DarkForest
10.69 made me a shill
And god forbid you’re going the same speed as the car in front of you but the gap is TOO BIG.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD Beta 10.69
- Thread starter Buckminster
- Start date
willow_hiller
Well-Known Member
This morning I think I can confirm Chuck Cook's advice regarding the report button:
Uploads max out at about 8-9 GB, and this morning I had a drive without pressing the report button, but I did have a few wheel-turn and brake disengagements. Only 2.3 MB up, so I'm not sure if it sent any training data back to Tesla at all.
So personally, I'm going to be hitting the report button along with brake disengagements whenever possible.
Uploads max out at about 8-9 GB, and this morning I had a drive without pressing the report button, but I did have a few wheel-turn and brake disengagements. Only 2.3 MB up, so I'm not sure if it sent any training data back to Tesla at all.
So personally, I'm going to be hitting the report button along with brake disengagements whenever possible.
Knightshade
Well-Known Member
AFAIK we always knew the reports get sent- just not if/how they're actually used.
The 5-7 is interesting, IIRC Green had previously said it could only cache 5 of em in available storage for a single drive so I wonder if they increased that?
The 5-7 is interesting, IIRC Green had previously said it could only cache 5 of em in available storage for a single drive so I wonder if they increased that?
willow_hiller
Well-Known Member
The thing that I had heard but am not so certain about anymore is that all interventions are sent back to Tesla. Previously Chuck had said that disengagements by turning the wheel or pressing the brake, and intervening by pressing the accelerator were all reported back. So if you wanted to disengage without creating a report, you had to press up on the stalk.
But my drive this morning had both steering wheel and brake disengagements, and it did not upload a lot of data. So unless interventions are sent back in metadata form only, I don't think they're all sent back automatically. Maybe certain interventions, if they match a campaign set by Tesla. But certainly not all of them.
powertoold
Active Member
I still doubt they prioritize snapshots. I see so many people snapping preferential / comfort / silly issues constantly. If I were Tesla I'd prioritize hard brakes or steering disengages (more likely safety issues). Between 160k users, you'd probably get 10,000+ high priority disengages daily. Then you also need someone to sift through them.
Knightshade
Well-Known Member
AFAIK disengagements are sent, but VERY basic data--type/location/etc... not any video which would be the bulk of size of snapshot uploads.... see this from Green on that topic
Next tweet after that one has him adding "no pictures in those fixed disengagement reports so they are really small (under 1kb)."
I think the auto-labeling can "sift" through them, however. Then when they need snapshots for something particular, they can search and retrieve the auto-labeled clips and add them to the training data. Same with shadow-mode clips.
....but that leads to people like me and our potential problem. I don't have WiFi for Tesla to upload my info. I suspect that all my Snapshots are just emptied into a blackhole (deleted) because of this. Wish Tesla would generally define what info it gathers and who from (WiFi and Cell or WiFi only). Would LOVE if then sent some type of confirmation Notification stating in General terms what and when they got from your car. This would make us all feel more rewarded, important/useful and would probably encourage us to use Beta even more (though I use it almost all the time).
I will still keep hitting the Camera icon button even though it feels like it is in vain since I don't have WiFi.
That's good info.
It feels like they weigh training more heavily on the most current issue to be solved as I can feel the effects of Chuck's ULT in totally unrelated scenarios. If so training might have an inherent forgetting factor for previous successful scenarios. It's odd how FSDb occasionally allows a turn into plainly obvious oncoming traffic. And at times a well laid out road construction bollard is ignored for FSDb to instead follow roadway lines. Or the over compensated herky-jerky steering wheel action causing a crossing over the double lines to the other side of the road. It's as if NN training fails to properly converge to the optimal solution. Looking forward to the 69.3 rollout.
Last edited:
willow_hiller
Well-Known Member
Chuck went on to say it does upload over LTE. Just more slowly. So some of your clips may be overwritten before they could be uploaded:
There's a fixed amount of space allocated, and snapshots can vary in size:
I wonder if it would be worth it to expose more options to public testers for tagging snapshots and so forth. I remember seeing screenshots of dev firmware that had a whole bunch of options for classifying the snapshots. I.e. "comfort, lane selection, blinker usage, etc."
Yeah I'm thinking that too, would need to add a clip browser and let people annotate them after a drive. Maybe too much work.
If it worked like the dashcam clip browser then it would be a useless feature too. That sucker never works for me in the car unless I’ve recently deleted almost all the clips from my ssd.
DarkForest
10.69 made me a shill
That's exactly how I assume they do it. I thought in last week's AI Day, they briefly mentioned how they process the snapshots that get uploaded, so they all do get used to some degree or another. And like you said, if they need to find something in particular, they can search through their readily available collection of snapshots.
I don't see taking a snapshot as a "FSD did something wrong" per se—more that, I think this snapshot is something the algorithm should train on to become better. So I try to do my part with snapshoting a few things per drive (see attached haha).
Regarding 10.69.2.2, something I've started letting it do more is turn onto a street when I would not have. I figure worst case scenario, if the gap isn't big enough, I can floor it, but so far I haven't had to do that. What used to happen to me all the time was I'd wait for a big gap because I'd judge cars to be going faster than they really were. Every time FSD has made a turn when I wouldn't have, it turns out traffic was going much slower than I thought, and FSD had judged it correctly. I don't plan on becoming complacent on this, but I am willing to give it some leeway if I think I can do so safely.
Attachments
gcsutton
Member
2022.20.18 installing now.
Natal
Member
Similar threads
- Article
