Just did first v12.3 drive and it was my best fsd drive ever. It didn't try to park, but did pull over in parking lots. It was superbly human-like with only a single acceleration intervention just for my comfort. And it passed the Wife Test!
6:20 "In all the road closure situations I put it in, it almost feels like it knows that it's not the right thing to do but always ends up just trying it anyway".... Wow.. so very much driving like a human. 🤣🤣🤣
Exactly what I was thinking. This is probably a topic where specific "staged" drives need to be recorded for the training data, because even the "good" drivers are sneaking around closed road signs. 😇
Honestly, the driving in this video was incredible. There still a lot of progress to achieve, adding the missing bits like parking, backing up, how to deal with construction, etc. and especially if self-driving has to work across different parts of the world. But it feels like this is the way to do it. Thank you for the video!
Holy cow. There are so many terrible, scary af intersections in this town. The number of roads you have to blindly pull out on where traffic could easily be going 45 is insane. We don't need FSD, we need better road design...
I watch most of the FSD channels out there. The production quality (and by production quality, I mean all of it... the video quality, audio, narration of scenarios, your narration voice, the editing... ALL of it) is the best there is. Just outstanding work. Thank you for this. As an FSD beta user since the first round of Safety Score adds, I for one, welcome our FSD 12 overlords.
Forgot to ask... You do a great job of running FSD 12 through ALL the scenarios. One that I haven't seen a lot of on your channel is traffic circles. Where I'm at, traffic circles are the bane of FSD. If you have any that are local to you, any chance you could take us along for the ride through a few of them? I know traffic circle density is highly variable according to geographic location, so if there aren't any local to you no worries; don't want you going cross country just for a few circles. : )
Internal Family Systems (IFS) theory suggests the human body is the same way. One part will be like "GO!!!", while another part (usually a protector) will stop us in our tracks.
@@KCJbomberFTWI’ve never understood why people post stupid stuff like OP. I mean, is it for “likes” on the comment? If so, why? Is there something you gain from posting like-bait? It’s always seemed so strange to me.
I am so impressed with your narration and your focus on describing how the car handles every situation you gave it! Excellent job! First time I have watched your channel. You are easily the best FSD test driver on RUclips and the most interesting to “ride along with”. I completely agree with your assessments. This is the year! People who think we are 10 years away from FSD will be blown away by 2024. I don’t know how many more FSD test videos you will be doing, but I will be along for the ride on all of them. Subscribed. 👍
@@AIDRIVR I have to add that some FSD test drivers I have watched take their time to make a RUclips video, but then everyone in the car chatters about irrelevant topics while the car is dealing with drunk drivers and doing other amazing things under 12.3. Unbelievable.
Not strange at all. Camera placement. He is filming from driver seat. The car has a camera in the middle. If the pedestrian was on the opposite side of the car in front, the car would not have seen it as early as the human in the drivers seat.
I don't think there's actually anyone there? When it first appears, we have a full view of the whole sidewalk going into that intersection and I can't see anyone. As he gets closer there's still an unobstructed view of the full width of the entrance the van is pulling into, until the last few seconds before the van stops, and there's still nobody visible.
If you back up the video a bit you can see the pedestrian walking, then the person gets obscured behind a parked car. The system has memory now and correctly infers that the pedestrian is going to continue walking across the road and slows down accordingly.
@@richardjones8395 No, it does not have any memory. But, the camera the car uses is placed more to the right compared to the camera that is filming the ride, and thus it can see things that are obscured by the video camera. Note that as soon as they get too close to the car in front, the pedestrian is gone.
@@Xanthopteryx Yeah, sorry, I totally mixed up with another section of the video here (I was talking about another bit in the video at around 21:43 where a pedestrian was crossing and the car waited). In the particular case at 3:00, it does see the pedestrian due to camera positioning as you say. But going back to the statement about memory, I should have qualified what I mean by 'memory': as I understand it, V12 learns from short video clips (around 30 seconds maybe), so when learning what a human driver would do in a given scenario, the neural nets will be taking account of what happens during the period of the clip and how the human driver responds. Whilst perhaps not a ‘memory’ in the conventional sense, it does give it the capability for things like ‘remembering’ a moving car that gets occluded by another car at an intersection- as long as it's just a few seconds it'll keep track of it. On previous versions, the FSD team did talk specifically about giving the system a short memory to take account of such situations. I’m not as clear yet how that works with the ‘end to end’ system and whether the team have taken specific steps to give it ‘memory’, but as mentioned above, the fact that it learns from video clips does give it the capability to infer that type of information (at least over the duration of the clip, whatever length it may be).
I watch a bunch of RUclipsrs for Tesla news,including FSD info, and i've been watching a few of your videos for a few months. I really love your commentary, editing, and the way you have multiple cameras E.G. interior vs exterior, and how you transition between them to illustrate your points! Makes for great "flow" in your communication, and really easy for the viewer to follow along, thanks for being awesome :)
00:01 FSD using human data for AI is working well 01:58 Tesla FSD 12.3 executes complex maneuvers well. 05:59 Tesla FSD 12.3 faces challenges with road closures 08:03 Tesla FSD 12.3 navigates tricky intersections well 12:03 Tesla FSD 12.3 exhibits extremely human-like driving decisions. 14:03 Tesla FSD 12.3 faced challenges with U-turns 18:06 Tesla FSD exhibits advanced decision-making skills in complex driving scenarios 19:52 Tesla FSD 12.3 handles challenges beautifully 23:35 Tesla FSD 12.3 stops early for potential hazards 25:30 Tesla FSD 12.3 brings significant improvements
I just went from 11.4.9 to 12.3. ALL of my pet peeves are gone! Late signaling, twitchy turns, incorrect changes into turn lanes, insanely fast RR crossings, all gone. Even new capabilities. An aggressive driver quickly moved into a turn lane from behind. FSD just let it by and pulled in behind. While waiting for a left turn, a car pull up and blocked my view to the right. FSD crept forward to regain visibility. Waiting on a right turn traffic passing traffic blocked my view to the left. FSD remembered to wait for a last car to go by and continued the turn! I need to completely reset my expectations to catch up to the progress of this release. Thanks for the effort to show it on video for all to see!!!
