Search results

Just test drove 12.3.3 model Y HW4 at Tesla today It was noticeably smoother with stopping and accelerating than my old 2020 HW3 Y Part of it may be the improved suspension, but I felt like it was more than that The fsd 12.3.3 performance was very impressive on that HW4 Y... I can see why...

12.3.3 is a regression for me vs 12.3.1, more disengagements in my area, more robotic motions once again

Glad u included a piece of your junk

I don't think Tesla will train on HW4 videos until they've reached a local maxima with HW3.

Elon confirms my thoughts about HW4 being run in emulation:

During the first autonomy day, Elon responded to an analyst's question about whether he meant level 5, Elon answered "yes," although I don't think the acronym "SAE" was included in that response, but that's probably what he meant I don't know the specific time, but it was during q&a I believe...

That's V11 though?

12.3.1 improves curbing risk, but it still gets uncomfortably close

End-to-end pretty much rules out L3 I guess they can run the V11 stack as the ODD referee, but it's a chicken and egg problem: if the V11 stack is good enough to reliably identify the ODD limits, then V12 should be able to handle it.

After my post, I gave it some thought: 1) those who try the trial will have some residual interest in fsd 2) as fsd gets better in the coming months, the ones who finished the trial will have some lingering interest to try the future improvements, see how they're better than 12.3.1, which...

Curious decision to make 12.3.1 the ChatGPT moment. If there will be significant improvements within the next 1-2 months, why not wait until after those?

Just fyi, Camera calibration is currently two parts 1) autopilot / NoA calibration - takes about 20-30 miles of highway 2) fsdb calibration - takes an additional ~20-30 miles of highway During calibration parts 1 and/or 2, it will seem like it's struck at a certain percentage for miles. It's...

12.3.1 tends to speed up to the limit more liberally, but it also depends on how fast surrounding cars are going I don't think the slow speed issue is resolved, but it's improved with 12.3.1

Wow check out this 12.3.1 closed road handling:

12.3.1's speed control is more robotic, but overall it makes fewer disengagement-worthy mistakes vs 12.3. I think 12.3.1 would have a lower tendency to curb the rims/wheels Also Tesla keeps tweaking the V11 highway stack, so 12.3.1 behaves differently vs 12.3, 3.1 tends to get in the rightmost...

Pretty sure Ms Jill has HW3, I remember looking it up before, when she bought her Plaid

More decision wobbling with 12.3.1:

Based on my drives and Omar's video, 12.3.1 has more decision wobbling like this:

12.3.1 sucks so far compared to 12.3, experiencing a lot of hesitations and multiple stops at stop signs Just overall feels more robotic

My Model Y also gets very close to curbs and whatnot

Sounds like the decision wobble, it's much improved from 12.2.1 -> 12.3, it was fairly common with 12.2.1, hopefully it continues to improve

Stupid question: does he mean 3 improvements every 2 weeks or 1 improvement every 2 weeks, 3 in total? Smart question: does it even matter? :)

The most surprising thing about V12 is how consistent it is. The behaviors have been fairly predictable and consistent for me. It's easy to anticipate when and where it might fail If the improvements are only coming from increased training and compute, then it's very promising

The amazing thing is that V11 was/is required to get to V12... In order to curate / test / simulate / etc. the data, you need V11 heuristics

I wonder if Tesla trains videos specific to each car model. We have some evidence that each model is separate? 1) For Chuck's turn, he saw all kinds of Tesla models. 2) During the Elon livestream, Ashok mentioned some issue specific to Model S.

It's funny, no matter how good V12 gets, I want more and more. I want it to pull out of my driveway, out of parking spots, everything! I want to get in my car and turn it on and not have to touch anything. But first, if it can gracefully deal with road closures, that would be amazing and make...

