RayK
Active Member
I hope that's a very close friend of yours. Might you have meant, pucker?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Elon says: Fleet feature usage tremendoulsy helps. Your thougts.
- Thread starter Garlan Garner
- Start date
Knightshade
Well-Known Member
I don't believe that is entirely the case- especially going forward...
The 2.5 computer is incapable of processing all 8 cameras at full frame rate.... from greentheonly we saw it's sometimes only showing every 4th frame for example on some of the cameras. effectively running each 36 fps camera at 9fps....while others run full speed.... (this makes some sense at at MAX HW2.5 can only process 200 total frames per second... and all 8 cameras at full frame rate would be 282 fps (7x36 and the backup cam at 30). So it has to reduce framerate on at least some cameras due to hardware limits.
This isn't a big deal today because they can still tell either car hardware "Send me pictures of stop signs" for a single camera and unless you're going REALLY fast 9 fps is good enough.
And since that's the way the "old" NN data was being labeled and used for training- that worked ok.
But the fundamental re-write is based on stitching all 8 cameras together, full frame rate full time all the time, to create a 360 degree real-time picture and understanding.
Which HW2.5 can't do.
So for example if they want to collect a bunch of intersection data to really teach the car how to handle "which light is relevant to which lane" or "what happens at 4 way stops" in situations where more than 1 camera worth of frame is required- HW2.5 can't get them that data in the way they need it.
That said, the % of AP2+ cars that still "only" have HW2.x is going to become an increasingly small % of the fleet over time- I don't suspect there'd be enough value in upgrading them without getting FSD money to be worth bothering with data wise.
Last edited:
You are explaining how transactional relationships work. Some companies prefer to build higher loyalty by making things that doesn’t make sense at first glance (for instance, upgrading 2.5 -> 3.0), but in reality can be beneficial in the long run. Such actions can change consumers’ perception of the company and compound over time. This is the only way to build a really successful business.
So wouldn't it be safe to say that the percentage of folks who:
1. Didn't order FSD
2. Don't have HW3
Won't be used for mapping?
I believe there are more people "right now" that have FSD and HW3.
destructure00
Active Member
Easy workaround, just keep your foot on the accelerator. You can go whatever speed you want, lane keeping and stop light detection continue to work, just lift off when approaching an intersection you actually want to stop at. Also don't have to confirm green lights this way.
Knightshade
Well-Known Member
I seriously doubt there's more people with FSD than without it.... I've never seen any serious take-rate estimate anywhere near 50% or above it.
I think the best/highest estimates have been in the 30-something percent range?
But they don't need you to have FSD for passive data collection.
As to HW2 vs HW3 cars-
Tough to say EXACTLY but we can estimate.
From end of Q3 2016 to end of Q1 2019 Tesla produced about 455,000 cars.
This is the very rough ballpark # of HW2.x cars built. Since the cutover HW dates aren't exactly at quarter break (and they made some 2.x cars after the official cut in date) this number is a decent guess but might be off a bit.
Since end of Q1 2019 Tesla had produced about 380,000 cars.
So if we assume at least 1/4 of HW2.x cars have been upgraded (which might be high- hard to know for sure)
Then we're roughly just a little bit more HW3 cars than HW2 cars out there today.
It might be a bit the other way still.
Obviously every quarter going forward though the HW3 percentage goes up.
End of Q2 2020 it should at least be a solid majority without any guessing needed.
(above also entirely removes any pre-HW2 cars from the totals for obvious reasons)
Knightshade
Well-Known Member
That's not really an "easy workaround"
That's "making TACC essentially worthless"
You'll also need to lift off anytime there's traffic in front of you too of course- since the car won't normally slow on its own for ANYTHING if you're on the accelerator. (if you forget this don't worry- emergency braking will remind you!)
I agree with you, but.....
I'm thinking differently.
The Data Collection I'm talking about concerns folks that are actually using the features of 2020.12.6. Feature testing not mapping.
For example:
How many people disengage the features of 2020.12.6 when approaching a light?
How many people disengage the features of 2020.12.6 when the wipers turn on?
How many people disengage the features of 2020.12.6 at certain speeds?
How many people won't engage the features of 2020.12.6 at night?
How many people use the features of 2020.12.6 in rural vs. urban areas?
Currently that's FSD + HW3 only folks right?
That automatically eliminates anyone without FSD right?
That automatically eliminates anyone without HW3 right?
Knightshade
Well-Known Member
Only FSD+HW3 PLUS explicitly turned on this new specific feature in settings too.
Yup- plus everyone who has both but chooses to leave it off for whatever reason.
KetchupKilla
Member
I'm not a big fan of using he new update. I hate how it ends up using the brakes to stop. Ive noticed it starts slowing down slowly at first, then speeds up and uses brakes.
2018 Model 3, AWD, FSD, HW3, 12.6
TACC - does allow me to go above the speed limit listed on screen, by scrolling the wheel, I've gone 10mph above without issue and it continued to work for stop signs and street lights.
Autopilot - will not let me go above the speed limit listed on screen, you are locked into what ever the speed limit is on the screen even if you had set +5 in the relative screen setting.
If in either one above when it starts to slow down and you press accelerator it continues on without issue. HOWEVER... if you press accelerator too soon and light turns yellow way before you get there, you have to manually brake. I got a little ambitious and hit accelerator a bit early, light turned yellow just after I pressed accelerator and I knew it would be a red going into intersection so I had to manually brake. By pressing on the accelerator too early I advised car it was safe to go, so it canceled stopping and continued on. Opps.
