic flow that it struggles with.
Good edge cases. Also I can see where the car would lose GPS-location signal under overpasses. I used to survey with Trimble GPS units. Tesla needs to determine a fix for lost GPS signals.
NGL, sub lower wacker was my favorite bit
😆👊👍
It’s all about correcting the errors now , overall it drives better than human already.
I enjoyed your video. Can you clarify something for me? Re screenshot from the beginning of your YT video, the video title is "Tesla's FSD 12.4.3 First Impressions & Major Challenge in Chicago" and your car's display states FSD v12.4.3. But, your video description states v12.4.2 was used in the video. Was all of the video of 12.4.3 or 12.4.2? If 12.4.2, why would you state 12.4.3 in the title and show a screen shot at the beginning of your car on v12.4.3?
Thanks for sharing!! There were definitely some edge situations, map data issues or connectivity or both. That you normally wouldn't run into. It's a very busy area where you were driving and it definitely needs to learn how to handle those. This is the third point release they've done on 12.4...1/2/3 I believe most of us are still on 12.3.6 This is really different how how they have done the releases this time around. Slow or cautious on go wide and still it looks like it's going to be some more time. It makes me wonder that if changing the release format from the past is have an effect on the progression for 12.4... Since I've had FSD 2 1/2 plus years. This is the slowest to my knowledge on going wide .. Is it Robo taxi or protecting share price or both? 🤔 Thx 🇺🇸✌️
❤❤
I think you reported several things that in my opinion are fine.
The reason it doesn't initiate right away is because the local environment (occupancy grid, close-up obstacles, etc) are "recalled" from the last time you parked there. There are dead zones near the car (especially in front) that the cameras don't observe, and all that data is "stale" from the previous drive (ex: parts of the curb where the camera can't see, close up objects front and back). The system is not sure if the data is recent enough to safely proceed. Therefore, you need to move out and away from nearby obstacles manually so you can "clear" the dead zones and the local environment gets sufficiently refreshed around the car for it to be confident that the information it has is up to date before going into autonomy. From my experience it's a small buffer directly in the path where the car wants to go, and in your case it's too close to the curb so it's not going forward in order to not curb the wheels in case it was parked by a human too close. It's putting the liability on you to get out of that situation
You had mentioned, in your last video, that you are more relaxed, driving by yourself and that it is not relaxing. It’s a different kind of relaxation. On the expressway, FSD takes out 90% of the stress and I am much less fatigued if I go on a five hour or 6 Hour drive, for example. However, When driving in the city and so forth, it is more stressful because it is like driving with someone who is getting their hours for their driving permit. You have to be on the ball and ready for anything so that you can take over. That said, I am still behind you on updates and I would say that there are absolutely no critical or dangerous times with my car. Almost every time I intervene or take over it is because it is just waiting too long to do some behavior or action or it is in a lane that I don’t want to be in, or route that I am not interested in taking. Additionally, it does not avoid potholes and obstacles of that size ever!
If it’s no critical disengagements , means it’s ready for robotaxi , a trial version 😮?
Thanks for not being a cult member about fsd. 12.3.6 is ok for me it’s mostly the navigation that chooses the most insane route thus making the drive more unpredictable and more difficult for fsd. Wish it could follow Apple Maps
You, Sir, have brass cajones. I would have disengaged like 6 times before the video was 60% through.
Thanks Jon for the great video!
Minute 6 shows the difficulty when manual and autonomous driving is parallel, humans squeeze themself per brute force into the lane, should robots also do this, or should they obey the law, when others are not doing it by keeping to close distances?
@@peterhelm522 Fairmont Hotel is a really great destination to force the route to Lower Wacker Dr. Without you, I would not have known this, so thank you! You have a really great question with a not so simple answer. It’s a very difficult one. Isn’t it?!?!
Lower Wacker Drive (LWD) is kind of "uncharted" on map, since it is 3-dimensional. In planning the route, there should be pre-determinations on each exits about LWD.
