Tesla’s Real FSD Breakthrough Isn’t What You Think--FSD 14.3.3

本站添加: 2026-05-25

[00:00:00]Hey y'all, it's Dr. Knowit All. So, I've got some really interesting news from Tesla. Their full self-driving 14.3 software just came out. We still haven't gotten access on our Model Y or our Cybertruck yet. Hopefully, we will get that soon and then I can do a beta test drive. But anyway, in the meantime, a big thank you to Chuck Cook for posting the release notes to 14.3. And even though it's only a dot dot release, they actually have some new release notes, which is very, very cool. And of course, a big thanks to Chuck for actually going to work and going flying this week because Tesla seems to love to drop these new releases right when Chuck is on the road and he can't test drive the system. Anyway, thank you, Chuck, for working so that we get the updates. That seems to be a connected thing in the universe. All joking aside, most of the release notes in 14.3 are the same as 14.3.2. I just went back and checked a few minutes ago, but there are some new things and they are very important.

[00:00:51]So, I'm going to do a super quick recap of what the previous notes were. If you haven't seen that original video, I will leave it up here and at the end of this video so you can check it out. I go into a lot more detail there, of course, but I want to get to the new ones as quickly as possible. Anyway, you can see this is 2026.14.6.6 and FSD supervised 14.3. Full self-driving supervised includes upgraded the reinforcement learning stage of training the FSD neural network resulting in improvements in a wide variety of driving scenarios.

[00:01:20]Reinforcement learning of course is gamifying things for the neural network.

[00:01:24]The neural network gets better by attempting answers most often in simulation of course. So you give it a scenario, it gives you what the answer is. And by giving the answer, what I mean of course is turning the steering wheel, pressing the accelerator or pressing the brake, something like that.

[00:01:37]That's the answer. and then it grades itself against what the reality was or what happens in a driving simulation type environment. That's reinforcement learning. They upgraded that as they said, which results in improvements in a wide variety of driving scenarios. They also upgraded the neural network vision encoder, improving understanding in rare and low visibility scenarios. In other words, nighttime things like that.

[00:01:57]Strengthening 3D geometry understanding and expanding traffic sign understanding. That all is very important as well. In other words, it's learning how to read better. Next up, and in my previous video, I go into a lot of detail about this because this was the key upgrade for 14.3.2, rewrote the AI compiler and runtime from the ground up. In other words, they created a new compiler for this with MLIR, resulting in 20% faster reaction time and improving model iteration speed. In other words, they're able to improve the model more quickly, which means that we should see updates more rapidly. There was a really big gap between 14.3.1, I believe, and 14.3.2. too. But the primary reason for that was creating this new compiler. And while 20% faster reaction time is great, that is a huge advantage. The really big advantage now is that they're going to be able to iterate and create new models faster and faster because they have this new compiler. Before we continue, a quick shout out to my channel sponsor, Joah.

[00:02:51]They make the best accessories for your Tesla and other EVs, and they have amazing warranties and customer service, too. Check out the link in the description and grab a Tesla roof shade, a fan cooled phone charger, a portable tire inflator, a foldout lap table, and so much more. Oh, and they make perfect gifts for your EV loving friends, too.

[00:03:07]Be sure to check out the link to save 5% off your Joa order. And now, let's get back to it. Next up, I think was one that was designed specifically for the Cybertruck. Mitigated unnecessary lane biasing and minor tailgating behaviors.

[00:03:19]Specifically for our Cybert truck, it loves to hug the left hand side of the lane sometimes, like right on the line or something, which feels very, very scary because you want to be centered in that. And this update corrected a lot of that behavior. Next up, and I'm going to speedrun these things pretty quickly.

[00:03:34]Increased decisiveness of parking spot selection and maneuvering. Improved parking location pin prediction now shown on a map with a P icon, which actually does happen once in a while.

[00:03:43]It's not very consistent, but it does happen. Enhanced response to emergency vehicles, school buses, right-of-way violators, and other rare vehicles. I have definitely noticed emergency vehicles and school buses. It's much better with that so far. Improved handling of small animals by focusing reinforcement learning training on harder examples and adding rewards for better proactive safety. I've definitely noticed that as well to the extent that we actually had a really cool intervention. I will try to overlay that over the top of this. We were coming off the highway at about, you know, 50 55 miles an hour, something like that, and a cat ran across the exit right in front of us. And by the time I had actually seen the cat, registered it, and put my foot on the brake, and I did disengage just because it all happened in an instant, and I didn't realize that the car was already doing it, but the car was already breaking, and had safely avoided the cat before I was even able to step on the brake. It probably saved that cat's life. That's pretty remarkable. So, I definitely noticed that improvement in 14.3.2. Next up is improved traffic light handling at complex intersections with compound lights. In other words, there could be series of lights and things like that.

