Maybe I need a flip off counter? Sponsored: Use my special link www.privateinternetaccess.com/AIDRIVR for an 83% discount, 4 months free and 30-day money-back guarantee on the best VPN - Private Internet Access
Man i appreciate the effort you take for your videos so much! Recording the center display and cropping each part out like the Indicators as well as speed, navigation and the text "creeping forward for visibility" just so you can put them in the bottom, as the Model S did! It looks so perfect. Actually had to check a few times if that really was the Model Y, because was so unsure. Well done! It is so awesome to see your passion for what you do! Loving every video! Keep it up :D
I'm suffering from ADHD, muscle spasms and sleeplessness. Your videos and especially the humor in your voice are so smoothing that I sometimes watch videos again just to smooth down. Maybe not the reason why you started that channel but for me the reason to watch it. 😄
It's fascinating to see the progress. I remember not long ago most of the time it was being disengaged for going too fast, etc. and now it's down to these edge cases where it just needs a bit more assertiveness and nuance. Amazing to experience vicariously. Keep up the great work.
10:20 The car bailed and turned left because the road ahead LOOKS like a one way street. Proof the neural nets are actually working. It's actually impressive.
You continue to be the most informative and balanced FSD reviewer, hands down! FSD RU-vidrs are a dime a dozen, and the big names quickly trend to the uber bull / shill end of the spectrum. But you present a balanced view, and a deeper technical view. Finally someone discussing the end of the occupancy network, and the language token routing model shown at the last AI day. I suspect you have a tech background, plus high EQ, leading to a really good perspective for a reviewer of this tech. v12 is truly a big change, and it terrifies me how much is thrown away in the process, but has huge promise at the same time. I can't wait for you to get your hands on it!
The turn signal problem is definitely worse than no turn signals. There has been so many times where my car has perversely signaled "right turn" because the road widens right before an intersection, but it is not actually going to turn. When there is a car waiting on the cross street they could reasonably take the signal at face value and start to go into my right of way. I usually quickly hit the stalk making it signal "left turn" making it (more safely) look like a crazy person is driving. I know I tend to believe other drivers more when they start their signal right before an intersection, so it's worrisome.
I really like how the visualizations are displayed, and the camera behind the car. This Model Y recording setup looks even better than the previous AI Driver Model S setup. With HW4, it will be interesting to see how FSD progresses on this channel.
Calling it "generalized artificial intelligence" (1:25) is inaccurate. What separates these attention based models from generalized AI is its ability to exceed its training, i.e. the car starting to exhibit behaviors outside of the driving it has been trained to do. IMO that isn't desirable, but it bothers me a bit when people misuse this established terminology in order to overhype their tech on a certain social media platform. Side note - love your videos as somebody who doesn't want to drop the cash on FSD but loves to see this tech progress.
I'm really glad you're back making these videos with your new hardware! You deserve more views for sure. People NEED to see what AI driving can actually do. We need more acceptance of this technology as soon as possible - it will save lives. Even though it still behaves way too much like a beginner in their first driving lesson lol
Wow. The traffic there is nuts. The people parking all over the place making it barely a single lane is crazy. And OMG why so many people think it's OK to park immediately in front of a stop sign?! I couldn't live there. I would have a nervous breakdown.
@20:43 I have the impression that it calculates a "pinch point" with the possibility of a parked car driver opening their door, so the path planning has to leave a few feet of room for that unlikely event. That may be why you're seeing hard braking for pinch points that would actually have plenty of room to drive through.
Every 6 months or so I check in to see how it's improving, I kinda feel like things are moving quite slowly, meaning 3-10 years until Tesla FSD is at human-like driving.
To my understanding, Occupancy network is just a part of vision model, it is a neural net system, and it is still there in v12. v12 mostly replaces the control part from manual code to net, and it's now all net, and can be trained end to end, which is huge. anyway, great video!
Is the control neural net in v12 making decisions from vector space data? Or directly from video footage? If it's using video footage, does that mean it's replaced the vision system?