Once again you have produced the best FSD review video of the bunch. I like the voiceover style and how you analyze each situation. Very well done as usual!
Your mention about going back to 11 FSD for that one area is interesting. Would be interesting to be able to move back to previous versions to show the changes. The improvements are not so clear until comparisons are made.
The problem is that each time you drive the environment has changed completely. Cars are different, position of them are different, your own car is placed different, speed differ, the light differ, the weather differ, shadows differ, ... The only way to truly compare is by doing like Waymo does: Tons of computer simulations. You let the virtual car drive a virtual environment and then you can replay the EXACT environment with different software versions.
I get really excited every time I see you posted a new video and I have to watch it on the big screen tv right away... It's a profound understanding of the baby steps of human ingenuity towards safer roads... You could make 1 hour videos and I would watch it... I love your voice and comments... I wish you would go for 2160 instead of 1440. Thank you!
At 21:31, the reason FSD will be safer than a human is that human drivers exhibit numerous dangerous behaviors, which FSD will never exhibit, and which cause up to 90% of accidents by some estimates. These include texting while driving; other distracted driving; drunk and drug-impaired driving; road rage; fatigue; falling asleep at the wheel; medical emergencies; running red/yellow lights and stop signs; wrong-way driving; following too closely; unsafe lane changes; speeding; street racing; reckless driving; loss of control; and several other categories I probably haven't thought of. FSD can be significantly less skilled than an average driver, perhaps similar to a beginning student driver, and still show a much lower accident, injury, and fatality rate that human drivers, for this reason.
I can't wait for the day we get true FSD! This will give so much freedom to my disabled wife. But, it absolutely needs to be six sigma, requiring zero intervention from the human riding inside the car, who is incapable of helping out. But this will be a game changer for so many people!!!
sorry to say, but this will never happen they are on it for 10 years now and it wouldn't pass a simple drivers test that any human with 12h of training could pass hell, in the video you see at least 10 examples where it makes terrible decisions; blocks the road; ignores tight spots until breaking in panic etc or simply breaks traffic laws
For the first versions using end to end neural nets its capabilities are amazing. Just think how much improvement we might see after two years of trained neural nets!
What was the year make and model of the car? And in your experience do you think it makes a difference if it's the Intel or the AMD based computer? @@AIDRIVR
@@KuramaKitsune1the intel and amd computers are just for the infotainment. The FSD computer is a separate Tesla designed system. For now both HW3 and HW4 perform the same with FSD
Excellent - plus a great summary. V12 has a very good understanding of the cars dimensions and can squeeze through spots (not sure if that's always a good thing). Teaching my niece to drive (I had her go through cones and between boxes etc) and now watching people more, I notice a lot of people have poor spatial understanding and not as good as V12
even if probably legal, that maneuver at 1:20 where you're driving into the exit of the drive thru is something i would have disengaged for. both 'lanes' have arrows pointing in the oncoming direction and it looks like the lane widening for 2-way traffic past those arrows is for drivers exiting the drive thru, based on the road markings. which is probably why that pedestrian didn't look both ways, because the road markings are indicating one way traffic.
Great video as usual so informative and entertaining. IMHO the thing that needs improving next is the navigation data. When a road is closed that contradicts the data, a new route needs to be quickly determined. Also, the closure could be updated to other drivers in the area with a few hours of persistence.
Simple: rerouting is another type of AI action. Find clips of people driving with road closures in them, not necessarily in the route path, then edit the route to go through the road closure, and add the reroute action as the target action. Lots of issues can be solved by creative data sampling and modification like this. You can do the same thing for the issue where the car is unable to safely follow the route. This stuff will be solved so fast.
The one thing I'm learning from these videos is that FSD really needs a way of detecting "road closed" signs and feeding that into the navigation system, like maybe even just ignoring the navigation and picking a different direction to go instead, and letting the gps reroute to the destination afterwards. Also, while I really don't like tesla as a company (mostly due to the owner of said company), and self driving cars still seem like an accident waiting to happen as soon as supervision isn't legally required (not to mention that unsupervised self driving cars solve a problem that would be much better served by widely available public transit), I can't deny that this is really cool tech, I'm kinda shocked how much this system manages to do correctly, even if it's definitely still at a point where intervention needs to be available at all times, it can drive better than most self driving tech I've yet seen.
I know the way you edit takes a lot of work but your approach - showing the boom-deleted car when appropriate, and the voice-over makes for a really useful video. Many thanks.
I got 12.3 this morning, coming from 11.4.9 🥳. They gotta get speed control in order 😅. Other than that it is a great leap forward as far as I can tell.
Great video, as usual. I'm really impressed by the improvements you've shown from FSD bets 12.2.1 to 12.3. I sure hope this starts rolling out to a much wider group of FSD beta testers soon so that we can get it on both of our Teslas.
