I'm not sure, normally when there are gaps it because the video seems to skip ahead a few seconds and you can't get the in car viewer to play at the correct spot. On the clip im watching the video does not skip ahead, the time stamp never stops on it, it really just seems like it was cut out buy tesla.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Wiki MASTER THREAD: Actual FSD Beta downloads and experiences
- Thread starter jebinc
- Start date
FSDtester#1
Craves Electrolytes
Installed Beta in my Plaid a few minutes ago... going to spend a long time in the bathroom to make sure I am driving on "empty" before I test it. I hope all of you other 100% and yokers got it also. Be safe and empty out 1st or wear a diaper!
Since i still have the safety score (at least for a little while) I can see what I scored on my drive into work this morning on mostly FSD. I had one instance where I took over and pretty sure I did some hard braking, but I didn't get a ding for it. I think the 3 second rule that applies to autopilot also applies to FSD.
Before safety score went away, I had to intervene twice by swerving abruptly, and I was dinged for those. Definitely less than three seconds, as it was the swerve that disconnected FSD Beta.
Last edited:
Well I had the Plaid drive me to Panera this morning from home. Stress level 9000. This thing made many mistakes including right off the bat hogging the road in my subdivision and likely pissing off 2 people coming the opposite direction. I had to take over. When existing the subdivision, it again hogged the road and made it too tight for people trying to turn in to the subdivision. At the next big intersection, it didn't want to get in to the turning lane and did some weird hard braking before the light. I again had to take over. Then it was supposed to make a left turn after 2 more stoplights but decided to try to take the turn at the first stoplight. This is a brand new intersection/lights so Tesla probably has no data on it. I had to take over a few more times before reaching destination just 2 miles away. The way home was much better but I still had to take over a couple of time. It did much better in my subdivision coming home by leaving plenty of room for oncoming traffic and driving around parked cars in tight spots. Overall it is cool but clearly not ready for prime time... at least in my area. I hope it learns quickly, otherwise this thing may age me 1 year every drive.
Also, the yoke spinning around like mad is an experience...lol
Also, the yoke spinning around like mad is an experience...lol
FSDtester#1
Craves Electrolytes
FSDtester#1
Craves Electrolytes
I am still in the bathroom,... I couldn't find an adult sized diaper for my first Beta test
Callahan27
Member
The best things I appreciate about 10.2 are the visualizations, no more confirming stoplights and better straight line driving. If pathing was better and FSD would be able to think ahead more it would be much more enjoyable. Looking forward to 10.3.
FSDtester#1
Craves Electrolytes
Okay thanks, I bet there weren't many beta testing being done there before. You should be a good test case for the system. I am in a pretty dense area, but not a lot of the city like driving experiences a lot of you have. My commute is 85% highway going against traffic. So there are almost no cars to and from work. So should be interesting for all of us.
The 10.2 release reminds me of the first release of Navigate on Autopilot. It behaved much differently than people expected and many said they’d never use it again.
Of course, NAV got better after a few releases, and I think those of us who used it contributed to the data used to improve the neural net.
But 10.2 is high stress. I’ve had it for only five days, and I‘m learning how it behaves and when to anticipate taking over or tapping the accelerator to speed turns. It’s a learning curve since it specifically does not drive the way I normally do, even tho I did score a 100 on the Safety Score…
Stay safe out there, I know I’m trying…
Of course, NAV got better after a few releases, and I think those of us who used it contributed to the data used to improve the neural net.
But 10.2 is high stress. I’ve had it for only five days, and I‘m learning how it behaves and when to anticipate taking over or tapping the accelerator to speed turns. It’s a learning curve since it specifically does not drive the way I normally do, even tho I did score a 100 on the Safety Score…
Stay safe out there, I know I’m trying…
scottf200
Well-Known Member
Am I confused as I thought the 2nd number implied the minimum number of miles you had to drive. 100/100 was mentioned by Elon in an early tweet and then at a later point he mentioned a score of 100 with at least 100 miles of driving. So I thought that 99/100 meant SS of 99+ and 100+ miles.
