Tesla HW3: You Can FEEL When AI Runs Out of Compute

本站添加: 2026-05-26

[00:00:00]A few days ago, I resubscribed to full self-driving on our older hardware 3 Tesla Model 3. While I've been driving full self-driving 14.3.2 on our Model Y and Cybert truck, a 5-hour trip back from the beach got me thinking. Let's give hardware 3 a try again in our Model 3. Well, it only took about 20 minutes for us to realize something very shocking. The older system doesn't just feel slightly worse, it feels like a completely different species of intelligence from our AI4 cars. What kept jumping out at me over and over again wasn't the intersections or the lane choices so much as a really ugly oscilly behavior. At highway speeds, the car weaves and corrects back and forth inside the lane like it's constantly trying to catch up with reality itself.

[00:00:41]And worse, the faster you go, the more pronounced this oscillation gets. At 30 mph, not such a big deal, but at 70, the behavior becomes almost seasickness inducing. And this put me in mind of two things from aerospace. Number one, pilot induced oscillation in aircraft. And number two, the Apollo lunar module ascent from the moon where delayed control corrections created visible oscillations during takeoff. In my opinion, what we're seeing here is the physical limits of embodied AI compute in real time. Because unlike chat GPT, when an AI system controlling a physical object reacts too slowly, the result is not a slightly slower answer, but late overcorrection in the car. And that's not good at all. And as it turns out, Elon Musk himself has now admitted publicly that hardware 3 will never achieve unsupervised full self-driving.

[00:01:27]And after driving it for a bit, I can absolutely see why. So today, let's have a chat about what I'm actually seeing in the real world driving hardware 3 and four systems backtoback and why if you own a hardware 3 vehicle, you're never going to get true full self-driving.

[00:01:41]Let's take a look. Before we start, a quick shout out to my channel sponsor Joah. They make amazing accessories for your Tesla and other EVs and also have great warranties and customer service, too. In fact, I use their accessories daily. Be sure to check the link in the description to get 5% off a fan cooled phone charger, a portable tire inflator, a foldout lap table, and a lot of other items. Oh, and they make perfect gifts for you and your Eevee loving friends, too. So, be sure to check out Joah. And now, let's get back to it.

[00:02:10]Hey y'all, it's Dr. Knowit All. First of all, happy Memorial Day to those who served in the military forces in the United States and everywhere in the globe. I appreciate your service. Thank you so much. Second of all, a big shout out to my friend Mark who drives an older hardware 3 car and said he recently resubscribed to full self-driving after not driving it for a while and he thought the experience was way, way better. Now, I think he originally was driving like version 11 of the software or something like that.

[00:02:33]So, obviously version 12.6.4 that he's driving currently is significantly better. But anyway, that combined with us driving back from the beach in our Model 3. I was like, you know what? We got a 5-hour trip. Let's just resubscribe. We'll spend a hundred bucks. We'll see what it's like. Those two things combined were what got me to pony up the hundred bucks. Am I going to continue the subscription? That actually depends a lot on my son. If you haven't seen that video yet about him having our Model 3 at college, you can definitely check it out up here. I'll leave it at the end of this video. But anyway, we're still deciding about that. It looks like we are going to side with keeping the Model 3 cuz it turns out there's some local charging really close to where he's staying. We actually stopped in Statesboro on the way back from St. Simon's Island. So, it was actually super convenient and a really good scouting mission.

[00:03:15]So, before we move on to the main issue, a few other issues I saw with 12.6.4 that did not surprise me nearly as much cuz I kind of expected that stuff.

[00:03:23]Number one, it's got very, very late lane planning as opposed to 14 that I've gotten used to that really plans ahead now. It thinks about what lane it wants to be in a significant distance ahead and places itself there. With 12.6.4, for it's still very very late in lane planning. It also completely misses railroad tracks and it also breaks a little bit late as well. It feels a little bit uncomfortable because it comes up on cars quickly and then breaks fast. And then to some extent there's some hesitation in the car as well, but it's not as pronounced. But overall, AI4 is significantly better in all of this stuff. It thinks ahead more. It's got a more humanlike reaction to situations.

[00:03:57]It sees train tracks, things like that.

[00:03:59]It's actually a little bit jerky once in a while because the cycle time is actually so fast that it actually makes multiple decisions a second. And sometimes you can feel that, which I find fascinating because it's the exact opposite of hardware 3. But overall, FSD 12.6.4 on hardware 3 is not terrible in city streets and at relatively low speeds below about 45 or 50 m an hour.

[00:04:18]But when we get to highway speeds, when we're going 65, 70, 75 mph, the limits of hardware 3 are really on display. You start to see some significant oscillation. I'm going to show you a quick video clip here. I don't know if you'll be able to tell particularly well. It's always hard to tell from a video, but just imagine yourself in the car and the steering wheel turning like that and you rocking back and forth. It really does start to make you feel like you're going to become seasick.

