-
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
Update
Just came back from a 100 mile drive with a combination of city/residential and highway.
Good
- On ramp and off ramp continues to improve for me. Biggest improvement is slower exit speeds on sharp exit ramps.
- FSD actually turned onto my street. Slowed down earlier and made a reasonable right hand turn. Maybe they actually addressed several emails I've sent.
- Phantom braking continues. Wasn't worse just continues to be very annoying. Happened on both city/residential and highway. Nice sunny day.
- Wheel was very jerky. This seemed to improve in 10.3 but 10.4 reverted back to 10.2.
- Lane changes were more abrupt then 10.3 and not always correct.
- Tried to pull out of side street with a car extremely close. Had to break hard. It's done this before on this street but had been better on 10.3.
- City driving was in Manchester NH where I haven't driven before so I was in unfamiliar territory just like FSD. Several disengagements.
Last edited:
I have had to restart my BT on the Android phone a few times with this release. But the problems were a complete failure to connect while inside the car, requiring the key card to drive. No intervening BT connections or changes on the phone. That has not happened for a year or so. Have not had problems with unlocking, however.
While everyone is waiting for 10.4, I've time stamped a particularly interesting part of Tesla's AI Day video. It shows how objects were classified four years ago, before the complete re-write of FSD:
And how driving over the same area multiple times helps the system learn:
And how driving over the same area multiple times helps the system learn:
Last edited:
EVNow
Well-Known Member
I had a Target run yesterday (10.3.1) - longest fsd beta run I've taken to date.
- Only one disengagement when it failed to correctly turn at a somewhat complicated traffic signal. I just disengaged in multiple roundabouts.
- Handled all the other traffic lights & stops correctly.
- Had to disengage a couple of times to force the route I prefer. They really need to get that waypoints feature out.
- Non-rainy in daylight.
- All in all, I think I'm getting used to fsd beta, like I got used to AP.
- Only one disengagement when it failed to correctly turn at a somewhat complicated traffic signal. I just disengaged in multiple roundabouts.
- Handled all the other traffic lights & stops correctly.
- Had to disengage a couple of times to force the route I prefer. They really need to get that waypoints feature out.
- Non-rainy in daylight.
- All in all, I think I'm getting used to fsd beta, like I got used to AP.
I’ve had the walk up and no recognition no unlock issue with this 10.3.1 since the beginning. I’m not sure if the sleep timer, sleep algorithm or something like Tezlab is forcing sleep more often than in prior releases.. Tiny bit annoying at 05:00 but for the rest of the day its fine.
But i will note, PAK works fine once I use the app to unlock the car.
OK, just one last moment from the AI Day video, promise. 
It deals with one of those "how in the world are they going to program for that?" type scenarios.
These guys are just phenomenally brilliant.
So I've been reading around the 'net a little bit, and it is very easy to get discouraged by the current discourse; guys that have been in the beta since version 8 (which was released in October of '20) are saying that there hasn't been a lot of significant progress when you compare 10.4 to 8.0.
That might (and probably is) the case from a user (driver) standpoint, but when you look at what they have already accomplished and are accomplishing behind the scenes, the amount of "progress" is far, far beyond what is visible and noticeable from a user's standpoint.
If you'll take the time to watch the AI Day video, you'll see what I mean. Yeah, at three hours long, it's a long watch. So break it down into pieces and watch it over several days or even a week.
It is truly eye opening, and will completely change your perspective on FSD. I'm finding that I have a lot more patience with FSD when it does something silly now.
IMO, Tesla now has solutions in place for the vast majority of driving situations; it's just a matter of time until they get the neural nets and AI trained/tuned to use those solutions.
After watching the video, you'll really get a good understanding of how far ahead of the competition Tesla truly is. Guys that argue that Waymo, et al are further ahead because they are running L5 in certain geo-fenced areas need to watch the AI Day video... they'll change their tune in a hurry.
Yeah, there's still a lot of work left to be done, but the framework is there. I gotta admit, after I first got FSD Beta, I was left wondering if we were going to have true FSD in my lifetime. After watching the vid, I'm convinced that Elon and crew will get this done.
As a guy said much earlier in this thread, we're seeing FSD at the worst it'll ever be. And when you put it in that context, it's pretty amazing at what it is capable of doing, even today.
It deals with one of those "how in the world are they going to program for that?" type scenarios.
These guys are just phenomenally brilliant.
So I've been reading around the 'net a little bit, and it is very easy to get discouraged by the current discourse; guys that have been in the beta since version 8 (which was released in October of '20) are saying that there hasn't been a lot of significant progress when you compare 10.4 to 8.0.
