Yes, but unfortunately some need to remove their fanboy tinted glasses while doing so
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD rewrite will go out on Oct 20 to limited beta
- Thread starter diplomat33
- Start date
powertoold
Active Member
Yup, most of the fails are related to the car going to fast towards something it isn't sure about (mostly intersections and hill crests). This will improve over time, very quickly IMO. 3 months later, many of the fails we see today will be solved. Then what?
Anyone who's wondering what the progression is like, take a look at when verygreen first showed off / enabled the traffic control automatic stopping. Then compare it to what it is today.
Last edited:
powertoold
Active Member
What you guys underestimate and what I constantly try to emphasize is that Tesla has been working very hard for very long on achieving a fleet input and perceptual training feedback loop that is accurate and reliable. Once they've been able to essentially perfect one aspect of the perceptual feedback loop, which they call the "data engine", they're able to apply the data engine to all manner of perceptual and prediction challenges involving FSD. They've demonstrated that they have a working data engine with the traffic control feature. That's why I said that the traffic control feature was a proof of concept for FSD.
Y'all should heed Elon's word this time:
Y'all should heed Elon's word this time:
Last edited:
The traffic is moving swiftly. It is not stop and go.
And he is passing all the traffic by being on the right side of passed cars not the left. In a lot of it he is in the right most lane of three lanes and is still passing everyone. Other drivers could miss he is approaching that fast on their right side.
AlanSubie4Life
Efficiency Obsessed Member
Maybe it is a mismatch on what we define as FSD?
I don’t really see that L4 will be coming from FSD in the next year or five.
I think it’s entirely possible it could be pretty good and go a couple or maybe even 10 miles without interventions in complicated situations relatively soon, and be perhaps 10x or 100x better than it is currently, in the next year, though!
I just don’t see how that is all that useful (for me). I mean, maybe it is useful for some people but I am not sure how safe it might be.
To be clear, this is fine as far as I am concerned. I just wanted better freeway driving and the better active safety features, which will likely be possible with HW3 at some point. I don’t ever expect my car to drive itself anywhere point to point while I read a book or play with my phone (even if that is legal somehow), and that is fine!
I think there’s a better than 1 in 4 chance of the HW3 computer saving me from doing 12k in damage to property, or bodily injury, in the next 5 years. (Assuming I am not using city FSD of course.) I guess that doesn’t say much about my driving? Probably the chances are lower than that. But I think that with reasonable further progress, the tradeoff here might be somewhat rational. It sucks to be in an accident. Especially an at fault one.
Last edited:
willow_hiller
Well-Known Member
Wow, just browsing through Twitter, it seems like some vehement Tesla critics have very quickly shifted from "FSD isn't real" to "Tesla pre-mapped every single road shown in FSD beta with LIDAR because the path prediction is too good for vision."
powertoold
Active Member
When I first saw how much the traffic control feature initially failed, I thought it'd take years for it to get widely deployed, since we actually don't have *any* NN-derived feature that we can rely on right now, except for many of Tesla's NN-derived features. Do you know of any NN feature that you use and find to be super reliable? Even autocorrect and dictation are still unreliable after 5+ years.
The reason Elon says level 5 is coming faster than we think is because we have underestimated and/or don't fully appreciate what machine learning techniques can do, once they are "perfected." If you look at the recent history of ML techniques, you'd see surprising things happening that people thought were years away. AlphaGo and Deepfakes and OpenAI Dota to name a few.
Yeah, I think I'm going to be wearing my Tesla fanboy hat for a while. Looking forward to seeing how this progresses.
DanCar
Active Member
Is because Elon doesn't understand what Level 5 is. It won't be here this decade. A strong level 4 will be amazing. Level 5 means it will work in snow, ice, rugged terrain, pass on a two lane highway with limited view, etc...
Last edited by a moderator:
Yup.
To be fair, I don't know exactly how far this process can go and operate on HW3.0, but it's clear theres a lack of data driven algorithm development experience here.
Gains can indeed be rapid. Definitely much faster than any other AV company's has been.
Labeling might still end up being a bottleneck when thisnis released to all paying drivers though.
powertoold
Active Member
I don’t think that there is any defined threshold for statistical safety of a level 5 system. From my cursory understanding, level 5 just means the system is capable of navigating through any routinely driven roads, which as you pointed out includes snow, heavy rain, not sure about ice though.
From what I’m seeing of the FSD beta, Tesla is very close to a level 5 system that is nowhere as safe as a human but *is capable* of navigating through all of the routinely driven roads.
captain_sahaj
Member
Hallelujah. Someone gets it. Someone finally calls out Musky
That's pretty much the same point I was making a bit further up thread - the difference between a Level 2 system that works in all conditions and a Level 3 or 4 or even 5 system in those conditions is a business decision, not a technical boundary - the company choosing to make it a higher level system and accepting liability for any accidents it has.
Obviously, they won't do that until they're convinced that number of accidents will be fairly small. But it seems like they're on that spectrum now, where the system is capable of doing everything in at least a mediocre fashion and they have to decide if it's good enough and how much it can improve.
As much as the re-write has restored my faith in FSD, I can't see L5 being possible with the current hardware. Remember, L5 means no controls necessary and should be able to operate in any condition. L3? Yes, with a geofenced L4 being a distant possibility.
AlanSubie4Life
Efficiency Obsessed Member
You can safely ignore these critics. Not only is this kind of a dumb initial position for a critic to have (since it was just obviously wrong even before FSD initial release), the critics without a financial stake in the endeavor actually are critical because FSD IS real.
Last edited:
Another video from Mr Locke and his wife on the new update .11. Mostly pretty good, but it’s clear that the system has no idea how to handle bike ones between normal lanes and turn lanes.
haha, yes exactly
to be honest, i had the reinforcement learning intuition before watching that talk (just due to the nature of the problem) and I wasn't sure why others weren't doing it already, but I was quite happy when geohotz mentioned it's the future.
Has anyone seen a video indicating how the beta handles freeway lane merges and known phantom braking situations (such as certain freeway overpasses that “scare” cars)? In addition to those issues, I have a couple sections of my commute with soft curves on a divided highway where the car suddenly tries to change lanes without warning and wonder if anyone has videos showing how the beta treats those known glitches?
i understand why the videos seem focused on how the beta handles the new stuff — especially intersections — but hope someone addresses how the beta handles previous shortcomings that most heavy users experience each day.
i understand why the videos seem focused on how the beta handles the new stuff — especially intersections — but hope someone addresses how the beta handles previous shortcomings that most heavy users experience each day.
Last edited:
Similar threads
- Question
