This Is What Happens When You Disengage Tesla FSD

Añadido: 2026-05-13

[00:00:00]So, Tesla's full-self driving data logging and collection system is one of the most unique and complex orchestration services, in my opinion, in the vehicle autonomous vehicle marketplace. Let's put it that way. So, the first thing is that shadow mode. Let me cover that with you with Let me cover that first right off the bat, real quick, because that is one of the most interesting processes that I think any company has implemented, and it's so smart. So, what shadow mode is is if you're driving your car, let's say if is not engaged, you don't even pay for full-self driving, you don't have full-self driving, you're still training the model. They're still comparing what the car is doing, what you're doing, rather, if you're driving manually, to what their AI would do. So, if you have a disagreement between what you're doing as the driver, manually driving the vehicle, and what the model would have done, that is technically a disengagement. That is how they're able to get so much data in a short period of time. Gosh, that's a great-looking Cybertruck. I really like that in in white. But, that is how much That's how they're able to get so much data in a short period of time, because it's not just if you're using FSD, this shadow mode is sort of like the silent background, which is fascinating to me.

[00:01:19]So, again, when the AI predicts actions and it disagrees with the human, that is flagged. That is a technical disengagement, and it sends that data back to Tesla, process it. We'll get to that in a second. But, I felt as if it was important for me to lead with the shadow mode, because it's not just when you're using FSD, but it's also when you're driving manually.

[00:01:42]Now, the second thing is active logging.

[00:01:44]When FSD is engaged or events trigger.

[00:01:47]So, any disengagement, think about when we tap the brake, when we use the scroll wheel, when we um take over the steering wheel, for For we get this new pop-up now. I'll just show you. When I disengage, we get this new pop-up. And what's fascinating to me here is that I just disengaged.

[00:02:04]This pop-up now pops up as as a new feature, very new feature. Let me re-engage full self-driving here. And I can click on why I disengaged. For now, I'll I'll click on preference. What I just did is I'm helping label the clip, which is so fascinating. Now, we're even more integrated as a customer into training the model, making it better.

[00:02:23]We're more integrated into the system.

[00:02:25]And what happened there is that Oh, wow, that's a That's a tough break here. But I think it did a good job here. As you can see right in front, this dashed line, I think prevents or or is is supposed to prevent people from stopping. So, anyways, what I just did is I disengaged, labeled it, and now that data is going back to Tesla to to be processed. And what's what's fascinating here is that there is an element of human labeling and automated labeling at the Tesla headquarters, or let's call it Tesla AI, their supercomputers. It's not fully automated. There are teams of what we call labelers that look at clips and identify, okay, this was a disengagement because the car in front was pulling too close. This is a disengagement because there was a debris in the road. This is a disengagement because there was road kill. Etcetera, etcetera, etcetera. So, it's it's There's Yes, it's it's a lot of it is automated when you disengage.

[00:03:24]But there's also this human element of See that truck right there? That truck is kind of taking space where it shouldn't. There It's rear end is basically sticking out in this dashed line. At least that's what I think it is. Let me know down in the comments if that's what it is. But that's kind of the labeling perspective.

[00:03:42]We're helping to label with this new box. And you also have label labelers at Tesla, which is quite fascinating to me.

[00:03:50]How the data physically gets sent to Tesla. I did not know this. I did not know this is how it actually works. I'm very curious to know if this is how your understanding of it is as well, or if you knew this, or if I'm just getting I'm I'm I'm getting caught up to speed.

[00:04:06]When I was talking to Grok about this, I was quite fascinated. So, here's what happens.

[00:04:11]There's storage on board of your Tesla.

[00:04:13]The video is buffered locally in your solid state driver, your SSD.

[00:04:18]And then what happens is when you connect to your Wi-Fi network or a Wi-Fi network, that's when data gets sent to Tesla.

[00:04:30]The fallback to that is cellular. This is apparently, allegedly, used sparingly for very critical or extremely high priority clips, for example, crashes. Or if Wi-Fi isn't available soon. Tesla's modem handle this handles this efficiently. I had none no idea either of those actually is how it works. I just thought it gets sent to Tesla remotely all of the time at the time of disengagement. But it actually seems to me, correct me if I'm wrong in the comments, it seems to me like they weigh the the clip. So, if it's a critical disengagement, as far as maybe a crash, an incident where it's a very high priority in terms of safety, that gets sent and shipped almost instantly over cellular data.

[00:05:19]If it's lower priority, if it's road kill, if it's just your average disengagement, if it's a preference, if it's a navigation-related disengagement, those all get locally stored. As far as like you can think of it as cache. It's essentially just cached on your local system in your car.

[00:05:38]And then when you get into your um your home, if you connect to a Wi-Fi network at home, which most of us do, that's how most of us receive updates, Your car then ships those clips to Tesla, which is very interesting. I had no [clears throat] idea that's how it works.

[00:05:55]And then Grok made sure to tell me a little bit about the compression and selection.

[00:06:00]It said here that not every frame or second is is sent. Only curated clips via shadow triggers or events.