I really do like these videos. People watching can learn to maybe drive better by example. This system doesn't get road rage. Being a professional driver for close to twenty years has taught me a lot of patience. I hope to continue watching videos of yours more ahead because so many people have a fuse set to road rage. If anything, it can be a great example of how to drive well.🙏
I only count them as critical if it was in any real danger. The car just stopped itself and needed me to re-engage, no danger to itself or anyone around
I like that you let FSD v12 do its thing to the very end instead of disengaging early. Gives us greater insight into what the car is trying to do/what its thinking.
Quite a fresh driver here, got my license in mid-February and I have got to say, that Tesla FSD drives similarly to me, so hopefully we will improve together xd
22:00 when the car slows down when it sees the light change to yellow, but then changes its mind and accelerates is mindblowing! It really shows signs of intelligence.
Not only inevitable - that's Tesla intentional strategy. That's a good thing. It doesn't mean the car will make mistakes like humans necessarily. The car is 100% attentive at all times (which is a large source of human error) and the training set is curated good driving.
When you look at crash compilations, it's most of the time because of either reckless driving or fatigue (microsleep, general distraction). This version shows no signs of doing mistakes like this and thus I would feel much safer knowing the cars around me are AI driven and don't just cut my lane, drunkenly swerve around or overlook pedestrians.
I think about this all the time on the highway, especially when someone is tailgating me and glancing down at their phone. Using FSD lets me watch other drivers around more and it’s terrifying
I'm sorry no one told you John but we took a vote and decided that due to your impatience, you are being slow walked on FSD. Perhaps a better attitude would help your case ...
Got my 12.3 yesterday night and drove it today.. really good improvement.. one thing I could notice is that lane centering is not perfect .. it’s hugging the left side a lot more than 11
After watching in awe, I've finally gotten the v12 update tonight. Love living vicariously through your videos with all the updates and now I can live it myself haha. I'm pretty cautious with FSD in general and don't actually use it much but am excited again for this update.
21:30 It's not like human drivers are bad at their core. We do that shit well, until we get distracted, inattentive, tired, impatient, reckless, ... So an AI, with its infinite attention span and patience, that is trained only on the best behavior should in theory be among the safest drivers out there.
(2:29 ) Just because the self-driving car needs to make a left turn immediately after the right turn doesn't mean it should put the left turn signal on before doing the right turn. The proper way is to signal right first and turn right. When the car is about 45 degree pointing to the right, change the turn signal to the left for moving to the left and for left turn at the traffic light. Improper use of signal is a traffic offense.
It's ok, the driver will happily pay the fine for the offence the AI made. An AI that is 100% property of Tesla, trained by Tesla, the customer only has a limited license to use.
@@_BangDroid_The driver is in control of the motor vehicle. They choose when the AI is driving and when they take control. The human should always be responsible for the AI’s actions until the company says that it’s safe to not actively monitor the AI, at which point the company is responsible for
Regarding the moment at 16:20 where it crept up on the guy who didn’t start going when the light turned green, basic Autopilot actually does the same thing! I’ve noticed for several months now that base Autopilot will release the brakes and start rolling up on the car in front of it if it can see the light turn green. It’s really cool to see.
Thank you for all the vids. The government has the final say but the following is my pov. I am not a numbers guy but insurance companies are. When they get the numbers that FSD saves money and life's the pressure will be on not only government but ALL auto manufacturers to license this tech.
I'm puzzled why some features have still not been implemented. 1. Road Closed training: it would be straightforward to sample incidences where the planned route was changed by human drivers deviating from navigation and where there's 'road closed' text (there will be vanishingly few examples of people literally driving up to the sign before responding!) 2. Dead ends etc: trivially easy to find and train on instances where the driver selects reverse (esp repeatedly for a multi point turn) and then proceeds in the opposite direction to that before reverse was selected. Having said that, should we be expecting a much faster rate of improvement from now on? Cheers dude, literally top content, as ever. Hoping the new eqpt is making your day a lot easier.
I think it's not wise to only count on visual queues. If waze (or Tesla's version of crowd-sourced navigation) tells you the road is closed, you should have come around anyway...
It is so stupid that it doesn't know how to properly use navigation. If it used any mapping program for routes it would have calculated from the start that route wasn't an option. Heck even my 2013 car does a better job at that.
Just experienced FSD 12.3 in a HW 3 Model 3 yesterday. Car was going through a merge with cones on right hand side, the construction vehicle on the right let out a huge bloom of steam and FSD seems to be blinded and NOT avoiding the cone. The driver manually took over and swerved, just clipping the cone or it would have plowed through the cones. Vision only system just will not cover the rare cases that it's blinded. Same with fog.
V12 is nutty coming from v11. Especially in my area. Gawd dayummm it’s like a human driving the car. Especially with turning into side streets and unprotected left turns.
21:44 pedestrian was perfectly visible for car, it's funny that non end2end display related neural network did lose that pedestrian later, but end2end did not lose it and tracked, so in this situation it was ok. 17:44 it's actually example why current approarch will not work (at least not as safe as waymo or mobileye one) the problem here is not car turning wrong way on road, it's the attempt to drive over static obstacle, i should point out that 12.2 did try to drive over pedestrian same way. while 12.3 does have safety net (4:18) it far from being good enough compared to waymo and mobileye, i.e. in some situations it will be too late for that, and waymo and mobileye safety net does not result in automation disengagement, just limiting neural nets not to do unsafe actions. p.s. i do understand that safety net will result in more robotic behavior (like 11) but it would be much much safer, and obviously there should be a way for a driver to temporary disable some limitation - for example to allow car to cross solid line, but extreme attention for human driver is required at such moments. p.s. waymo do drive without serious risk to human life and without human overwatch(there is no way to stop immediately for any human) for several years now, but it less smart than fsd 12.