So FSD sucks?
Is it not possible that disabling your peripheral cameras is affecting the whole system?
Very impressive but a little bit of frustration(FSD). What do you think of commercializing robo taxi at this moment??
I beg to differ. We humans, when confused by a weird situation, we take cues from other drivers around us. If FSD is learning how to drive from the habits of good drivers, it should take cues from other drivers experiencing the same dilemma or confusion. Even if we are wrong, there is strength in numbers as long as it is safe.
All the driving on the highway is using the old 11.x software stack so not relevant to 12.4.x. Elon has said v12.5 will be the first end-to-end neural net version to include highway driving.
!
FSD is terrible in Chicago. I live here and rarely have an intervention free drive.
Would you consider taking it out to a geography that might be less well sampled, such as Mexico? i’m quite curious to know just how generalizable its capabilities are!!
I haven't seen your vlog before, but I like this video. I've been testing FSD for a little over a year. Regarding the green light at 17:50 in the video, I'm going to disagree a little bit about this edge case being that important to fix. If FSD is going to be a driver assistance system, Tesla doesn't need to handle this problem (or many other edge cases), the driver stepping on the accelerator pedal fixes it. However, a RoboCab will have to handle this problem. The Tesla AI is going to have to recognize quickly when its stuck and develop strategies for what to do - especially when the other cars around them seem to be able to go. I expect a portion of the AI engine is going to have to be rearchitected to handle a plethora of these special cases. That won't happen until V13 or V14, when they've finally accumulated enough special cases to even approach reengineering an architecture for solving them all.
I was distressed to see FSD driving over what are clearly diagonal parking spaces. In a grocery store parking lot, that behavior is asking for an accident.
Overall, I think Tesla is very close to turning FSD V12 into a good driver assistance tool. However, I think they are years away at FSD running RoboCabs. Cruise had people at the back office intervening - about one person for 4 Cruise vehicles. I believe Waymo is doing the same thing. Neither of these companies is/was executing on true autonomous driving.
The edge case is important to fix if Elon is telling the truth about everyone being able to deploy their cars to the RoboTaxi network in the future. He has been talking about this ever since 2019 - how Tesla vehicles will be the only vehicles that will become appreciating assets. This is achieved by allowing them to travel around autonomously to make money for you.
I can't understand the enthusiasm for this newer version of FSD. You say it performs well on the highway 'except for merging, lane selection, and getting off at exits'. What else is there, besides lane keeping, which was good enough in AP 1.0?
Highway still on V11 stack...
That was absolutely legendary when it went to sub whacker drive.
5:00 still stops way too short of every stop line at intersections. Something they haven't fixed since V12.1 sadly...
I hope Tesla will take advantage of the difficult driving in Chicago. They should activate more Tesla's with the v12.4.3 in Chicago. I agree if FSD can master Chicago it can master driving almost anywhere.
They have to do atleast 5 to 10x more training. With 70% training on normal driving with huge variety of places and 20% training on unique scenarios and difficult roads and 9% on disengagement and 1% on edge case.
Note - I am a "Bored" certified "Expert" 😂.
Just speculating I am just a tech geek.
I agree. We know next to nothing about the way they train it; but I'm absolutely confident they're developing order within the chaos, the Tesla way, and someday soon, they'll crack it and win!
I'm assuming it wouldn't work in the beginning because you were parked beside a solid white line.
seriously, it needs to automatically reverse out if going into that dead end/blocked gate instead of stuck there
now watching this video and understand why it is hard to roll out in Asia , many of the situation we would get honked at, people here are impatient when they drive
Interesting how there are so many different assessments of what is acceptable with FSD. I hope we are all careful to evaluate it based on safety and getting you where you want to go rather than saying it’s not working because it doesn’t do everything the way you would. I think although less and less frequent, a lot of people disengage because it’s behaving in a way that they are not used to, i.e. waiting too long to stop sign or something. It’s hard to feel like you have to monitor but you have to give up control about how the car is driven- a lesson I wish my girlfriend could learn!