[00:04:43]Curved roads and yellow light stopping and yellow light stopping. I've noticed it's actually slightly more conservative than it used to be with yellow light stopping driven by training on hard reinforcement learning examples sourced from the Tesla fleet. So again, I think that's the third time they've mentioned reinforcement learning in these notes.

[00:04:58]Reinforcement learning is super important for them and they've obviously improved things a lot on that front.

[00:05:03]Next up is improved handling for rare and unusual objects extending, hanging, or leaning into the vehicle path by sourcing infrequent events from the fleet. That would be something like a tree branch hanging over or maybe a wire or something like that. Something that's not on the ground but up in the air. And I definitely have noticed that we have trees that kind of grow over the road sometimes and I've noticed that the vehicle is able to very easily get out of the way of those now, whereas it used to often drive through those branches and you get that nasty little scraping sound as the branch goes across your windshield or something. So, it's very nice that they've adjusted for that. And then the last older note is improved handling of temporary system degradation by maintaining control and automatically recovering without driver intervention, reducing unnecessary disengagements. And what that means, of course, is if there's a failure or something, if something goes wrong with the full self-driving system, it autocorrects. It gracefully fails. Instead of throwing up that red steering wheel thing and beeping at you really, really loud, it's actually able to take care of it itself.

[00:06:00]So now we get to the four new notes that they have added to this version 14.3 as opposed to 14.3.2. One of them actually came out during 14.3.2, but they hadn't added a note at that point. Anyway, the first and really important part here is unified the model between actually smart summon, full self-driving, and robo taxi for more capable and reliable behavior.

[00:06:21]So what does that mean? That means that actually smart summon now is part of the full self-driving stack. If you don't know, historically Tesla had a highway stack, a city stack, a parking stack, and an actually smart summon stack.

[00:06:32]Those were all separate stacks. They have been putting them together one at a time, and actually smart summon was the last thing that was still sort of hanging out there doing its own thing.

[00:06:40]So, this is actually a really big upgrade that all of these are now part of the same stack. And also that full self-driving that we consumers get is now part of the same stack as the robo taxi stack, which means they're able to develop all of them at the same time and not have to split their resources. I expect that means we will get substantially faster iterations of consumer level full self-driving going into the future and we will get all the latest new features which is also super cool.

[00:07:07]Next up is a biggie and I think they actually listened to us because all of us were complaining about actually smart summon only going 6 m an hour. It was very very slow. They actually increase the speed to 8 m an hour now and that should actually make it feel a lot better. And the couple of people that I've seen online that have tested it said it not only goes faster, but it also appears to be more confident. It's not as sort of creeping around kind of apologizing for its existence kind of thing. It's actually willing to go out there and drive in a parking lot. If you don't know what actually smart summon is, it's basically if your car is in a parking spot and you're at the exit of the store like Walmart or Costco or whatever, you can just push the button and the car comes to you and you can load the car up with your groceries and then you can go on home. So, you don't have to walk all the way out to your car. I personally never use it unless it's a rainstorm or something like that and I don't want to have to go out and get really wet in the rain. In that case, I will actually use it. But for the most part, I don't. But of course, the better that actually smart summon gets, the more frequently I will probably use it. So that of course is nice, but the really important part is that it's now part of the same stack as full self-driving itself. And that actually is a big unlock for Tesla.

[00:08:13]Next up, and we actually saw this in 14.3.2 two iterations. You can now help Tesla improve self-driving. In fact, you have to help them improve self-driving by selecting an intervention reason on the main screen after taking over. Now, in theory, this is a great idea, but I think that the implementation, it was really bad when they first started cuz it was a non-modal dialogue box that appeared in front of your navigation and you had to interact with it and it was like taking up a lot of space and it was very very annoying. And the microphone UI was in a different place. In later versions, they've moved it down into the bottom lefthand corner, which is significantly better. And of course, you can interact with the full self-driving and with the navigation and stuff like that. While the dialogue is still up there, but there could be improvements.

[00:08:53]Number one, there's only four options and other is not one of them. It's navigation, preference, something else I'm not remembering right now, and critical, but I think they should also include other options like speed and other so that you can explain what it is. And they should also have an explicit closure box that if you don't want to interact, you can just close it by hitting an X. I did a video on that if you want to see that. I will leave it at the end of this video as well. But anyway, it's quite clear that Tesla wants very very specific feedback for their full self-driving system. And they would prefer to tag the interventions with very specific data. I just don't think it's granular enough. And I think that that's going to cause them problems cuz people are just going to push a button and not really care what the actual reason is. And by the way, Dr. Scott Walter found out that there is a way to dismiss that dialogue box without interacting with it. You just have to push the microphone button twice in rapid succession and it will go away.