It'll be interesting to see how they address variations in traffic control from state to state. Based on what you said, in California, a red right arrow is the same as no right turn on red. But I live in Oregon and here, a red arrow simply indicates a mandatory turning direction and legal turns on red are allowed unless there is a NO TURN ON RED sign.
Great video! Hope the next models get cameras on all four corners to better manage intersections and backing out of tight spaces. Seems to be room for improvement there and it is unlikely they can do much when there is no visibility.
13:10 In those situations, you really shouldn’t leave a gap. It can be very dangerous because cars in the other lanes can’t see them and if they attempt to turn past your open lane while you are stopped cars to your left can’t see them and could cause a collision
That guy just waived at a robot
This shows that the robot already behaves very much like a human being
@@tom_skip3523shutcho corny ahh up 👨🏭🧜♂️🕺
😂and doesn’t even know it …
Please tell the robot, the human is saying thank you. (Robot saying "what an odd behaviour")
@@tom_skip3523 Surprisingly, doesn't take a lot for a robot to behave like a human - we are not that far ahead on a cosmic scale
Great video. At 12:40 in to the video consider this:
A human driver waved “thank you” to an A.I. without knowing it. Turing test complete! :-)
Thanks for the laugh 😂, we sure are living in an interesting time.
For people confused it is not at 13 minutes, but around 12:40
@@Wigglylove Thanks for the clarification. I edited my comment to put your time in. Thank you.
You are, by far, the best at describing and analyzing the latest version of FSD. Can't wait for you next one!
Nobody else has Version 12 or has that changed?
@@cafl9844 Everyone should by now. I got it last night. My son got it last night. We never get updates until wide release.
💯 The commentary is great!
Heavily biased takes though
Agreed most of the rest of the FSD videos are weak compared to AIDRIVER
22:15 , Straight up says no and goes for it like every human LOL. This is amazing!
Exceeding the speed limit by 10 mph!
There was no danger at all@@ThePixelize
What subscribing autopilot
@@ThePixelize yep, human like 😂
I thought the goal was to be safer than human drivers, not equal
Just did first v12.3 drive and it was my best fsd drive ever. It didn't try to park, but did pull over in parking lots. It was superbly human-like with only a single acceleration intervention just for my comfort. And it passed the Wife Test!
Passing the wife test is the hardest challenge
Where did you do it?
The only harder test is the "would you let your kid ride in it" test
@@tateisaacs1048better to teach them how to drive first.
I would like to test it too. Where can I find one? Wife I mean.
No human honks! But now we have to add construction worker yells to replace this category
6:20 "In all the road closure situations I put it in, it almost feels like it knows that it's not the right thing to do but always ends up just trying it anyway".... Wow.. so very much driving like a human. 🤣🤣🤣
Facts😂
Exactly what I was thinking. This is probably a topic where specific "staged" drives need to be recorded for the training data, because even the "good" drivers are sneaking around closed road signs. 😇
The best of the best. Car and driver.
Rare for this 80 old to say “It blew my mind!”
V12 truly is mind blowing. Every drive I have with it, it does something incredible. Thank you so much!
"Same thing I would have done" is the new benchmark . Incredible progress!
I know it's been said before but your camera setup and transitions are absolutely insane 💯🔥
Appreciate it! Takes a while but end result is worth it
Honestly, the driving in this video was incredible. There still a lot of progress to achieve, adding the missing bits like parking, backing up, how to deal with construction, etc. and especially if self-driving has to work across different parts of the world. But it feels like this is the way to do it. Thank you for the video!
Thanks for watching and leaving a comment!
Holy cow. There are so many terrible, scary af intersections in this town. The number of roads you have to blindly pull out on where traffic could easily be going 45 is insane. We don't need FSD, we need better road design...
And get rid/move signs that block your view, as well as shrubbery that blocks your view.
yes, this, and less car dependency in general so that roads don't have to be so wide and fast to serve the number of people that need to use them
I watch most of the FSD channels out there. The production quality (and by production quality, I mean all of it... the video quality, audio, narration of scenarios, your narration voice, the editing... ALL of it) is the best there is.
Just outstanding work. Thank you for this. As an FSD beta user since the first round of Safety Score adds, I for one, welcome our FSD 12 overlords.
Forgot to ask... You do a great job of running FSD 12 through ALL the scenarios. One that I haven't seen a lot of on your channel is traffic circles. Where I'm at, traffic circles are the bane of FSD. If you have any that are local to you, any chance you could take us along for the ride through a few of them? I know traffic circle density is highly variable according to geographic location, so if there aren't any local to you no worries; don't want you going cross country just for a few circles. : )
FSD causing the rest of the car to turn on AEB is hilarious
In short: Human-like...
Shows paint is not cycling infrastructure too lol
Internal Family Systems (IFS) theory suggests the human body is the same way. One part will be like "GO!!!", while another part (usually a protector) will stop us in our tracks.
It's hilariously bad
@@MiesvanderLippeYeah, not super thrilled with the idea of teslas cutting into bike lanes early like that.
Your videos are there most informative about FSD. Focusing on moments of interest and highlighting differences to previous versions. Thank you.