Wowwww if 12.4 performance jump is similar to 12.2.1 -> 12.3, I've run out of mind blowing words

Fsd just needs to 1) handle closed roads gracefully 2) not squeeze into bad situations Then it's golden FSD ADAS It feels so close (for the bay area anyway) Just drove through SF during st paddy's traffic and peds, it was incredible except for a closed road (I didn't want to let it try :) )...

BTW, be careful out there, 12.3 does get very close to curbs:

12.3 in a nutshell for bay area, CA: 1) less nags 2) does extremely well with "typical" city driving 3) fails with uncommonly encountered scenarios. Doesn't fail horribly, but it's annoying when it tries to get in the wrong lane and you have places to be, etc. 4) parking lot behavior is not as...

12.3 is definitely amazing. In the bay area, it makes the city driving performance and reliability close to the experience we get on highways. With 12.3, I activate it the moment I back out of my parking spot or driveway, so I give it a lot more chances to mess up. I expect it to do everything...

12.3 drives very very well, But you guys are right too, driverless still seems a long long way to go, especially for HW3 V12 is amazing ADAS, but...

Maybe has something to do with it being a parking lot as well? I wonder if the driver pressed the accel or it went by itself

Y'alls don't care about my opinion, and you haven't tried 12.2.1, so you don't have reference, but 12.3: 1) Way better than 12.2.1 2) Not perfect in parking lots, but way better, it's logical but drives like some alien that hasn't driven many parking lots 3) In general way better 4) Doesn't...

Wow these people are starting to sound like me

Here's confirmation that 12.3 *can* go to HW3 cars:

An fsd feeling is worth a thousand words :) Btw in that first drive, I see a lot of lane change decision wobbling Speed stuttering has much improved I prefer the average auto-max setting, Omar uses assertive auto-max, so I'm not sure how that affects the drive performance

12.3 looks to be HW4 only so far, don't hold your breath

Well, then they can simply remove the posts. It's odd that these people are deleting their entire accounts, along with YouTube in the case of teslaaigirl I guess Tesla is asking them to do it or risk termination

If so, why would they need to delete their accounts? They'd just get beta revoked

You could also argue that the left lane is supposed to be left turn only :) but you're more likely correct. It would be confusing for many drivers new to that intersection though.

I think 12.3 performed very well on that intersection as well. The adjacent truck started moving but ego was obstructed on the left by the truck and also there was only 1 lane to go into, so ego waited until truck passed and committed.

Here's a new 12.3 video from Kim, it's more city streets. The only real juicy part is 14:40 onwards, does very well with the construction closed lane:

You mean the Kim guy? He was driving mostly highway in that video, only about 4-5 mins of city. Almost all of his commentary were highway situations if I understand correctly.

When you guys get V12, you'll be surprised how stable, coherent, and reliable its decision making is. I had my doubts about end to end video training initially, but V12 is way more predictable and stable than I anticipated. It doesn't randomly hallucinate a car and swerve or stop for no reason...

BTW, I'm pretty sure TeslaAiGirl was using 12.3 on HW4, based on the smoothed out interior camera housing. HW3 has a more crescenty shaped interior housing: HW3:

BTW what's your assessment of V12 right now? Does it look promising for geofenced robotaxis within 2 years, once it gains ability to reverse?

You're right. There can be overfitting wrt data curation: how many examples of each, etc. Even using hired drivers for Chuck's UPL is an example of possible overfitting. I was mostly referring to the idea of overfitting wrt V11 heuristics that were most familiar with the engineers in the Bay...

Yea I don't know why this is so difficult to understand. Perhaps during testing, the driver had to disengage and complete the maneuver manually? And maybe they're trained to constantly keep their hands on the wheel as if they're manually driving? Also, "looks like manual driving" doesn't mean I...

Yea except even chuck thought it was for validation in his post today. You say that they had drivers manually driving for weeks, but Chuck didn't know that (Chuck even said so himself that he didn't know). Nobody knew for sure until Elon confirmed it today. That was my point: Before Elon...