Tesla FSD 13.2.1 Update Comes To Cybertruck!

[@Jaw0lf]

Amazing how different the two vehicles are for FSD.

[@freddydad1]

the only thing that matters is that there is continuous improvement.

[@CyberTahoe]

Just got my update overnight. Stoked!

[@diamond_h0us]

Same. Installed it when I got to work on my Y and can’t wait to try on the way back home. I take a rural route that’s a pretty good test for it.

[@mariusmeyer14]

They should call it Version 11 for Cybertruck

[@w0nd3rlu573r]

Hmmm, what's with the tinted right and left windows, but not the right small one? Is it just light polarization. or I am missing smth? Is it the glass that spontaneously broke?

[@JohnEFly]

<a href="#" class="seekto" data-time="415">6:55</a> I think cybertruck hit the brakes because of the car to your right, not the car straight across. It was still moving when the truck braked. I think it thought it may be running through the stop sign and was being very cautious. It stopped just after, and just beyond the line, but the timing of it makes me think it was worried about a possible collision, even though it was your turn to move.

[@D8V1Dx]

Prob still needs more training since all 4 wheels on the CT can turn.

[@RedPowerStation]

I wonder if they are testing strategies... theoretically you could attempt to map the Model Y training data to the Cybertruck with a transformation layer either at training of inference. However the physics of both vehicles is so different. This would save on having to gather specific Cybertruck FSD data... however I expect this could be a short term solution until Tesla can get enough Cybertruck specific FSD data for a Cybertruck only model. This could explain the performance differences between CT and Model Y.

[@wademt]

@@RedPowerStation my guess is that they are curating Cybertruck specific data and not transforming video from other vehicle types. And I’m also guessing that they just need more curated video content to train their model.

[@RobertLoPinto]

"Cybertruck is lagging" HW3 owners be like 🤔

[@canngo123]

In Oregon, it’s actually a traffic offense to wait in the intersection to make a left. It’s a stupid law but I wonder if that’s why it does that.

[@3nd04]

I would assume in that case the driver is forced to go straight until they find another place to turn or make a u turn, that does sound silly, but it makes sense for single lane roads

[@gocybertruck8189]

I'm excited to see FSD v13 on the non-foundation Cybertruck. It's great that you're pointing out some annoyances. I've tried FSD v12 a few times and was impressed. Surely, it will improve just like it did with the Model Y.

[@martindbp]

If I had to guess, they're probably fine-tuning the model for Cybertruck with much less data and it destroys some of the ability. What you want to do is to parameterize the model for different cars, i.e. the car, or the attributes of the car (size, weight etc) is an input into the model, along side the video stream and route/map data. I wouldn't worry too much, they'll figure out a way to have a single "driver" for all vehicles.

[@marcusschou1843]

I Think so too must be due to different camera angels on the cybertruck.

[@didierpuzenat7280]

The issue is how much new examples are needed to fine tune, I would say quite a lot since the characteristics of the car impacts many quite different scenarios. It will be a good indicator of the future difficulty for Tesla to launch the Cybercab and even to sell trained models to other brands.

[@bear4278]

That makes sense to me with getting too close to stuff etc. (obviously it is bigger than the Y etc.), but not moving from a stop sign and running a red type behaviour is harder for me to explain (I would think the existing data should be usable since the size of the vehicle shouldn’t really matter) 🤷‍♂️

[@didierpuzenat7280]

@@bear4278 If they do not have enough examples (even to fine tune), a solution is to use a model with less parameters and so it will be less efficient on every thing. But of course it is just an hypothesis.

[@martindbp]

​@@didierpuzenat7280 Different camera placement may be the biggest issue unless you have some preprocessing step to normalize them. It may be difficult for the model to generalize over different placements (i.e. be able to handle a new one out of the box), since there are essentially only a few "examples". It's easier to generate data in simulation for lots of different camera placements, that could help, but not ideal.