That might (and probably is) the case from a user (driver) standpoint, but when you look at what they have already accomplished and are accomplishing behind the scenes, the amount of "progress" is far, far beyond what is visible and noticeable from a user's standpoint.
If you'll take the time to watch the AI Day video, you'll see what I mean. Yeah, at three hours long, it's a long watch. So break it down into pieces and watch it over several days or even a week.
It is truly eye opening, and will completely change your perspective on FSD. I'm finding that I have a lot more patience with FSD when it does something silly now.
IMO, Tesla now has solutions in place for the vast majority of driving situations; it's just a matter of time until they get the neural nets and AI trained/tuned to use those solutions.
After watching the video, you'll really get a good understanding of how far ahead of the competition Tesla truly is. Guys that argue that Waymo, et al are further ahead because they are running L5 in certain geo-fenced areas need to watch the AI Day video... they'll change their tune in a hurry.
Yeah, there's still a lot of work left to be done, but the framework is there. I gotta admit, after I first got FSD Beta, I was left wondering if we were going to have true FSD in my lifetime. After watching the vid, I'm convinced that Elon and crew will get this done.
As a guy said much earlier in this thread, we're seeing FSD at the worst it'll ever be. And when you put it in that context, it's pretty amazing at what it is capable of doing, even today.
CharleyBC
Active Member
I have 36.8.5 and Sentry Live works. You do have to turn it on in the security config screen in the car. And you need Premium Connectivity.
It's not present on 36.8.5 for Model S or X models right now, just on Model 3s and Ys, much to the chagrin of S and X owners...
brunla
MS-LR / MSM / Black / Arachnid / FSD
New S/X don't even have Car Wash mode yet. And what happened to ANC? I thought that was being released as well.
rxlawdude
Active Member
As I posted earlier, I'm expecting Ohmman to do some OT housecleaning. Maybe at 0000 hours tomorrow.
rxlawdude
Active Member
But you've physically changed the condition, so it is not "doing the same thing." You are not doing the same thing when you bang on virgin vs. degraded glass (from prior attempts). Hammering a nail into a board results in the expectation that the nail head will be flush with the surface of the board (in most cases). Yet sometimes it takes several strikes on the nail to beat it into that position.
Every batter in MLB expects (dreams of) hitting a home run when contact with the ball is made. The correct expectation is EITHER (a) a swinging strike, (b) a foul tip, (c) a [fair or foul] grounder, (d) a [fair or foul] line drive, or (e) a [fair or foul] popup fly. (I'm ignoring balls and called strikes that don't involve contact.)
Expecting your batter's hit ball to land in McCovey Cove (for Giants fans) every time is a definition of insanity because it does not take other likely outcomes into consideration.
Each pitch and swing is unique, and each pitch thrown resulting in contact must simply be an expectation of one of (a)-(f) occurring. One would certainly not expect a line drive from a called strike.
Last edited:
I love this discussion although I'm not sure it's the right venue so I'll make an effort to connect to the topic of the thread here.
The action you are taking is the same despite the changed conditions of the object of your action. It's reasonable to expect different outcomes to the same action because the context in which an action is taken is always changing. Once can never really take the same action twice with exactly the same conditions. This is exactly what makes the FSD problem so difficult; making accurate predictions is incredibly difficult given the ever-changing nature of the context.
For your specific example, a batter should reasonably expect a different probability distribution of outcomes from each swing, both because each pitch is different and because his own internal models could be updated with each miss to change the success rate of his swing. It would be inaccurate for a batter to expect exactly the same distribution of outcomes in every game, or in every month, or in every season.
But more to the point, actual insanity has very little to do with making incorrect predictions about the effects of one's actions*. If you make bad predictions about the social consequences of your actions you are socially awkward, if you make bad predictions about the consequences of your motor movements you are a klutz, if you make bad predictions about the consequences of your acceleration and steering, you are just a Tesla. None of these things make you insane, which involves a more severe disfunction of your cognition.
*Although, interestingly one of the best theories of auditory hallucinations in schizophrenia does have to do with an impaired forward modeling system in the brain -- voices in our heads that we don't predict are more surprising to us and therefore attributed to external sources by the brain.
rxlawdude
Active Member
We're not talking clinically insane here, but your point is taken.
There are ideologies foreign to us that seem (to us) insanely ridiculous, without attributing frank mental illness to their actions and beliefs.
2021.36.8.5 installing now, unless of course I'm insane
Last edited:
Same here2021.36.8.5 installing now, unless of course I'm insane
Just got 36.8.5 now as well. Looks like it’s finally rolling wide!
Similar threads