@@spiderbite7849 I don't totally understand what you've asked, but (as a computer guy) I might be able to clarify that the v12 system uses video footage as its training input rather than hard coded rules (i.e. If -this happens- then -do such and such-.). So they thrown out those rules that are in v11 and given the machine learning algorithm oodles of examples of good driving behavior in the hopes it will generalize and generate the right "rules" on its own. The drawback is you don't really know how it's going to apply its learning/decisions until it has been presented with a novel or new situation. The drawback is also an advantage as it's more likely to respond after the fashion of the great examples it has been given (i.e. better than the average driver would respond). Neither the hard coded rules or the oodles of examples can anticipate all situations. If the machine learning screws up, you give it more examples of the specific behavior you'd like. It's similar to trusting a person to make decisions based on their experience versus writing a law for every situation they might encounter, if the person's experience was limited to only having done the "right" thing in the experiences they have had.
Ok thank you, I understand that v12 using a neural net to make desicions. But I'm confused as to why it's trained on raw video data, and not on the 3D vector space visulisation. So previously v11 used a vision neural net for recognising objects and placing them in 3D vector space space. Once everything was in vector space, the decisions were made from hard coded rules. But now, it seems v12 skips the object recognition vision>vector space part. If v12 runs a red light, how do we know if if vision failed or if control failed. In other words, did it not detect the red light (due to rain, haze, angle) or did it detect the red light, but failed in it's decision (due to insufficient training). In V11 this is easy to distinguish, but if v12 is video data in and controls out, I am not sure if this can be distinguished anymore. @@carabaoboy
@@carabaoboybut I think it is still doing steps for it's training. When those super computers get the video they still do all the labeling and occupancy network and then train the behave on that "translated" world.
Regarding the stop sign at 17:15. I don't actually think the software registers an upside down stop sign as a stop sign. If you look at applications that do facial recognition, they usually fail if you flip the camera upside down, as they don't actually understand concepts such as faces, only patterns such as eye, nose, ear, and mouth placements.
There must be some benefit for Tesla and AI V12 team to keep releasing V11 versions, so yes please keep recording more videos on 11. It’s still very interesting to watch what changes. Also showing the differences to HW3 and 4. Maybe at some point you’ll say, now its better.
Correct me if I'm wrong but I don't think they're getting rid of all the occupancy networks and stuff they did before on perception, the only thing they're replacing now is control which was the remaining coded stuff. I'm pretty sure the old perception networks are still in there but now just able to be trained end to end with the rest of the model
I was kind of thinking the same for the occupancy netwrok but then again we know there is no more lebeling happening. The car doesn t know what a stop sign is in v12, or a round about etc. If that is the case, and it is, then why have the occupancy network. The car will learn from the video training what is drivable space and what is not.
@@lucadellasciucca967 the occupancy network also helps with just knowing if there is an object in the space, even if it is not recognized. Presumably v12 will understand “don’t hit stuff” but maybe the occupancy network will still be needed
9:48 - my car does this on FSD beta 11.4.4 sometimes. When I end a trip it will see a parked car and alert me to take over thinking the car is headed toward me. Engages auto steer and alerts.
Winter is coming on in NW Wisconsin. FSD has, thus far, been totally unacceptable as soon as the snow falls. I'm doubting that I'll trust it this winter because the car does not completely brake before going into turns. Doing so is absolutely critical in the snow and important, but not as much so, in wet weather.
Here in Missouri, we have LOTS of roundabouts, and I have to disable FSD when I enter them. I have HW3 and I don't think the processor has the horsepower to handle 2 lane roundabouts it slows down, and folks honk at me and have LOTS of trouble figuring out if the oncoming traffic is going straight through or following the circle - I've been in many crashes where it guessed wrong. It just is too slow to figure out what the traffic is doing.
Communities are making laws that you have to put on your turn signals at least 300ft before your turn. If laws like that are part of FSD it can cause confusion in managing turn signals.
