Cybertruck FSD Supervised v12.5.5.2 - First Impressions

[@DirtyTesla]

Y’all Chuck is NOT kidding that that turn takes guts

[@user-hg6nz1cz4l]

Chuck Cool

[@ChuckCook]

Thanks for the support Chris. You would know because we did it together first hand. Thanks for the comment.

[@Someone_Cool.x3]

You both are the best ❤🎉

[@CyberTahoe]

@@DirtyTesla I feel gutsy letting it do a left turn in a 4-way stop intersection 😂 This is next level!

[@ericchild3363]

Chuck, as someone without either FSD or a tesla, I find your commentary to be just right, please keep going as you are

[@ChuckCook]

I appreciate that!

[@Plexipal]

Keep doing what you're doing, longer content the better. Maybe a coast to coast run Tesla said they would do one years ago but haven't yet.

[@flyshacker]

Chuck, you, are, fearless!! Excellent testing job! That first unprotected left, I think, was a problem exactly because of what you surmised - your CT was waiting for that car in the median to clear the median, and meanwhile that car was waiting for you! I don’t think Tesla could pay for a better tester than you are! I think your intense pilot chops help a lot.

[@ChuckCook]

I appreciate that!

[@deanlevang6031]

Gutter guards. 😅 Good description of action taken.

[@Someone_Cool.x3]

Love all your videos and commentary❤🎉

[@mikemulrooney4824]

Gosh,that truck is so kool.

[@hugh3257]

Chuck, I have been through all the FSD updates since 2019 and submitted most of my driving time to helping FSD get exposure to my driving. However, I am only at 12.5.4.1 on updates! Should I not be kept more in the loop with more updates?

[@hollowmoose]

The U-turn was slow but he didn't stop for the stop sign. If he had it wouldn't have been note worthy

[@SteveSmith88]

At time 4:34 there appears in the rear view camera display - a “refracted light” triangle. ?? But it did, shortly thereafter handle, correctly commit to the turn. Just sayin’ well done.

[@MatthewKruegeriscool]

I’ve personally thought the model was in separate steps, like checkpoints. So the output is a specific net to CT, but the path planning is universal, but vision may be CT specific (but easy to train with just a few minutes of lidar from a reference car)

[@SpaceTravel1776]

@5:20 The car turning right from a stop sign usually has the right of way in that situation, but it was very close as the car was slow-rolling the stop sign just as FSD decided to go for the U-turn. These are coin flip situations I think FSD will always get blamed for if there is an accident, even though the other car was illegally slow rolling and didn't come to a stop.

[@Test6837]

Hey Chuck, will you go again to the super speed roundabout with the Cybertruck?

[@ChuckCook]

yes!

[@mikemulrooney4824]

Good vid

[@ChuckCook]

Appreciate the feedback!

[@shallenberger57]

Chuck, those little jerky adjustments are 3 times more noticeable in the Cybertruck because of the steer by wire and only 350 degree wheel movement. It wouldn't have been noticeable in the model Y..

[@ChuckCook]

It was the acceleration I felt, not the steering. My body moved.

[@shallenberger57]

@ChuckCook I was talking about that left hand turn down the short street

[@AVerySmallTinyMinisculeH-bc1lv]

9:35 Your commentary on a neutral vehicle makes me wonder how much improvement will come when Giga Texas’ Cortex comes online. Right now your truck is emulating the model rather than using a model specifically trained for AI4/HW4.

[@seussiii]

I'm kinda surprised they haven't push the angled parking yet.

[@bencoman]

Interesting idea about the braking curve being affected by adapting training from neutral vehicle to the heavier Cybertruck. I wonder if such vehicle parameters as mass and shock absorber performance would be useful inputs to the training model. In terms of adjusting to different camera positions.... If Tesla are no longer compute constrained, it seems less optimal to do most of the training on a single fixed "neutral" vehicle and then as a final step fine tune for a specific vehicle. Neural networks become more robust with noisy augmented training data. So instead a there could be a training preprocessing step to build a 3D world and from that generate the views from different camera positions to use as augmented training data during the main training cycle. Encoding the camera position with the augmented data may allow the neural network to be adjusted to different vehicles by "simply" defining new camera positions.

[@sillystuff6247]

on first failed ULT, i wonder if the car behind CT, was the factor that caused FSD to not commit?

[@sillystuff6247]

i think cybercab can & will circumvent tragically tricky traffic turns, like chuck's ULT. cybercab will reduce its "degree of difficulty" to become a commercial product sooner.

[@adamnowak926]

👍

[@CyberTahoe]

Definitely still rough around the edges, hopefully catches up to 12.6 status quickly...Speed profiles and set max speed will make it much better for me, in addition to them hopefully fixing the hard braking/lack of following distance. Or i guess we could jump straight to V13 🤷‍♂

[@Someone_Cool.x3]

V13🙌

[@johnmoe6055]

Wouldn't the criticism of the FSD/Autopilot safety numbers be that perhaps users turn on the system mainly in the easy cases?

[@davec2211]

Just wondering, Would adjusting the gap setting to the vehicle in front have an effect on its breaking intensity?