[00:06:08]And the data of course is anonymized. So all that good stuff that we already know. Um wow, this this Dodge Dart is really darting in the traffic there. Um but I found that to be super interesting. Now the processing pipeline at Tesla from clip to the neural net improvement. What actually happens when it gets there? This is also quite fascinating.

[00:06:30]When the clip arrives at Tesla's data center, shadow mode and disengagement data are triaged by importance. And this is a human feedback loop. This is kind of where we as the end users now contribute to the labeling process in addition to the human labelers that they have.

[00:06:48]Again, Grok mentioned this as well.

[00:06:50]There's a hybrid plus a hybrid human plus AI process. The early days were heavily manual. Now it's heavily automated with unsupervised and self-supervised techniques. The neural nets learn. It's It's basically a pattern recognition directly from the raw video clips.

[00:07:10]Tesla's custom supercomputer is purpose-built for video data at petabyte scale. Petabyte. That's just an insane amount of data.

[00:07:21]So the iteration is train models, validation on held out fleet data, then safety checks, then over-the-air deployments. This kind of closes the loop and why FSD improves almost weekly if not monthly. Let me make sure my camera's still up there. Good. It's still good. And I appreciate I'll I'll I'll pause here. I'll I appreciate everybody's feedback on the previous video. I'm kind of iterating and trying to make these this dual setup better.

[00:07:50]So, just bear with me. I think in in this video it should be better. It should be more stable. Thank you for those of you that recommended turning off the stabilization on the interior camera. I really appreciate that. That's exactly what I did. So, you should see a little less movement, but let me know in the comments. How do you guys like it? Do you want me to continue this dual setup?

[00:08:11]If not, I'm happy to revert back to just the windshield. But for me personally, I do enjoy watching the 360 the ultra wide view on top. But again, this is all part of the feedback loop for me. Just like feedback loop for Tesla to get FSD better, there's certainly a feedback loop for my channel.

[00:08:30]Um and so, let's let's pivot back to Tesla here for a second and how FSD disengages.

[00:08:36]There's a privacy component to this as well.

[00:08:40]You must agree to data sharing when enabling FSD or certain features.

[00:08:45]Um it Grok went out of its way to tell me that there are edge cases. So, parked car, limited processing, no Wi-Fi for days, older hardware, hardware 3, still supported but processes shadow mode less aggressively in some cases. Uh intentional differences or data handling compiles with local laws. That's also interesting. I had no idea that there's I mean, it makes sense now that I think about it. I think it varies state to state, jurisdiction to jurisdiction even. So, that's interesting.

[00:09:15]Um the implications positively it accelerates safety of course. There's a rapid pipeline, very fast and aggressive iteration mechanism. So, that's all really good. As far as the critiques, it says that some orders find disengagement prompts intrusive, which I think what they mean now specifically as of the most recent update is the disengagement box that we've really started to complain significantly about, but have since uh provided feedback to Tesla and they've tried to improve it to a degree.

[00:09:45]It still doesn't go away, as we all know. I'm not going to harp on that.

[00:09:49]I've harped on it enough, but I think that's kind of where Grok is coming from.

[00:09:53]So, in a nutshell, in a very fast boiled down version, there's shadow mode, there's human labelers, there's AI fully automated labeling systems in place. And what I found most interesting, not that it's a crazy data point or anything, not that it that it really even impacts FSD or the experience of the day-to-day or the safety, but it's just the fact that I always thought that when you disengage full self-driving, that clip gets immediately sent to Tesla, but that's not the case. It actually stores it locally in your car and cached, and then when you connect to Wi-Fi later that night or maybe it does uh uploads on a weekly basis, I don't know. Maybe it only uploads as needed based on the storage. I have no clue how that works. But I always thought it was instant and ships delivers and then instantly gets uh forked over to Tesla for processing. Um again, not like that's a huge data point in any which way, doesn't impact the day-to-day. I just thought that was interesting. So, that's kind of what I wanted to touch on today. What actually happens when you disengage full self-driving. From the moment you disengage to shadow mode to processing to labelers, it's all really fascinating stuff. It's very interesting. Oh, yeah, and then the last thing that I remember reading, it's not in my notes here, but I do recall reading somewhere um in my Grok report that believe it or not, they don't differentiate between supervised or unsupervised, as far as I can tell.

[00:11:25]Obviously, there has to be an element of tagging in the system somewhere for them to know that this came from an unsupervised model versus a supervised model, but they might treat it very similar or more similarly than you might think, which is fascinating. That's just food for thought.

[00:11:42]Anyways, I'm not going to make this video as long as my previous one. If you've made it this far, first of all, thank you. Thank you so much for listening to me ramble about disengagements and nerding out about FSD. I really appreciate the support. I really do. If you haven't yet subscribed to the channel, this is another friendly reminder to hit that subscribe button down below. It helps me get to the next goal, which is 50,000 subscribers on my channel. I really appreciate all the support, guys. If you have any questions, comments, concerns, if you want to reach out to me via email, it's [email protected].

[00:12:16]You could find the email in the description below. We're pulling into Whole Foods now, so with that, guys, I'm going to let you go. Thanks so much.

[00:12:22]I'll catch you guys in the next one.

[00:12:24]Bye-bye.