[00:09:41]So, there's a little hint for you. If you don't want to actually interact with it, you can just do that instead.

[00:09:48]And then finally, and this is a big one with 14.3 and it's actually causing a little bit of controversy out there in the interwebs, you can now see distance traveled in full self-driving supervised without an intervention. In other words, they're gifying how far you go without intervening. The self-driving app will also show you your longest intervention-free streak. So, is gamifying this a good idea? Yeah. I mean, people love to get their gold stars. People love to play games. You like the endorphin rush of going like, you've traveled a 100 miles without intervening, things like that. But, as Chuck Cook says, there is a definite downside to this. And the downside is that you might need to intervene. There might be a situation where you're like, "Oh, boy. I need to intervene right here." but you're also like, I've got 125 miles since my last intervention. I don't want to do it. And that hesitation for that one or two seconds or something like that, could cause you to actually get into an accident or a bad situation that you wouldn't if you didn't care about this intervention streak. So, there is a downside to this. My suggestion, although it's not easy to implement, would actually be for them to look at the data and give you extra points if you actually intervened in a timely manner in a critical environment.

[00:10:55]So, of course, I would be interested in knowing what you all think about this.

[00:10:57]Let me know in the comments. It is kind of divisive online. There are a lot of people that think it's a good idea.

[00:11:03]There's a lot of people that don't. And by the way, if you really want to help out this channel and help us get to 100,000 subscribers in 2026, we're getting really, really close. Please do consider subscribing and hitting that bell notification icon. Thank you.

[00:11:16]And finally, let's turn to upcoming improvements. These are actually the same as 14.3.2. So they didn't implement these things in 14.3.3. Number one is expand reasoning to all behaviors beyond destination handling. Nobody's been able to figure it out exactly what that means. But in my mind, what I think that is is kind of system 2 thinking. In other words, Grock level thinking because Grock is integrated more and more into the vehicle. You can talk to Grock now. You can actually say, "Hey, Grock now with this version of the software." So, I think what they're talking about here is expanding having Grock capable of reasoning about things beyond just simply saying like, "Hey, I need a five-star restaurant. Can you suggest something?" And then saying like, "Oh, yeah, navigate to that one or something." Right? That's destination handling. And I think what they're talking about is expanding Grock's capabilities within the Tesla vehicle to beyond that. Next up is one that's very, very obvious, which is add pothole avoidance. The more they can do that, the better. Even in Georgia, we do have potholes. And I know that's nothing like up in the north where you get the freezing and unfreezing and you get gigantic potholes. Avoiding potholes would be a really good thing for your Tesla's tires. And then finally is improved driver monitoring system sensitivity with better eye gaze tracking, eyewear handling, and higher accuracy in variable lighting conditions. And by the way, I think they've already started to do this. I've noticed that even when I'm wearing sunglasses, it seems to be able to see my eyes better. And if I take my eyes off the road and kind of looking over there or something like that, it's much faster about saying like, "Hey, pay attention." I used to be able to look pretty much anywhere I wanted to when I was wearing sunglasses, but that does not seem to be the case anymore. So, I think they're implementing this over time, but in a future version, obviously, they will have a step change that is significantly better than what is implemented now. Editor John here, after I recorded my video yesterday, you can see here that Elon Musk posted that 14.3 nags less, too. This was at nine o'clock last night. So, that indicates to me that they're already starting to work on the eye gaze tracking as well. I just thought I'd update you with that late breaking news. All righty, folks.

[00:13:11]So, that's what 14.3 brings to us. There are some nice new improvements. What do you think about the gamifying thing, like I said before, let me know in the comments. I think that's the most divisive part of this upgrade. I will be interested to try it and to see if it does have a negative effect on me actually, you know, waiting to intervene when maybe I should have done that sooner. If that's the case, that's not so great. But if it keeps me on full self-driving more, that would be great as well. But like I keep saying, every single time at the beginning of my drive, I always take a back route instead of the main road because it saves a bunch of time. So basically almost every trip I'm going to have to intervene, which means that my streaks are not going to be that long. And that's going to remain until Tesla allows me to do something like save a trip or show them the way that I want to drive or something like that. Until then, I will have to intervene a lot, which is unfortunate for my streaks, but I have a feeling a lot of people will be very happy that they get hundreds or thousands of miles without interventions. Anyway, while you're down there leaving comments, please do like the video so other people can find it.

[00:14:08]And like I said previously, if you want to consider subscribing and helping the channel out, that's fantastic. Thank you so much, and I will see you in the [music] next video. Bye-bye.

[00:14:23]>> [music]