Appreciate that, thank you!
wow, love your enthusiasm for 12.3 and am beyond excited to try 12.3 for myself when I get the OTA.
It’s amazing, I can finally use FSD with my wife in the car
Yep, looking forward to my wife's reaction. She hates V-11 drives.@@AIDRIVR
i’ll say it again.
Best FSD videos on the internet
You put in a lot of work, it shows and is appreciated.
Great job bro!
You are Chinese
2:01 Choosing "the perfect moment to pull out" is TIGHT!
Not yielding to every vehicle is super easy, barely an inconvenience
😂
Babe wake up, ai driver dropped!
I was elbow deep in my wife and stopped the second I saw this notification. I have my priorities, she can take care of herself!
This is the fourth unrelated video with this as top comment
@@KCJbomberFTWI’ve never understood why people post stupid stuff like OP. I mean, is it for “likes” on the comment? If so, why? Is there something you gain from posting like-bait?
It’s always seemed so strange to me.
@@codycastDanny Mullen and told in stone both posted today and the top comment is this same thing
How much of a cult member do you need to be to act like this?
Wow! What a journey from what FSD was like in your early videos. Thanks taking us along for the rides.
It's definitely been quite the journey, thanks for sticking around Paul!!
Please make compilation of what it used to do and what it does now.
That would be great.
Always my favorite FSD channel!
I am so impressed with your narration and your focus on describing how the car handles every situation you gave it! Excellent job! First time I have watched your channel. You are easily the best FSD test driver on RUclips and the most interesting to “ride along with”. I completely agree with your assessments. This is the year! People who think we are 10 years away from FSD will be blown away by 2024. I don’t know how many more FSD test videos you will be doing, but I will be along for the ride on all of them. Subscribed. 👍
Appreciate the kind words flyshacker! Thanks for the sub :]
@@AIDRIVR I have to add that some FSD test drivers I have watched take their time to make a RUclips video, but then everyone in the car chatters about irrelevant topics while the car is dealing with drunk drivers and doing other amazing things under 12.3. Unbelievable.
The fact that it spotted the pedestrian around 3:00 was impressive; went back twice and still didn't see the person walking on the video
Not strange at all. Camera placement. He is filming from driver seat. The car has a camera in the middle. If the pedestrian was on the opposite side of the car in front, the car would not have seen it as early as the human in the drivers seat.
I don't think there's actually anyone there? When it first appears, we have a full view of the whole sidewalk going into that intersection and I can't see anyone. As he gets closer there's still an unobstructed view of the full width of the entrance the van is pulling into, until the last few seconds before the van stops, and there's still nobody visible.
If you back up the video a bit you can see the pedestrian walking, then the person gets obscured behind a parked car. The system has memory now and correctly infers that the pedestrian is going to continue walking across the road and slows down accordingly.
@@richardjones8395 No, it does not have any memory. But, the camera the car uses is placed more to the right compared to the camera that is filming the ride, and thus it can see things that are obscured by the video camera.
Note that as soon as they get too close to the car in front, the pedestrian is gone.
@@Xanthopteryx Yeah, sorry, I totally mixed up with another section of the video here (I was talking about another bit in the video at around 21:43 where a pedestrian was crossing and the car waited). In the particular case at 3:00, it does see the pedestrian due to camera positioning as you say. But going back to the statement about memory, I should have qualified what I mean by 'memory': as I understand it, V12 learns from short video clips (around 30 seconds maybe), so when learning what a human driver would do in a given scenario, the neural nets will be taking account of what happens during the period of the clip and how the human driver responds. Whilst perhaps not a ‘memory’ in the conventional sense, it does give it the capability for things like ‘remembering’ a moving car that gets occluded by another car at an intersection- as long as it's just a few seconds it'll keep track of it. On previous versions, the FSD team did talk specifically about giving the system a short memory to take account of such situations. I’m not as clear yet how that works with the ‘end to end’ system and whether the team have taken specific steps to give it ‘memory’, but as mentioned above, the fact that it learns from video clips does give it the capability to infer that type of information (at least over the duration of the clip, whatever length it may be).
I just got the FSD 12.3 update on my car. After watching this video, I am so itchy to take the car out for a spin! Thanks for the video!
Awesome, let me know what you think!
I watch a bunch of RUclipsrs for Tesla news,including FSD info, and i've been watching a few of your videos for a few months.
I really love your commentary, editing, and the way you have multiple cameras E.G. interior vs exterior, and how you transition between them to illustrate your points! Makes for great "flow" in your communication, and really easy for the viewer to follow along, thanks for being awesome :)
00:01 FSD using human data for AI is working well
01:58 Tesla FSD 12.3 executes complex maneuvers well.
05:59 Tesla FSD 12.3 faces challenges with road closures
08:03 Tesla FSD 12.3 navigates tricky intersections well
12:03 Tesla FSD 12.3 exhibits extremely human-like driving decisions.
14:03 Tesla FSD 12.3 faced challenges with U-turns
18:06 Tesla FSD exhibits advanced decision-making skills in complex driving scenarios
19:52 Tesla FSD 12.3 handles challenges beautifully
23:35 Tesla FSD 12.3 stops early for potential hazards
25:30 Tesla FSD 12.3 brings significant improvements
Bullshit Tesla car AI should be banned from public road traffic.. Is is is so dangerous for humans and animals too.
Excellent video and narration, as always.
Appreciate it!