[@Wankidy]

There is no data to train CT

[@khaleefax9553]

*How is the ride quality on the Cybertruck if you have driven a ICE truck before and can compare?* *I drive a Maybach and a Lexus IS300, picked Tesla Model Y as our third car - I think the Tesla's ride quality needs a lot of getting used to. I mean, even in small potholes, the suspension seems to hit you pretty sharply.*

[@BongoWongoOG]

@<a href="#" class="seekto" data-time="70">1:10</a> ?

[@chrisanderson687]

Speculation: there is just a massively smaller amount of training data for the Cybertruck, since the X/Y were out for so much longer. This should improve massively over the next year though. Luckily CT is insanely fun to drive so I don't mind keeping it in "manual" for awhile :)

[@SkullEZClap]

Ayyy finally

[@DouglasJMark]

Loved your navigation oopsie! 😂

[@Scott-sm9nm]

<a href="#" class="seekto" data-time="1090">18:10</a> I can't you did not disengage when that big black SUV was moving into your lane ... even tho they didn't have a turn signal on it felt like they were making a last minute decision to turn.

[@EthanNelson-1]

I'm 47yrs old. $73,000 biweekly and I'm retired, this video have inspired me greatly in many ways!!!!❤️

[@James1xr]

I'm feeling really motivated. Could you share some details about the bi-weekly topic you brought up?

[@EthanNelson-1]

All thanks to Maria Luisa Clare

[@EthanNelson-1]

After I raised up to 325k trading with her I bought a new House and a car here in the states 🇺🇸🇺🇸 also paid for my daughter's surgery (Joey). Glory to God.shalom.

[@Michelle_59]

Absolutely! I've heard stories of people who started with little to no knowledge but made it out victoriously thanks to Maria Luisa Clare.

[@Jasmine12-i3f]

Wow 😱 I know her too Miss Maria Luisa Clare is a remarkable individual whom has brought immense positivity and inspiration into my life.

[@byGDur]

Cybertruck beeing more timid is probably just a phase. If the other models can do it better than so will Cybertruck eventually :)

[@happycamperinc.]

The whole "Cybertruck FSD" Program Sucks from the Start of cybertruck delivery event.

[@didierpuzenat7280]

Any new car will need some new training, it is not relative to the Cybertruck. It will probably be worse for the Cybercab since it cannot be released to the public before being operational so all the training will have to be done by Tesla staff on private roads (and of course in simulators).

[@GreylanderTV]

<a href="#" class="seekto" data-time="75">1:15</a> She thought it was weird that you stopped. CT's path plan was a normal turn, which would only be a problem if she stepped into the driving lane, same as if this happened while you were going straight now a road. FSD correctly assessed that she very likely did not intend to do so, and was taking the turn slowly enough it could stop if she did. Note that she had already turned her head and was aware you were there making the turn, FSD could see the body language as well, thus reinforcing that she was unlikely to step out (as opposed to, say, if she had her nose buried in a smartphone oblivious to all around her). Put it this way, if you'd been making the turn yourself, would you have stopped? I don't think so. Just take it slow and careful. 6 mph is pretty slow...

[@joshcarter-com]

Thank you for showing us the limits of CT FSD, both the good and the bad. Mine isn’t offering the v13 update yet. My biggest issue is awful turns at large intersections-it appears to not know where the lanes are so it hunts mid-turn. Fortunately I haven’t seen the late-braking problems you mention.

[@briansmith8320]

Do you think the rain has any effect on it being less smooth than the Y?

[@plixie1]

Im on hw3. The setting sun was so intense that the car beeped and cut on its hazard lights and told me to take over. I will be canceling until hw3 upgrade. The shadows causing hard phantom braking and difficulties handling morning communtes ( school busses and school zones).

[@quadsis5569]

They obviously need more Cybertruck data

[@appl314]

Question. If people use fsd for cybertruck, will they get enough good data? I thought they'd use the same data, but use an algorithm for different hardware interfaces....guess not

[@kenion2166]

@@appl314 FSD runs in shadow mode if its not active and still collect data in the background

[@appl314]

@kenion2166 Yes, but if everyone is using fsd, no human shadow mode data.

[@jmendiola222]

Yes, but it's not promising that they can't abstract the form/weight factor and handling of the vehicle from the actual behavior. If they have to "duplicate" all the data from every different model, it wont scale.