[00:04:44]Okay, this is not good. Look at the steering wheel.

[00:04:47]It's like it's a little hard to tell, but I don't know if you can see, but it is absolutely not cool.

[00:04:54]It was to the extent that we thought the car was going to tip itself over. Okay.

[00:04:59]Yeah. So, this is full self-driving 12.6. 6.4 and the old hardware 3. Not good at all.

[00:05:06]So, with that, even though you can't experience it yourself because you're not getting that rocking back and forth motion, I'm sure you can imagine just how bad that experience would be being in the car. If you actually have full self-driving with hardware 3, let me know in the comments what you think about all of this. Do you experience the same thing? I actually did some research and I found out from Reddit and X and a whole bunch of other places that a lot of people are describing the same problem. They're describing it with terms such as pingponging like it's going back and forth, swaying, quivering, motion sickness, hunting for the center line, things like that. It tends to go back and forth and back and forth inside the lane. And clearly from the number of folks out there that are complaining about this, we are not the only ones. I thought maybe our car was messed up. I actually recalibrated the cameras on the car just to see if that would solve the problem, but it didn't.

[00:05:52]So that's the observed behavior. Why am I saying that? This is indicative of the fact that hardware 3 will never be able to solve full self-driving and a car with hardware 3 in it will never be able to go unsupervised. It will never be able to drive you by itself. The answer to this goes down to control loop theory. And this doesn't even involve neural networks or things like that.

[00:06:10]It's something that's existed in computer science and robotics for decades since the earliest days of robotics. Actually, if you talk to Dr. Scott Walter, I'm sure he would talk your ear off about control loop theory.

[00:06:20]Anyway, control loop theory is perceive, predict, decide, act, re-evaluate. So basically, you perceive the world with whatever system you've got. You then predict what the next action is. You decide on that action. You act and then you re-evaluate the situation. And of course, this is not just for robotics.

[00:06:36]It's for human beings as well. We do the same thing all the time. If we're driving or flying, for example, you perceive, you predict, you decide, you act, and then you re-evaluate on a consistent looping basis all of the time while you're doing something like driving or flying. And potentially, there are two major problems with control loop. Number one is overcorrections. In other words, you make too large of a decision. So, you're driving and you want to turn just a little bit to the left or something like that and you overturn to the left.

[00:07:02]That's a significant problem. But for our situation, the bigger problem in my mind, the problem that the Model 3 with hardware 3 is having is that the loop is too slow. It's not running fast enough.

[00:07:13]I would take a guess that it's running at about two hertz or it's making decisions somewhere around two times per second. And it needs to be making decisions more like 25 or 30 times a second, not two times a second. So, it's about an order of magnitude too slow, at least at highway speeds. So, the problem with this delay is if you get delayed corrections, you tend to overshoot. So, right, you're on the right hand side of the lane. You turn the steering wheel.

[00:07:36]You start moving to the left. You re-evaluate, but you re-evaluate too late. And by the time you re-evaluate, you're left of the center line. So, then you turn the steering wheel back to the right again. You counterorrect. And then you re-evaluate, but by the time you've re-evaluated, you're too far to the right of the center line. You turn the steering wheel back right. So, this creates this oscillation, this weaving motion. And if it gets really bad, you'll enter a positive feedback loop where you will correct to the left.

[00:07:59]You'll overcorrect back to the right and it will be bigger than your correction to the left. You'll overorrect to the left, it'll be bigger, overcorrect to the right, over correct to the left, and before you know it, you're off the road.

[00:08:08]You're weaving around off the road or something. So, why is this a bigger deal at higher speeds? Well, it just comes down to the fact that you can make corrections more slowly when you're going slowly. If you're going 34 mph, that's about 44 ft per second or somewhere around 12 m or so per second, somewhere along those lines. If you're going that speed, you're not covering that much ground. And a slower control loop at, let's say, two hertz or two times a second is able to make corrections that are reasonable. It doesn't travel that far every half a second. It's only going a little bit in and out. So, you don't really notice it.

[00:08:38]It's able to correct and be reasonable about it. On the other hand, at 70 mph, you're going about 103 ft per second or somewhere around 33 m/s, which means at a two herz loop frequency, you're going over 50 ft before it makes a correction.

[00:08:51]So, if you're traveling down the highway, 50 ft is a long way. That's multiple lengths of the car that you're going before the car re-evaluates where it is and makes a decision about where it should go next. And at that speed, two herz is just flat out too slow to keep the car centered. The correction loop is always going to be behind and you're always going to get oscillations.

[00:09:08]Now, this of course is my speculation.

[00:09:10]We don't have a lot of actual evidence from Tesla. But given the fact that hardware 3 is almost 10-year-old technology right now, and Tesla didn't even quite know what they were doing at the time they built hardware 3, it's not unreasonable to believe that this hardware is at the very, very limits of its capabilities. And it's just not able to operate fast enough to be able to keep the car feeling natural, centered in the lane, and not overcorrecting and oscillating as it drives, at least at highway speeds.

