-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
98 Score - Got FSD Beta?
- Thread starter Daniellane
- Start date
-
- Tags
- 10.5 98 safety score fsd bets
So it seems like resetting and driving a super slow and safe 100 miles is the way to go. Seems like they rolled out to 98’s for a day or two and then stopped.Yup got beta with the current batch of folks
(SS dropped to 99 after driving to Chapel Hill this morning. Was 100 before that. Was 98 before I reset last Friday pretty consistently since I opted in.)
My theory was that if they are adding fewer people to the beta at a time than the current size of the pool of people who meet the current bar to get beta, they're picking prioritized on SS. You might stand a good chance getting in at 98 when 10.6 comes out and they add a bunch of people, but I think these "yeah let's grow the population by like 5-10%" pushes, 100s are going to preempt you.
I'm not on TeslaFi so I don't know about the distribution of scores, but I'd imagine there are more 98s than 100s and I see no other good explanation of why we anecdotally see lots of people with 100s and 99s getting in but people with 98s seem really bummed out. (I guess the distribution could explain this and then I'd be wrong, but based on my limited knowledge, this is what my gut says.)
And I betcha they close the SS-doesn't-get-logged-if-you-reboot-the-car loophole in a future software update. If they're basing insurance premiums off of SS, it seems like a huge issue if people can drive like maniacs and pay a low premium. They might even close the loophole where they even let you reset the SS in the first place (which I'd imagine wouldn't even require a software update to the car). If SS was just for FSD beta, I don't think Tesla would care all too much about fixing these issues, but the insurance aspect might prioritize such a thing. (I don't think Tesla cares all too much about who they're adding to beta once someone asks for it, but I think since they readily have the ability to put a priority queue that sorts by even a flawed representation of safety, they did it. Even if it's not perfect and people are hacking their way to the front of the line, it's still, strictly speaking, better than nothing.)
PDX-Y
Member
That's my suspicion too. They have some internal model for how rapidly the rollout should happen (part based on statistics about accidents, part marketing, probably some of it having to do with how they want to present things to regulators if there's another rollback, etc...). And every few days they top up the number from whoever has done the best job qualifying.
I was at 98 (barely) on 11/21. On 11/22 I drove 530 miles at 100 (the vast majority on NoA) and returned on 11/28 (only got 370 miles at 100). But this raised my score to 99. Just got the 10.5 beta today.
No worry...I was at 99 for 3-4 days. I got it yesterday. You should be in the queue!
what was score before the point ?Looks like they pulled the scores a couple days before. I sadly dropped to 97 on Sunday and also got into 10.5.
johnmodely
Member
Aww someone didn't like being called a noob lol. Facts in my statement don't change because a few hundred extra people are getting beta a week. Im sure ill get it too soon, but the fact will still remain that wide release was supposed to happen end of 2020, and we are at end of 2021 and we have to play stupid games and waste our time to even have a chance at it.
And this will probably still remain the case for people with full FSD into 2022, a full 4 years after FSD was first promised..
johnmodely
Member
Did you just get it this morning? I think my 98 score dropped on Monday.Looks like they pulled the scores a couple days before. I sadly dropped to 97 on Sunday and also got into 10.5.
OK. So back to any new successful downloads of 10.5 occurring today? I’m currently at SS 100 (for 5 days) prior at 99 (for 25 days) and prior at 98 (for +2 months). An early ‘19 M3 performance at 36.5.1. Any ideas on what’s what?
Let's use Karpathy's example of training an AI to recognize pictures of dogs.
Before the AI has been shown its first picture of a dog, it (obviously) won't be able to look at a picture and successfully label it as a dog.
Now you show it one picture of a dog. As long as you show it only that picture, it'll have a 100% success rate at labeling that picture "dog."
Its accuracy at labeling an object as a dog gets far, far more accurate with every new dog picture it is shown. There are many situations where "chasing the long tail" or "march of 9's" is applicable, and this is one of them; show the machine a million different pictures of different dog breeds, sizes, context, etc. and you'll add more 9's to the tail of its accuracy.
Let's say you've shown the AI a million different dog pictures, but then you accidentally slip in a picture of a cat during training, so the AI thinks you just showed it a dog. How bad will that affect its accuracy? Not much, since the first million pictures were of dogs, the weighting that one picture will get is very small. Now obviously you're going to want your training set to be as accurate as possible, but the odd mis-label won't be the machines undoing. In fact, it should be possible (once the machine has seen a certain number of dog pictures) for the machine to actually question the input... "Hey, dummy, you've shown me a million pictures of dogs, and I'm telling ya man... this ain't one. You wanna confirm this one for me?"
The diminishing returns point is... interesting to ponder. Yes, you'll continue to march the 9's with more pictures, so at what point do you stop? That's a question for the experts, but it isn't hard to imagine a dog peeking its head out from behind a car, such that only the head is visible. How many pictures are you going to need to feed the machine so that it can accurately label the object as a dog from just this limited view? I'm not sure the answer to that question is currently known, but a more general answer of "the more the better" definitely applies.
And so far, we're only talking about dogs! Think of the thousands of other objects the machine needs to be able to correctly identify/label, and think of the mind boggling different contexts possible for each of those objects. They're gonna need a lot of cameras producing a lot of videos, and some serious computational horse power.
Even if they fail at achieving a self driving car, they'll end up having a NN with an insanely accurate ability to pick out dogs. So at least there's that.
Sorry for the OT post... I've been watching too many AI/NN videos lately. It's a fascinating subject.
There's been quite a bit of discussion here about the need for Tesla to be accumulating such a huge video library at this point of the game, what with Dojo not up and running yet. IMO, they need them all. Data storage ain't free, and Elon isn't known for spending money without there being a good reason for it.
I just really hope that Elon doesn't find himself in the position of Tony Stark's father: "I'm limited by the technology of my time."
And thus ends this home roasted coffee fueled wall of off topic text.
Grats to you guys/gals/others that have gotten the Beta! Welcome to the E Ticket ride (you might be old if you get that reference).
Just out of curiosity, what Autopilot chip do you have? Need version 3, FSD chip. I think a lot of people don't have the right hardware and are somehow expecting to get it.
Similar threads
- Question