My two biggest problems with 11.4.7 has been the random stops during turns, or at random times. Not sure why this happens but its very frustrating and happens a lot, but i report each one. and then the problem with the windshield wipers going off even when there's no rain....that's extremely annoying also. i hope they fix that quick.
Interesting, as someone who lives in the UK it is nice to see people use full self driving on roads like this which are a lot more similar to the roads I drive on. Most videos seem to be on big wide American roads which we don't have much off here, which has always let me to be more of a self driving sceptic, but it handled some of the complex situations fairly well.
Tesla's "Autopilot" will never become available in Europe as it's technology fails even the lowest safety-expectations. A system that is SOLEY based on cameras will never achieve a statisfying reliability. It doesn't matter what the algorythms and 'AI' do with the data from the cameras since this data is inherently extremely unreliable. If Musk ever wants to sell "self driving" cars in Europe, it is absolute ESSENTIAL that he scraps his "LIDAR bad, Cameras good" attitude. Without a multitude of sensors that must either confirm or deny the cameras interpretation of information, self-driving will stay an american thing. Which I'm quite thankfull for. I have zero interest in sharing the street with a bunch of self-driving cars that rely solely on notoriously terrible video and picture interpretation, which can't even correctly distinguish between colors on a mildly sunny day. Nooo thanks!
Until v12 lands, in about "two weeks..." lol, I think it could be very instructive to see a video of a hardware 3 car and a hardware 4 car doing the exact same route. Keep up the good work! Love the videos.
@20:20 look at the red leaves just above the top of the red truck... it does kind of look like there is a stop sign behind leaves there (just barely). Just speculating on why the car stopped. Interesting that it didn't show a stop sign on its display.
I wouldn't worry about it taking a long time to train HW4, once Tesla figures out the most efficient way to do it with HW3 + they will have way way more computing power by then, training HW4 will happen very quickly
I think that when they have more compute power they will increase the data set, to the point of maxing out the available compute power. That is, I expect they will always choose bigger computation, for the purpose of capability, rather than having the compute process take less time.
@@cleon_teunissen yeah but why did HW3 take such a long time to compute? It wasn't that they were lacking data. They didn't have the computing power and they also didn't know how to make the AI algorithms, now that they do have the computing power with Dojo and they know the code cause HW3, HW4 should come quicker than people expect
Welcome back! For about 2 months I've seen dirtytesla FSD videos in my recommended and got excited it was you, but alas, no. I didn't drink that off brand diet coke though, I stayed faithful
😊Amazed at your patience with the driver who gave you the bird. If you were not on the camera and beta testing would you have been so kind with a driver who was so rude?
You’re hilarious with this helmet man😭 last time I think you rick rolled us😭 job well done. Don’t know if you’ll ever show and it’s funny. I bet you go to these Tesla events and can go unnoticed 😂
Great video!!! It was definitely worth waiting. Cup of tea and big screen with AI DRIVR on ❤. All the details, like the disengagements counter and all other info on the screen is great.
I don't think they abandoned the occupancy network and other neural nets in FSD 12. They just added more neural nets in place of handwritten logic. Like in human brain, there are specialized structures for specific tasks - not all neurons are the same.
The info I'm seeing is that v12 is end-to-end, one NN. Photons in, vehicle controls out. There will likely continue to be a separate (V11) perception neural net onboard for the on-screen visualizations.
17:46 was the white car's fault for stopping the the middle of the road so that the blue car couldn't turn in, yet they get irritated and flip the bird 🙄
Okay, so I think interactions like stopping at a stop sign for constructions will work properly in v12.. I do think the Honda driver flipping the bird to you was very unnecessary since she nearly literally crashed into you 😂
Looking at that bailed intersection I noticed that STOP seems to have been written in the middle of the road which it may have understood as it being a one way road and may have bailed to avoid oncoming traffic.