[00:09:36]Now, fascinatingly enough, I've seen this in two other places that really, really matter. Both of those are aerospace industry situations. The first one is pilot induced oscillation which actually can be very very bad. There's a thing called a Dutch roll where there's a thing called a Dutch roll in specific that's extremely dangerous especially with a Vtailed airplane which is where the airplane banks and yaws at the same time and the pilot overcorrects banks and yaws banks and yaws banks and yaws banks and yaws right and eventually the plane will actually flip over or it will get outside of its tolerances and a wing will break off or something like that.

[00:10:08]It's very very dangerous. But the root of the problem is that the pilot, the human pilot is correcting too slowly.

[00:10:14]And when they correct, they're overcorrecting. And then they're overcorrecting and overcorrecting. And it's the same sort of feeling I feel with the hardware 3 Model 3 in the car.

[00:10:23]And then the second comparison is actually computerized as well, 1960s technology instead of 2020s technology, but it's the Apollo Lunar Module Ascent.

[00:10:31]As you watch the Apollo Lunar module ascend from the moon, and I've got some original 16 mm footage here that I'll overlay. You can actually see through the window that the moon is kind of oscillating back and forth. And that's because while the Apollo guidance computer for the time was miraculously amazingly fast, it was deeply compute constraint and could only update its status a couple of times a second. It was just a couple of hertz. And you can see that delayed stabilization visibly as it's launching off the moon as it's rocking back and forth. Now, that oscillation was damped out so that it didn't get worse and worse. you didn't enter a positive feedback loop which would have caused the LEM ascent stage to get out of control and oscillate more and more wildly. Fortunately, that did not happen. But you can definitely see that it was oscillating because the compute technology, the control loop was too slow to keep it smooth on ascent.

[00:11:18]So, what I'm noticing here is that hardware 3 is apparently reaching a similar kind of boundary condition for AI that Apollo missions were experiencing back in the '60s.

[00:11:28]And that takes us back to about a month ago, April 22nd of 2026, where Elon during the Q1 earnings call admitted fully that hardware 3 cars will never get unsupervised full self-driving.

[00:11:39]Reading what Elon said, he said, "I wish it were otherwise, but hardware 3 simply does not have the capability to achieve unsupervised full self-driving. We did think at one point it would have that, but relative to hardware 4, it has only 1/8 of the memory bandwidth of hardware 4. And memory bandwidth is one of the key elements needed for unsupervised full self-driving. So what does that lower memory bandwidth translate into?

[00:12:01]It translates into data going much much more slowly between the compute units and the memory units. You can imagine sucking air through a tiny straw rather than a gigantic hose, right? You can't take in as much air, which means that the compute units are going to run slower and it's going to update more slowly, which exactly matches our experience driving back from St. Simon's Island and yesterday driving to Atlanta and back.

[00:12:22]So in the end, while this is highly disappointing, it's also not very surprising. Embodied AI has very very strict latency requirements that are just non-negotiable. With a chatbot, if you get a slower response, it's like that's okay. But with robotics, especially if you're driving a car at high speeds, low latency and efficient control loops just have to happen. And it turns out Tesla is unable to push this almost 10-year-old technology forward any more than it is already. And even with full self-driving 12.6.4, we're seeing significant limitations that lead to oscilly behavior that is just unacceptable at highway speeds. All righty, folks. That's what I've got for you today. Let me know in the comments what you think about all of this. I have a feeling for folks like us that own hardware 3 cars that, you know, you might not be very happy about this. And I have to say, you have a right to be upset about this because of course Elon and Tesla did promise that we would have access to full self-driving with hardware 3 cars. And eventually, it sounds like at least if you purchased full self-driving, they will try to make it right for you. But I have a feeling they will delay this and delay this and delay this and hope that you will buy a hardware 4 car instead, which will get them off the hook on having to upgrade your vehicle. Anyway, let me know in the comments what you think about all of this. If you have a hardware 3 car, are you considering getting an AI4 car instead in order to get real full self-driving? In my mind, what Tesla should do is take every hardware 3 car where the person purchased full self-driving with the vehicle and they should give you a free transfer of full self-driving to an AI4-based car if you sell your Tesla at the same time. That's what I would do if I were Tesla to make things right and make it go away more quickly than having to recall the cars and put in new boards, put in new cameras, all that stuff. That's very expensive. Much easier for them just to give you free full self-driving upgrades. Anyway, let me know what you think while you're down there. If you don't mind liking the video, it really helps other people to find it. And if you want to help out this channel the most you possibly can in 2026, help us get to 100,000 subscribers, please consider subscribing and hitting that bell notification icon. And I will see you in the next video. Bye-bye.