I hope eventually FSD will be a partner that we can help train to drive the way we are comfortable with. I’d certainly like to be able to have it to take a route that I prefer rather than always having to follow navigation. How great would it be if you could just turn it on and say take the next left orpull in here I want to get a soda?
So basically, it's not there yet again.. I'm gonna watch for the next update now.. I'm not interested at the moment, lol
I know you know, but highway stack is still 11.x version. I couldn't tell whether your merges happened after it transitioned to the highway stack or was still running the 12.4.3 stack with auto max. Also, I have mentioned before that I think some of FSDs intervention/engagement instances is a result of bad navigation data.
You stated "I know you know, but highway stack is still 11.x version." Have you seen that being explicitly stated in any of the Tesla official written documentations? If so, please share the Tesla official written document(s) with the rest of us. Thank you.
You can’t have a too taxi with bad nav data that is part of the gig
Any idea when 12.4 goes wide
What’s the big picture here? Where is this tech headed and how long will it take to get there
Great question. I can take a stab at answering that in the Part 2 video based on what Elon Musk has mentioned.
Well the driver of the red car did not have his turn signal on, so I guess the Tesla did not exptect him to change lanes?
Does it still think Route 45 and Route 50 signs are speed limits?
I'll call it the Whacky Drive Incident.
Is this now a single stack for freeway and surface street driving?
Not yet.
11:15 the GPS slips a fair amount on the visualization. 13:15 chicago has a deeper level? you have been holding out on us :)
17:45 I think is another example proving that it's not a pure end to end net because it relies so exactly on the visu take on the lights. So they gave it some of the basic filters as a scaffold for it to learn with. Sort of cheat sheets but it then becomes reliant on them and subject to their imperfections. The no right on red is not a red traffic light.
21:00 that could be an interesting case of FSD starting to go because a cheater did. Some learned herd mentality.
Great test. The others can retire now, it's all about chicago :)
Great comments and excellent observation with the red arrow slipping at 11:15. I also believe that there is still some underlying heuristics code for traffic lights behavior.
Its not about the car driving like humans would. Its about the car driving safer than humans but not to the point were its scared to drive or make moves it needs to make.
The high way software is still running the old version. I would think that high way driving will drastically improve when the two software stacks are finally merged.
Yes, only they were finally merged when v11 came out. I waited a long time for them to merge before, and it was a really big deal when it happened. Tesla was radio silent about the stacks separating again with v12, so I think it’s crazy that everyone gives FSD a pass just because the stacks haven’t been merged together.
I follow a half a dozen FSD channels. Like your take as well but am concerned for your safety. Not worth the 1/100 chance of an accident in those questionable FSD maneuvers. Please, please take over sooner.
Thank you for your comment and for your concern! I will do my best to stay as safe as possible. You’re right, I am probably a lot more trusting of the software than others. Just because I am filming doesn’t mean I should put my own car at risk. I have never felt like I was putting my life at risk though. I am always taking extreme care to monitor who is behind me at all times. Having a big huge camera hanging on top of my car really helps.
What happens if u have the blind spot cam taking up extra memory?
Nothing, probably. 😆 I just would rather not have my visualizations glitch. Sometimes it’s not very smooth with the older Intel processors.
The line changing isn't normal? people never hesitate?
I watch these on 2x speed. Going to 1x is sooo brutally slow. Does it feel this slow in person?
@@sald7485 smart!
The problem with release notes is there isn’t specific new changes. That’s not how the new versions are updated. It’s more training and I think Tesla is still fighting out how it works.
In most cases with neural nets you can’t figure out how it works.
Keep testing! They should be paying you per mile of testing driven.
Of course not. It can't even go backwards yet.
They getting paid from TH-cam :)
Where is the 5 to 10 times improvement 😢