If Tesla are purely relying on NN for both HW3 and HW4, it makes sense to me that they would “normalise” both sets of incoming data (that is colour correction and upscaling/downscaling) prior to training and executing on NN, otherwise they are going to need two sets of NN. I think after awhile, they will take advantage of the better hardware over time, but it’ll be lowest common denominator at the start to they to have a smoother transition
Some people make a fetish out of putting their turn signals on early. Turning your indicator on two or more intersections before you turn is nuts and yet people swear by it. I don't understand some people.
Yeah, and some does it too late... or never. Just remember why we're supposed to use them: we show what we're *thinking* about doing, so that our fellow road users can act if we do do not act correctly. You know; for safety. The earlier we can show our planning in an unambiguous (!) manner, the more time they have to use that information. Also, it's good for efficiency. But yes, the signalling should not leave any doubt as to what you're thinking about doing, so don't do it before it is unambiguous.
I paid $104 to subscribe to full self driving in my enhanced auto pilot model Y performance for two days. I was told even if I canceled I would get one month before it shut off but it shuts off in two days on November 17. I drove 8 miles before my first Shut down and red screen of death indicating full self driving would not be available for the remainder of the drive. After a reset stop, I made it 4 miles on the highway before the screen reappeared. Full self driving, at least in the model wide performance is not ready for prime time let alone $104 per month subscription.
im not a tesla engineer or something but my guess is that the visualization is being generated from an object classification ai but the driving ai gets the classified object data with the occupancy networks output. while object classification can give more false positives or just doesnt get it, occupancy network does a better job at defining if the road is obstructed consistently. the gray thing you see on the road edges or in roundabouts, is the occupancy network showing that car knows it cant drive there
I think that's why Tesla deleted the stalk and moved the turn signal buttons to the steering wheel, to bring human signalling in line with v11 signalling. :)
I've been thinking about car learning from human. It can be great but also it can be awful, especially in narrow streets, because humans like to take unnecessary risks, swirl on the road, not stopping on Stop signs, go through yellows and early red lights, drive a lot over the speed limit. Hopefully all of this will be take into account. My biggest "fear" are stop signs, as I can't remember when I saw someone stop on the Stop sign. Great video as always
One thing to note with the indicator signals is: what if the neural net is learning from successful indicators from other vehicles in order to validate responses
2 prius in the same minute of video for a total of 3 in one parking lot. more than I could keep count of by the end of the video. Can't wait to see how much better this gets!!
The moment I stopped using FSD was when an unnecessary turn signal confused a driver turning onto my street and I almost hit them. If I had, they would have full justification for blaming me. I'll have to subscribe to see when it's worth it to start using again.
My observation with 3 car is turn signal coming on when passing turn lane, also when enteringb a vnew roundabout it has no clue, finally on 2 lane highway the car does not know what to do with passing lane
V12 will have nothing to do with this. Hope Tesla stops wasting time & money on old style FSD. Hopefully you will get a 11.9 ( end to end Neutral net) Beta to test soon. I can’t wait.
My thinking, Tesla will need to keep both the neural net FSD and the programmed FSD. Both will supervise each other. Without the supervision of the programmed FSD you can't monitor neural net version or make higher level decisions.
Can’t wait for v12! I don’t care if its less reliable or makes more mistakes at first. It should be drastically more human like, with have more common sense. It should understand hand gestures, have smarter planning of lane changes, show more cooperative use of space, add parking lots and fix your turn signals. - Hopefully
the 10:20 crossroad turning left is probably due to NTHSA's regulated rule hardwired to force going in the direction of your marked lane. I bet you were on the "left turning" lane
I would love to see FSD attempt a driving test and see if it would pass. Don't know how they work in the US but in the UK you can have up to 15 minors (no more than 3 of the same type) and no majors in order to pass. Would be interesting to see how FSD would fare.
The key issue to self driving is liability law. The liability needs to be moved from buyer to the manufacturer. Without such liability shift for accidents caused by self-driving software manufacturers will keep trying to push half baked technology into our streets.