Stop signs and creep, if it stops way short of stop sign (it does sometimes) and if you are in either of the above and you press accelerator it will resume the set speed, either what you had in TACC or what autopilot says. There is no creeping up to get better view at intersection, you have to exit it by brake or stalk and then creep up. After clearing intersection reengage TACC or Autopilot and continue on your journey.
It's not fun in neighborhoods because of the stop sign distance difference, plus it can pick up the higher speed of an adjacent road and when you press accelerator telling it to resume, you better be ready for a jolt if you didn't notice the speed change on the screen.
So I think it's two way learning, what it records us doing and how we adjust to it's nuances.
TACC - does allow me to go above the speed limit listed on screen, by scrolling the wheel, I've gone 10mph above without issue and it continued to work for stop signs and street lights.
Autopilot - will not let me go above the speed limit listed on screen, you are locked into what ever the speed limit is on the screen even if you had set +5 in the relative screen setting.
If in either one above when it starts to slow down and you press accelerator it continues on without issue. HOWEVER... if you press accelerator too soon and light turns yellow way before you get there, you have to manually brake. I got a little ambitious and hit accelerator a bit early, light turned yellow just after I pressed accelerator and I knew it would be a red going into intersection so I had to manually brake. By pressing on the accelerator too early I advised car it was safe to go, so it canceled stopping and continued on. Opps.
Stop signs and creep, if it stops way short of stop sign (it does sometimes) and if you are in either of the above and you press accelerator it will resume the set speed, either what you had in TACC or what autopilot says. There is no creeping up to get better view at intersection, you have to exit it by brake or stalk and then creep up. After clearing intersection reengage TACC or Autopilot and continue on your journey.
It's not fun in neighborhoods because of the stop sign distance difference, plus it can pick up the higher speed of an adjacent road and when you press accelerator telling it to resume, you better be ready for a jolt if you didn't notice the speed change on the screen.
So I think it's two way learning, what it records us doing and how we adjust to it's nuances.
Has anyone seen Tesla's report on the FSD take rate?
If there was one, I missed it.
I would have guessed 10-15% rate, which is my unofficial sampling result from folks at work who have TM3's.
Exactly.
And the % of people who enable the "brake when traffic light is detected" feature is likely to be small, and exponentially decaying daily.
I get the value from collecting this date for Tesla. It makes sense for the company.
But if there ever was a nuisance feature for the driver, that detracts from AP driving experience, this is it.
My hat is off for folks who are willing to put this up for the benefit of Tesla, for free.
Thank you, guinea pigs!
Looks like there is more to come concerning FSD.
Elon said this in the earnings call.
“Regarding Autopilot, we released a new software update for traffic lights and stop signs to early access users in March and to all US customers with full self-driving package just last week. Our cars will now automatically stop at each stop sign or traffic light until the driver gets a confirmation to proceed. I should say that the car is actually capable of much more than this, but we are only exposing functionality that we feel quite good about and where we feel it is probably a safety improvement,” Musk said.
Elon said this in the earnings call.
“Regarding Autopilot, we released a new software update for traffic lights and stop signs to early access users in March and to all US customers with full self-driving package just last week. Our cars will now automatically stop at each stop sign or traffic light until the driver gets a confirmation to proceed. I should say that the car is actually capable of much more than this, but we are only exposing functionality that we feel quite good about and where we feel it is probably a safety improvement,” Musk said.
Knightshade
Well-Known Member
...you realize he's saying "There's other stuff the car can do- but it'd be MORE dangerous than not having it if we released it now" right?
That's not a promise- it's a threat
Troy has survey data from a decent # of buyers-
Over 3000 on his model 3 sheet covering basically Jan 2019 to today.
Teslike Model 3 Order Tracker #3
Take rate is actually up to 40.3% (the last time I looked at these numbers was before EAP went away, so makes sense they'd be higher now for FSD since that's where most of the EAP content ended up)
(EAP didn't go away till Feb/March 2019 which is why there's still a tiny # of that shown in the survey tab)
So still a minority of buyers but a decent bit higher than when EAP was available.
... you mean what’s in the pop up when you enable the feature? Because it pretty much says exactly this.
to quote: “This feature will be conserative, slowdown often at first, and will not attempt to turn through intersections. Over time, as we learn from the fleet, the feature will control more naturally.”
...you realize he's saying "There's other stuff the car can do- but it'd be MORE dangerous than not having it if we released it now" right?
That's not a promise- it's a threat
That might be true....
We will have to just wait and see.
I don't mind living on the edge.
I don’t disagree at all. I was replying to the poster who thought HW3 would improve the data collection for Tesla. We both know it doesn’t, because HW2.5 can collect the same data that HW3 can for training purposes. It just cannot process it on board. Even today, HW2.5 can detect the traffic lights and could allow a driver to tap the stalk to mark the data as an approach to an intersection and the conditions allow the car to proceed. It apparently just cannot do more than that sufficiently for Tesla to think it should allow HW2.5 to support full traffic light behavior. None of us know the details of that decision, though I suspect your point about processing is the main reason.
With the tapping of the stalk during approach to an intersection, the only thing the driver is doing is flagging that data as relevant to training the neural net for that condition. The training is done centrally, and not in the car. The ability to apply that “knowledge” in real time is a whole different thing and obviously needs HW3.