[@Atr6015]

😂 guts 😏

[@jeffbransky7966]

I wonder if the unprotected left turn will work in other similar situations. How about trying it out in other locations? Will the question mark turn work everywhere there is a median?

[@ChuckCook]

Good point, I can change this up.

[@benjaminkindle1841]

Do you think they should add a beep about 200ms before it "jumps" into traffic? Waiting for it to jump into traffic is definitely the most stressful part of FSD for me, so I think a quick warning would be helpful

[@gabrielcastlebary2836]

@@benjaminkindle1841 if the hardware is actually running neural nets like Tesla says, you will get a LOT of 200ms beeps. Or the car might commit to something it shouldn't if the 200ms beep happens prematurely. That is to say, the nature of neural nets isn't really supportive of a warning feature like this or a "why am I doing X". It's certainly possible, but not in the early stages that we are right now where NNs seem to be so close to the end interface. Also 200ms isn't a lot of time.

[@BenefitOfTheDoubtInquiry]

Insane the local government doesnt already have a "No Left Turns" sign.

[@wolfgangpreier9160]

In Europe such roads would not be allowed. But they are fascinating!

[@jimmccall8030]

Seems like added or relocated cameras would help with the first situation

[@jerryaguirre23]

Did not notice a screen change when entering and exiting freeway. Could they have merged the stacks? It said "CHILL" the entire time.

[@leanderleeco]

A lot of situations where it didn't feel like it was the Tesla's fault

[@collegepros]

Hey Chuck. I’ve been wondering for years, how Tesla is going to handle extreme fog conditions that I occasionally run into. How is a camera only system going to fix that. It seems that you’d want to design a system that can ‘see’ better than a human to avoid harsh snowfall and foggy conditions. Would a FLIR (sp?) solve this? Thanks, and enjoy your videos!

[@richardvivian3665]

It can’t work any better than a human in extreme fog because sonar is not 100% reliable. In any case, it’s better that the car behaves just like a human in all conditions because you don’t really want a superhuman car doing it’s own thing in low visibility. It could cause dangerous unexpected situations for other drivers.

[@BenefitOfTheDoubtInquiry]

Drive slower, only as far as it can see and safely stop, the same way we do. Along with maps of the area, it should be just fine. If we want the cars to drive faster than humans in fog, I think you'd have a harder time convincing humans that it's safer to do so than getting the cameras to exceed normal human speed.

[@0i0d]

Exactly the same way you drive in fog

[@gabrielcastlebary2836]

I read an InsideEVs article a moment before seeing this comment where the writer made the delusional suggestion to add lidar or radar to "make things more reliable". First Lidar has the same problem in dusty or foggy conditions. Second, I THINK radar just doesn't have the resolution required to even help supplement. That said, I think the answer to your question is simply that the system has to slow down like a human would.

[@Someone_Cool.x3]

Sandy Munro thinks so ❤

[@PeterQuistgard]

Cool, but is this a significant difference from 5.6 on Model Y?

[@ChuckCook]

Very different.

[@kadeembest4752]

@@ChuckCook Is it different in a good way or a bad way?

[@DonkeyKong96]

Siiiiiiik

[@Muonium1]

I woke up to a new free trial month on my M3LR today. It still camps out interminably in the fast lane, for miles and MILES and *MILES* on end, regardless of aggressiveness setting, while countless cars pass me on the right. It's still a hard no from me even at $100 a month, let alone the laughable $8K cost to buy permanently.

[@edhill8568]

Unless you have the new end to end highway update you have to make sure your setting is at chill or average. If on assertive FSD typically will not move out of the passing lane. Been this way for a long time. V12.5.6.x will address that based on recent reviews.

[@Someone_Cool.x3]

Do you have HW3? What FSD version? 12.5.6 seems good from beta tester perspective so, prob only a matter of time. Tesla has been pushing the subscription model for quite some time...makes no sense to purchase FSD upfront. I will turn on signal and force into the right lane. I have no patience for left lane campers lol.

[@Muonium1]

@@Someone_Cool.x3 HW4. I do push the right turn button to get it into the middle lane, but it's CONSTANT and often just exits immediately again.

[@EVentures888]

Why don’t the developers incorporate a new feature that I will call “record my trip” that will teach/train the FSD software how to drive a certain route. This feature could accelerate the learning for the algorithm. The FSD could learn how fast to execute a u-turn like a human driver, etc.

[@lecisko]

Apparently it's a lot more complicated than that. Mainly you need a shitload of data

[@wolfgangpreier9160]

They are not allowed to. The NHTSA says everyone must stop at the stop sign. Only problem: That is not how US drivers drive their cars. Therefor: Your route is not allowed to be fed to the AI.

[@0i0d]

That’s what already exists. They take data from our drives and incorporate it into new builds. Doing that in realtime isn’t possible atm due to huge hardware constraints (we’re talking many orders of magnitude)

[@billw.5964]

@@leciskoYes, my guess is that it isn’t that easy. Maybe in a year or so.

[@gabrielcastlebary2836]

The cars already record your trip. That data is uploaded to an AWS bucket where it can be auto labeled and processed for model training. How is what you are suggesting better?