[@joshcarter-com]

@@jmendiola222 that may be part of the trade-off with a machine learning stack; you can’t get in there and start adjusting the model. It’s much more of a “black box” than their prior hand-coded FSD stack.

[@rquindoy]

I have a hw3 ModelY on 12.5.4 2 and it drives so much better than my CT on 12.5.5.3. I totally understand your frustration and thanks for being honest in your review.

[@augustas1]

I still don't see the point of the cybertruck

[@Jens.M]

KZbin now also translates the video title to my language? Ohh man I don't want that feature! How can you even see trough these dark window tints? Crazy! I didn't noticed that until now

[@rmkep]

In the spirit of uninformed opinions ... it has to be steer by wire that causes many of the problems you clearly demonstrate. The CT is very different and all that machine learning video off the other Tesla models isn't perfectly applicable to the CT.

[@retro1120]

<a href="#" class="seekto" data-time="78">1:18</a> intervention was soooo unnecessary... Would you stop for that lady like that, 60 meters away from her if you were just driving normally?😭

[@JohnFabrega]

Installing now!

[@SkepticalCaveman]

It's a different vehicle, with less data. The Cybertruck will catch up eventually.

[@DanFrederiksen]

It might be the same 13.2 model that they finetune for cybertruck afterwards with tesla drivers driving a certain way, taking a page from LLMs. So based on the same understanding as 3 and Y but conditioned behavior for CT afterwards.

[@noahhogue9341]

I demo drove a cybertruck a few days ago. It was my first time even driving a Tesla. I tried the fsd and it was wild it made take over one time but it might have been because I was recording.

[@daniele_go]

Why FSD would behave differently on CT than on MY? and what about M3? I am SO SURPRISED that when FSD is installed on a new model it does different things ?!?!?

[@fox_3591]

Different AI model behaviors on different vehicles. Theirs alot of different things the AI model has to understand based on the vehicle Look at how they always release new versions first on the Y, because it has the most miles as the most prominent vehicle model on the road.

[@Someone_Cool.x3]

I'm guessing they have some training to do with CT moving through the environment hence, inability to park, unpark, etc.

[@w0nd3rlu573r]

Also - are there any heating elements in the front bumper bellow the headlights? It seems super illogical to put a shelf, that collects dirt and snow, in front of the only lights, that you can use for driving in the night.

[@upsguppy520]

top driving on road to test this

[@voelkela]

Would you know what the technical reason for that different behavior is? - Does CT need specific training? Can the training experience gathered with earlier existing models not be easily re-used? - Is it the different sensors' positions?? Could that mean that the training is NOT (yet) independent of the sensors' positions?

[@TannerBugatti]

These have to be different models per vehicle right? I can't imagine it's the same model that is used on the 3/Y.

[@wademt]

@@TannerBugatti I’m guessing it’s the same model structurally, but unique curated video data per vehicle.

[@voelkela]

Correct title photo!

[@byGDur]

<a href="#" class="seekto" data-time="1090">18:10</a> wat :D

[@TheMamonti1]

what do they do when the light turns yellow?? You know.... when your going that in-between speed where if you stop you'll squeal the brakes if you go the light is red before you pass the intersection.

[@happycamperinc.]

<a href="#" class="seekto" data-time="884">14:44</a>

[@gerardgalasso]

It’s definitely different brother. Noticeably more scared

[@Haffy1952]

I would think that the lack of data the Cyber Truck has at this time is a lot less data the 3 and Y have is maybe the reason that it isn't as smooth. Just my thoughts. Many more 3 and Y's are on the road compared to the CT.

[@LloydSchober]

Never seen FSD in action but 6 seconds into the drive and you have to intervene so you don't run over a pedestrian ?

[@l0v3nul]

Probably the model has to be trained on cybertrack and not model 3, Y data...

[@gerardgalasso]

I noticed FSD with the truck on the highway just doesn’t maintain consistent follow distance, doesn’t accelerate and I always have to

[@rogerheuckeroth7456]

Same

[@MrFrmartin]

still waiting for v13 for my Model Y

[@jamajcaman]

I feel like Cybertuck just need mora data for the model.