I just went from 11.4.9 to 12.3. ALL of my pet peeves are gone! Late signaling, twitchy turns, incorrect changes into turn lanes, insanely fast RR crossings, all gone.
Even new capabilities. An aggressive driver quickly moved into a turn lane from behind. FSD just let it by and pulled in behind. While waiting for a left turn, a car pull up and blocked my view to the right. FSD crept forward to regain visibility. Waiting on a right turn traffic passing traffic blocked my view to the left. FSD remembered to wait for a last car to go by and continued the turn!
I need to completely reset my expectations to catch up to the progress of this release. Thanks for the effort to show it on video for all to see!!!
V12.3 looks super impressive. I sure hope I will get it soon. Great video, AIDRIVR's camera setup, videography, and commentary are the best. 😀
Once again you have produced the best FSD review video of the bunch. I like the voiceover style and how you analyze each situation. Very well done as usual!
Appreciate it Dave, thank you!
A combination of your video , chuck cook, Brad Ferguson and Rob Mauer’s latest v12.3 convinced me to buy more Tesla stocks!!
If I had any spare money I'd be buying too
Your mention about going back to 11 FSD for that one area is interesting. Would be interesting to be able to move back to previous versions to show the changes. The improvements are not so clear until comparisons are made.
The problem is that each time you drive the environment has changed completely. Cars are different, position of them are different, your own car is placed different, speed differ, the light differ, the weather differ, shadows differ, ...
The only way to truly compare is by doing like Waymo does:
Tons of computer simulations. You let the virtual car drive a virtual environment and then you can replay the EXACT environment with different software versions.
Can you focus on roundabouts in the next video? They shouldn't be a problem (finally) for V12, but I'm very curious. Great video anyway, thank you!
He mentioned right in the video that there are none near him other than that one mini roundabout he showed.
I get really excited every time I see you posted a new video and I have to watch it on the big screen tv right away... It's a profound understanding of the baby steps of human ingenuity towards safer roads... You could make 1 hour videos and I would watch it... I love your voice and comments... I wish you would go for 2160 instead of 1440. Thank you!
At 21:31, the reason FSD will be safer than a human is that human drivers exhibit numerous dangerous behaviors, which FSD will never exhibit, and which cause up to 90% of accidents by some estimates. These include texting while driving; other distracted driving; drunk and drug-impaired driving; road rage; fatigue; falling asleep at the wheel; medical emergencies; running red/yellow lights and stop signs; wrong-way driving; following too closely; unsafe lane changes; speeding; street racing; reckless driving; loss of control; and several other categories I probably haven't thought of. FSD can be significantly less skilled than an average driver, perhaps similar to a beginning student driver, and still show a much lower accident, injury, and fatality rate that human drivers, for this reason.
I can't wait for the day we get true FSD! This will give so much freedom to my disabled wife. But, it absolutely needs to be six sigma, requiring zero intervention from the human riding inside the car, who is incapable of helping out. But this will be a game changer for so many people!!!
Good for old people that shouldn't be driving.
sorry to say, but this will never happen
they are on it for 10 years now and it wouldn't pass a simple drivers test that any human with 12h of training could pass
hell, in the video you see at least 10 examples where it makes terrible decisions; blocks the road; ignores tight spots until breaking in panic etc or simply breaks traffic laws
For the first versions using end to end neural nets its capabilities are amazing. Just think how much improvement we might see after two years of trained neural nets!
AI driver is still number 1.
Thank you for excellent content.
Thank you for watching!
What was the year make and model of the car?
And in your experience do you think it makes a difference if it's the Intel or the AMD based computer?
@@AIDRIVR
Bullshit Tesla car AI should be banned from public road traffic.. Is is is so dangerous for humans and animals too.
@@KuramaKitsune1the intel and amd computers are just for the infotainment. The FSD computer is a separate Tesla designed system. For now both HW3 and HW4 perform the same with FSD
just got mine upgraded to 12.3 tonight will try it out tomorrow very exciting after seeing this video of what it's capable of doing.
Excellent - plus a great summary.
V12 has a very good understanding of the cars dimensions and can squeeze through spots (not sure if that's always a good thing).
Teaching my niece to drive (I had her go through cones and between boxes etc) and now watching people more, I notice a lot of people have poor spatial understanding and not as good as V12
even if probably legal, that maneuver at 1:20 where you're driving into the exit of the drive thru is something i would have disengaged for. both 'lanes' have arrows pointing in the oncoming direction and it looks like the lane widening for 2-way traffic past those arrows is for drivers exiting the drive thru, based on the road markings. which is probably why that pedestrian didn't look both ways, because the road markings are indicating one way traffic.
I was going to say the same thing.
Thank you for the excellent work!
Right back at you my friend
GOOD JOB 👍🏾 Your visuals allow me to understand your observations & comment on the car's decisions.
Great video as usual so informative and entertaining.
IMHO the thing that needs improving next is the navigation data. When a road is closed that contradicts the data, a new route needs to be quickly determined. Also, the closure could be updated to other drivers in the area with a few hours of persistence.
Simple: rerouting is another type of AI action. Find clips of people driving with road closures in them, not necessarily in the route path, then edit the route to go through the road closure, and add the reroute action as the target action. Lots of issues can be solved by creative data sampling and modification like this. You can do the same thing for the issue where the car is unable to safely follow the route. This stuff will be solved so fast.
very interesting to see major improvements on fsd compared to early builds looking forward to more progress 😊
Crossing solid white line is legal in a lot of states (California and Washington for sure) as long as it's safe to do so.