He mentioned 2017 X with a MCU1-> MCU2 upgrade. I don't follow your question about HW2 or 2.5 since he would have to have upgraded to the HW3/FSD computer to get into the FSDBeta? I have a June 2017 X HW2 and did the MCU1->2 upgrade. I paid for FSD at the time of purchase Dec 2017 and had it upgraded to HW3/FSD computer when it became available.
Bill Price
Active Member
Holy sugar!! If you weren't so lazy? After reading your post, I've come to the conclusion that we are lucky you aren't industrious?That was a fantastic post, and exactly what I would've written myself if I weren't so lazy.
But touching on the "is Tesla really benefiting form this?", "How do we test?", etc. subject...
Let's imagine a scenario where there is a small, pre-defined set of rules that must be followed when streets are built. ALL of the specs of the road, from the width of each lane, the type of asphalt used, the color of the paints used, the markings.... ALL of it are preset rules that must be followed.
Programming the navigation of these streets would be fairly straight forward. But the part that isn't so easy, is the "what is the other driver going to do?" problem. That's hard.
Now add the fact that streets aren't vanilla. There are so many variations that you just can't go by a pre-programmed set of rules. What do you do for these instances? Elon, et al, refer to this as "chasing the nines" or "chasing the long tail." Programming for 99% of circumstances is easy. Programming for 99.999999999999% of circumstances is hard.
The solution to the "what is the other guy going to do?" problem and the navigation problem? AI/NN.
But how do you program an AI? One of the Tesla AI guys (sorry I don't remember his name), used a great example of training an AI to recognize pictures of dogs. The best way to do this is to show the AI as many different, varied pictures of dogs as possible. And the more dogs you show it, the better it becomes at recognizing dogs.
Tesla is using this approach for training the AI about driving; showing it as many variances to every day occurrences as possible so that it becomes better at recognizing its world.
This same thing can be used to help with "what is the other driver going to do?" Show it many instances of that situation, and look at the outcome of what the other driver did. And repeat this millions of times. You end up with a pretty good idea of how the average human is going to react in a given driving situation.
This approach requires mind boggling amounts of data.
Your car isn't wasting bandwidth when it spends the whole night uploading gigabytes of information to the Mothership. It really is needed to teach the AI about the world it sees, and how to interpret it.
And as others have mentioned, Tesla can set triggers for the car to automatically send video clips of things they are currently working on.
So really, the "best way to test" is to just drive the car as much as you can, in as many varied instances as you can find. No need to take my word for it, Tesla has actually said this several times. I'd find an instance of it, but I'm too lazy. Go look for yourself if you're interested, I'm sure it wouldn't take long to find.
Let the car and Tesla determine what they need.
But you are not wasting your time. You really are helping, one gigabyte at a time.
One guy mentioned that there is no way they could have enough engineers to manually watch all of these clips, and mark/catalog them all. That's completely correct... they can't. So they have developed a machine to teach the machine. And how accurately the teaching machine can label all this stuff correctly is, once again, dependent on showing it millions of examples.
I'm no expert by any means, but IMO, this is going to require YEARS worth of both data and manual labor. There's going to need to be a person manually teaching until the machine becomes accurate enough to label stuff without a human's help. Once that goal is reached, things will really take off, as (obviously) a machine can scrub through videos and label stuff far, far faster than a person could.
All of the above is taken from things that Tesla has already stated. This is just a TL/DR version of all of it.
Last edited:
I think the question is specifically about the cameras. Do you know if your config is 2.0 or 2.5 cameras?
Tesla upgraded the cameras sometime in 2017 and thus far nobody with 2.0 cameras/ MCU2/ HW3.0 has FSD beta.
You would be the first report.
Similar threads