[@AMPed.]

<a href="#" class="seekto" data-time="78">1:18</a>.. lady starts to move toward driver side as FSD making left. CT would have stopped. I wish you let it play out. as soon as left started lady was still kinda not going left. At that point not clear if going to right or left of car. Maybe FSd could have waited 2 seconds before going. hmm

[@Neukids]

So, its being released to ALL CYBER TRUCKS or only those in beta testing, influencers etc?

[@andrewvercillo7584]

That disengagement was not needed!! You did not need to do that!!! You were uncomfortable. Cybertruck was confident.

[@Benedict-01]

You have to get used to that people let the Cybertruck go first when they see such an imposing appearance 😉

[@PaulSage]

Have you considered wrapping your truck in a color that matches your local mud/dirt? 😂

[@rcguymike]

Just leave the mud/dirt on it🤷‍♂️😁

[@PreposterousPixel]

I think most of the issues with Cybertruck's FSD come down to Tesla's FSD architecture. Tesla's photons in controls out means FSD needs to be re-trained for each of the vehicles (or at least most). Cybertruck's a very different shape and size, so most of the previously gathered training data from the S3XY lineup can't be used, where most of Tesla's data is. Less data means Tesla can't be as picky about what goes into the model. While they do make a lot of synthetic data, it's not a perfect <a href="#" class="seekto" data-time="61">1:1</a> to the real world, however close it may be. Previously Tesla was mapping camera data to 3D space, space that could have been shared between all the vehicles, but I'm assuming they switched due to processing limitations and maybe compounding error.

[@avidavid9237]

Brian@FutureAzA talked to a Tesla engineer about this, he indicated that it's because the Cyberturck has 4-wheel steer. It's tracking geometry is different from all the other cars.

[@PreposterousPixel]

@ Makes sense, though it wouldn’t be the only contributing factor. The truck has a significantly different turning radius than sedans and crossovers without all wheel steering, but moves differently on highways with it. Using the same model had almost no shot of working

[@antoinepageau8336]

Still waiting for holiday update and still on version 12.5.5.3. Being in a larger city than Ann Arbour I have many lane selection issues at complex highway interchanges. It'll be in the correct lane that ends soon and it'll decide to change lanes for 200 ft and then try to get back into the lane we were in that is completely full 🤦🏻‍♂ I hope V13 addresses this issue.

[@happycamperinc.]

Cybertruck is amateur in Driving himself. Model Y kicks ass

[@Spatial_Computer]

If Tesla is using a nueral net AI like they said they started on FSD 12, then the cybertruck is still in its infant phase of training. Model y FSD 13 has way more hours trained. I wish the FSD version came with a release note for hours trained. I assume they changed the neurons from FSD 12 to 13. A release note saying what was added or removed would be great or quantity changed. Might be too much for customers. But, much easier to compare actual driving experience to see, oh FSD 13 CT has 100 hours trained and 50 neurons, model y 10,000 hours trained and 60 neurons. Just example, not evaluated numbers.

[@wademt]

I agree. The amount of video available for 3/Y, and the amount of curation that the FSD team has done has brought v13 to a really good functioning product. The FSD team just doesn’t have the same amount of video content and hasn’t had the time to curate the best content to really train the model. I’m speculating that the model they are using is the same structurally. Same number of layers, same transform data types in the layers. (I use a neural net model to forecast next hour energy). They just need better video data for Cybertruck.

[@yuniorsanchez8578]

It is not that close, based on the Camara there is like a good 6-10 inch away

[@no_more_free_nicks]

square box

[@aethernets9442]

achtually, its more of a rectangle 🤓

[@dev-11164]

#1 content creator

[@rwhirsch]

attach some bull bars and those wheel spikes from Ben Hur to your CT and let it go crazy.