Yet, even some of those may not allow lane changes within 100 feet of an intersection.
The one thing I'm learning from these videos is that FSD really needs a way of detecting "road closed" signs and feeding that into the navigation system, like maybe even just ignoring the navigation and picking a different direction to go instead, and letting the gps reroute to the destination afterwards.
Also, while I really don't like tesla as a company (mostly due to the owner of said company), and self driving cars still seem like an accident waiting to happen as soon as supervision isn't legally required (not to mention that unsupervised self driving cars solve a problem that would be much better served by widely available public transit), I can't deny that this is really cool tech, I'm kinda shocked how much this system manages to do correctly, even if it's definitely still at a point where intervention needs to be available at all times, it can drive better than most self driving tech I've yet seen.
11:08 I think 2024 will be the year when we have fully self driving cars tbh
the video literally came out 6 mins ago how are you there
Yeah. Probably. Or 2025. latest 2026.
@@tav9755Or someday. Maybe.
@@halophobic9550 He skipped a part
Or played the video at 2x speed
I know the way you edit takes a lot of work but your approach - showing the boom-deleted car when appropriate, and the voice-over makes for a really useful video. Many thanks.
This was a wonderful demonstration. Thank you and best wishes
Great graphics & great commentary - Thank you from London UK. 333am GMT.
Still the best!! Will have to cook up some fsd vs human scenarios to make it interesting pretty soon.
Peace.. Shalom.. Salam.. Namaste and Thank You Everybody for All that you are doing to Heal our Mother Earth 🙏🏻 😊 ✌ ☮ ❤
I got 12.3 this morning, coming from 11.4.9 🥳. They gotta get speed control in order 😅. Other than that it is a great leap forward as far as I can tell.
Meh, crossing a single solid white line is permitted. It just means to proceed with extra caution.
Still the king of FSD videos, thanks for the upload.
Thank you for watching and for the compliment!
the irony of me exhausted from driving for hours today, then after getting home and taking a nap, watching this video start to end
Another great drive. AI Diver really puts out quality content. By far my favorite.
Great video, as usual. I'm really impressed by the improvements you've shown from FSD bets 12.2.1 to 12.3. I sure hope this starts rolling out to a much wider group of FSD beta testers soon so that we can get it on both of our Teslas.
Now we need reversing it seems
I really do like these videos. People watching can learn to maybe drive better by example. This system doesn't get road rage.
Being a professional driver for close to twenty years has taught me a lot of patience.
I hope to continue watching videos of yours more ahead because so many people have a fuse set to road rage. If anything, it can be a great example of how to drive well.🙏
3:40 I think that should've been a critical disengagement, especially since the car is doubting itself
I only count them as critical if it was in any real danger. The car just stopped itself and needed me to re-engage, no danger to itself or anyone around
@@AIDRIVR It was kind of dangerous because just sitting in the road doing nothing is like, not very good.
I like that you let FSD v12 do its thing to the very end instead of disengaging early. Gives us greater insight into what the car is trying to do/what its thinking.
4:12. I think it hadn’t accounted fully for the second trash can initially.
Quite a fresh driver here, got my license in mid-February and I have got to say, that Tesla FSD drives similarly to me, so hopefully we will improve together xd
love your videos very entertaining and i enjoy the long ones especially
Appreciate the kind words! They take a while to make but comments like this make it worth it
I just got version 12.3 on my car last night. I drove around a bunch today around Austin, TX and the car did AMAZING.
Thanks!
I really appreciate that your video not only informs but is so visually pleasing to watch. Well done.
once FSD can read and interpret signs...
22:00 when the car slows down when it sees the light change to yellow, but then changes its mind and accelerates is mindblowing! It really shows signs of intelligence.
Data from human drivers in the fleet training the model was inevitable.
At least it didn't gleefully send the trash can flying!
Not only inevitable - that's Tesla intentional strategy.
That's a good thing. It doesn't mean the car will make mistakes like humans necessarily. The car is 100% attentive at all times (which is a large source of human error) and the training set is curated good driving.
13:23 it’s actually illegal to block an entrance like that and you are suppose to leave it opened for traffic to flow
When you look at crash compilations, it's most of the time because of either reckless driving or fatigue (microsleep, general distraction). This version shows no signs of doing mistakes like this and thus I would feel much safer knowing the cars around me are AI driven and don't just cut my lane, drunkenly swerve around or overlook pedestrians.
I think about this all the time on the highway, especially when someone is tailgating me and glancing down at their phone. Using FSD lets me watch other drivers around more and it’s terrifying
I drove my Tesla with the latest update and it was incredible, no intervention at all. Way to go Elon.
LETS GOOO!!!
Best FSD Tester on YT!
Ive been using 13.2.1 on my Model 3 and I have to say.. it's a tons more smooth and confident-inducing than when I tried it last time
It SUCKS that I paid full price for FSD and get these builds months after others!
Previous owner of my car paid full price for it too, and I haven't received any build of FSD whatsoever yet.
I'm sorry no one told you John but we took a vote and decided that due to your impatience, you are being slow walked on FSD. Perhaps a better attitude would help your case ...
@@rmkep I snuck in and voted multiple times to deny John FSD.
@@rmkep You’re amazing! And dense!
Got my 12.3 yesterday night and drove it today.. really good improvement.. one thing I could notice is that lane centering is not perfect .. it’s hugging the left side a lot more than 11
Extremely dangerous and irresponsible technology
How so?