[@byGDur]

LFG

[@jmendiola222]

It's gonna ruin fsd tracker stats, as we'll have mixed data from 13.2.1 on CT and the rest. :/

[@eruveo]

im guessing tesla is intentionally giving it more hesitation since they know it sucks atm and will lower it jn the coming months

[@chadwickwood9843]

Pretty bad

[@SkullEZClap]

Dang 😔 CyberTruck FSD update disappointing. I would’ve figured FSD on the truck would be better since there’s more cameras and they’re higher up than other Teslas…… I guess I’ll buy Model 3 or Y

[@Doriorio]

Huh .. i think itll updated tho

[@timepstein274]

It will get there. The biggest problem at the moment with all the larger models (S/X/CT) is that the training model is based on model 3/Y training and they have to continuously compensate for the larger diameter that they weren't trained on. Once the new super clusters come online, they will be able to train each model in parallel and then they should all become as smooth as the 3/Y are on Ver 13

[@SkullEZClap]

@@Doriorio CyberTruck, S and X have always been behind the 3 and Y when it comes to FSD. It’s because they’re not sold as much as the other 2 so Tesla doesn’t prioritize them. If one of the main reasons I was getting a Tesla was FSD then I’d rather just get 3 and Y since they’re always ahead in that department.

[@SkullEZClap]

@@timepstein274 I hope it’s soon.

[@CrumResearch]

How about getting Model Y Juniper (refresh)? Will it have AC outlets a k.a. PowerShare?

[@michaelscott4766]

I obviously don't know but my suspicion is that Tesla has far less data to train the AI for Cybertruck than it does for Model Y which could lead to it being less confident.

[@FlingerFilms]

First! Thanks for your vids!

[@aadityachourasia2124]

I hope we reach full autonomy around version 15 or 16

[@TheMamonti1]

got to be weird and non trusting..... I mean like being a passenger when the wife is driving!

[@bear4278]

I remember hearing reports about there being seperate, special “influencer versions” of FSD and this video jogged that memory for me! Like maybe your Y got the “influencer” version, but then they forgot and just gave you the “everybody else” version for Cybertruck or something 😝 I have no idea if those clams of an “influencer” version were true, and even if they are, I highly doubt my theory explains the discrepancy here, but man it is wild that the same version of FSD can be so drastically different and it does make me wonder how that is possible!

[@didierpuzenat7280]

Can you explain me what would be the point for Tesla to enable the download of a "bad version" of FSD if they have a better one ?

[@bear4278]

@ from what I remember of the reports (supposedly from ex-Tesla employees from memory), the point of the influencer versions was to make FSD look better than it actually was so people watching said influencers videos would be more likely to buy it (basically false advertising to get more users). As to what was different about it, it required Tesla staff to specially and manually map the areas around where the “selected” influencers live / places they often go to make sure FSD works especially well for them. Obviously this takes a lot of time for Tesla employees to do manually and it isn’t how FSD normally works so that is why they don’t do this for everyone. I’m sure I am forgetting some of the details but the story is online if you are interested 😊 Also, again, I have no idea if that is true or if Tesla do actually do that (I’m not trying to claim it is or they do), but this situation just reminded me about it due to the massive difference here despite being the same version 😊 Oh and am definitely not saying Dirty Tesla is doing anything wrong either! From what I understand, this is apparently being done without the influencers knowledge 😊 Just wanted to specifically point that out as he seems like a good dude and am sure he would be just as cranky as the rest of us if the claims are true 😊

[@321brianh]

It’s really annoying the HW3 cars still don’t have it. I only got 12.5.4.2 last week. I find it useful enough that I don’t want to cancel my subscription but knowing that I may never get v13 on my less than 2 yr Tesla pisses me off.

[@sector7offshore]

At the stop sign it stopped because the guy on the righht was creeping into the intersection. And the first pedestrian would be hard to guess what was happening even as a human and it still did nothing wrong cause I creep up on people wandering in the roadway. You are overly negative! Close to cars on the left disengagement does count cause it wasn't going to hit the cars on the left, just close, so thats your anxiety. How about leave it in hurry mode to make comparison since it passed all your expectations so your error of mode choice.

[@upsguppy520]

hey hole your messing up other peoples lives with this crap stop making your self a obstacle for others trying to live their lives

[@TeslaReferralProgram]

Doesn't look bad at all!!