After watching in awe, I've finally gotten the v12 update tonight. Love living vicariously through your videos with all the updates and now I can live it myself haha. I'm pretty cautious with FSD in general and don't actually use it much but am excited again for this update.
21:30 It's not like human drivers are bad at their core. We do that shit well, until we get distracted, inattentive, tired, impatient, reckless, ...
So an AI, with its infinite attention span and patience, that is trained only on the best behavior should in theory be among the safest drivers out there.
That theory is yet to be proven, so far all evidence suggests it wont
Great video. Absolutely love the visual effects etc.
Thank you, appreciate that! Takes a while but comments like yours make it worth it
(2:29 ) Just because the self-driving car needs to make a left turn immediately after the right turn doesn't mean it should put the left turn signal on before doing the right turn. The proper way is to signal right first and turn right. When the car is about 45 degree pointing to the right, change the turn signal to the left for moving to the left and for left turn at the traffic light.
Improper use of signal is a traffic offense.
It's ok, the driver will happily pay the fine for the offence the AI made. An AI that is 100% property of Tesla, trained by Tesla, the customer only has a limited license to use.
@@_BangDroid_The driver is in control of the motor vehicle. They choose when the AI is driving and when they take control. The human should always be responsible for the AI’s actions until the company says that it’s safe to not actively monitor the AI, at which point the company is responsible for
@@_BangDroid_ Poor driver getting fined for mistake that AI makes.
Remarkable it’s come such a long way in such a short period of time.
Most people enjoy driving their cars themselves. Never once have I thought I need it to drive itself whilst I have to hope it does not mess up.
Regarding the moment at 16:20 where it crept up on the guy who didn’t start going when the light turned green, basic Autopilot actually does the same thing! I’ve noticed for several months now that base Autopilot will release the brakes and start rolling up on the car in front of it if it can see the light turn green. It’s really cool to see.
hate when people do that to me . Why not honk ?
In Texas it is legal to cross not just a solid white line but even a double yellow line if you are turning in to a driveway.
Awesome video. Thanks for producing. So nicely done. Great graphics
Thank you for all the vids. The government has the final say but the following is my pov. I am not a numbers guy but insurance companies are. When they get the numbers that FSD saves money and life's the pressure will be on not only government but ALL auto manufacturers to license this tech.
I'm puzzled why some features have still not been implemented.
1. Road Closed training: it would be straightforward to sample incidences where the planned route was changed by human drivers deviating from navigation and where there's 'road closed' text (there will be vanishingly few examples of people literally driving up to the sign before responding!)
2. Dead ends etc: trivially easy to find and train on instances where the driver selects reverse (esp repeatedly for a multi point turn) and then proceeds in the opposite direction to that before reverse was selected.
Having said that, should we be expecting a much faster rate of improvement from now on?
Cheers dude, literally top content, as ever. Hoping the new eqpt is making your day a lot easier.
I think it's not wise to only count on visual queues. If waze (or Tesla's version of crowd-sourced navigation) tells you the road is closed, you should have come around anyway...
It is so stupid that it doesn't know how to properly use navigation. If it used any mapping program for routes it would have calculated from the start that route wasn't an option. Heck even my 2013 car does a better job at that.
Excellent video. Probably one of the first of yours I've watched but really did an awesome job of showing what this new version is capable of.
Just experienced FSD 12.3 in a HW 3 Model 3 yesterday. Car was going through a merge with cones on right hand side, the construction vehicle on the right let out a huge bloom of steam and FSD seems to be blinded and NOT avoiding the cone. The driver manually took over and swerved, just clipping the cone or it would have plowed through the cones. Vision only system just will not cover the rare cases that it's blinded. Same with fog.
You do an excellent job ! Many thanks.
V12 is nutty coming from v11. Especially in my area. Gawd dayummm it’s like a human driving the car. Especially with turning into side streets and unprotected left turns.
21:44 pedestrian was perfectly visible for car, it's funny that non end2end display related neural network did lose that pedestrian later, but end2end did not lose it and tracked, so in this situation it was ok.
17:44 it's actually example why current approarch will not work (at least not as safe as waymo or mobileye one) the problem here is not car turning wrong way on road, it's the attempt to drive over static obstacle, i should point out that 12.2 did try to drive over pedestrian same way.
while 12.3 does have safety net (4:18) it far from being good enough compared to waymo and mobileye, i.e. in some situations it will be too late for that, and waymo and mobileye safety net does not result in automation disengagement, just limiting neural nets not to do unsafe actions.
p.s. i do understand that safety net will result in more robotic behavior (like 11) but it would be much much safer, and obviously there should be a way for a driver to temporary disable some limitation - for example to allow car to cross solid line, but extreme attention for human driver is required at such moments.
p.s. waymo do drive without serious risk to human life and without human overwatch(there is no way to stop immediately for any human) for several years now, but it less smart than fsd 12.
It'll be interesting to see how they address variations in traffic control from state to state. Based on what you said, in California, a red right arrow is the same as no right turn on red. But I live in Oregon and here, a red arrow simply indicates a mandatory turning direction and legal turns on red are allowed unless there is a NO TURN ON RED sign.
Great video! Hope the next models get cameras on all four corners to better manage intersections and backing out of tight spaces. Seems to be room for improvement there and it is unlikely they can do much when there is no visibility.
Can we just appreciate how much editing effort is put into these videos 👏
13:10 In those situations, you really shouldn’t leave a gap. It can be very dangerous because cars in the other lanes can’t see them and if they attempt to turn past your open lane while you are stopped cars to your left can’t see them and could cause a collision