OFFICIAL BUTTON WATCH

[FalconFour]

Not bad! I used @FalconFour's "I'm probably not driving" magnet, I hope it has the same effect! hahaha

In case anyone wants it:

"I'm probably not driving" bumper MAGNET sticker/decal, Dark (OLD LISTING) | eBay

Find many great new & used options and get the best deals for "I'm probably not driving" bumper MAGNET sticker/decal, Dark (OLD LISTING) at the best online prices at eBay! Free shipping for many products!

www.ebay.com

Alternatively (in the style of @ChimpledPot above, haha), I have this

CAUTION: STUDENT ROBOTAXI bumper MAGNET sticker/decal for Tesla Autopilot FSD | eBay

Be the first to get this funny new take on the Autopilot warning stickers.

www.ebay.com

Both are available in either magnets (for 3/Y trunk lid) or stickers (for S/X/anywhere).

I feel like they at least reduce the angry reactions, but I do try and make AP behave like a reasonable human being. The worst part of AP right now is its incredibly laggy, half-asleep response to needing to accelerate... (scenario: nobody ahead or lead car is 10-spaced-car-lengths up and accelerating, set speed=70mph, currently going 53... 54... 55 at 1mph/sec). It's really hard to excuse that kind of behavior from AP

 

[acastal]

Regarding Unsafe Following. Is it possible that this is reported when on AP but not scored? I did a 22 mile drive today and had 7.9% Unsafe Following, but still got 100%.

 

[FalconFour]

Regarding Unsafe Following. Is it possible that this is reported when on AP but not scored? I did a 22 mile drive today and had 7.9% Unsafe Following, but still got 100%.

I'm constantly studying and fighting my arch nemesis, the "Unsafe Following" scoring. It's the only thing I can't figure out. So I've got some tips...

If it shows it on the screen, it "saw" it happen out of AP. I previously thought that it'd log everything, but only score what was out of AP. But that appears to not be the case.

Unsafe Following seems to be a relatively low-impact thing. So even though you got a bit of it logged, it still gives 100 because it rounds up to 100. So far, the worst I seem to do is 99, and all I get are occasional "unsafe following" blips. Trying to go for 100 every time, with some success recently.

That "Unsafe Following" only claims to be triggered (according to docs) above 50 MPH. So, it's most likely to happen as you're merging onto a freeway, I think. Probably behind some jerk that doesn't realize they're entering a freeway, so you're both leaving room for an inevitable oncoming car in the thru-lane to ram into the offender, and also not being merging-too-slow yourself... solution: enable AP before you get over 50mph on the freeway (and hope for the best merging situation)

 

[rpo]

I

I'm constantly studying and fighting my arch nemesis, the "Unsafe Following" scoring. It's the only thing I can't figure out. So I've got some tips...

If it shows it on the screen, it "saw" it happen out of AP. I previously thought that it'd log everything, but only score what was out of AP. But that appears to not be the case.

Unsafe Following seems to be a relatively low-impact thing. So even though you got a bit of it logged, it still gives 100 because it rounds up to 100. So far, the worst I seem to do is 99, and all I get are occasional "unsafe following" blips. Trying to go for 100 every time, with some success recently.

That "Unsafe Following" only claims to be triggered (according to docs) above 50 MPH. So, it's most likely to happen as you're merging onto a freeway, I think. Probably behind some jerk that doesn't realize they're entering a freeway, so you're both leaving room for an inevitable oncoming car in the thru-lane to ram into the offender, and also not being merging-too-slow yourself... solution: enable AP before you get over 50mph on the freeway (and hope for the best merging situation)

I can confirm that you can get dinged for unsafe following below 50 mph as I done multiple trips to the grocery store on surface streets without ever hitting 25 MPH AND without encountering another car on the road. How it is possible to get unsafe following in that situation is beyond me.

 

[GargantulaKon]

I can confirm that you can get dinged for unsafe following below 50 mph as I done multiple trips to the grocery store on surface streets without ever hitting 25 MPH AND without encountering another car on the road. How it is possible to get unsafe following in that situation is beyond me.

Yeah, I wonder how this is calculated because on one trip while I was mostly on autopilot, it marked me for 20% of the time at an unsafe following distance. I have my autopilot following distance set to 7 and when I drive manually, I absolutely can't stand when somebody tailgates me (don't want to be a hypocrite) so I'm really far from the person in front of me, to the point where people go pass me and then cut me off to let me know that I was driving too far for them. In other words, I have no clue why I got a yellow rating for unsafe following. It's only happened once so far.

I even think that the Tesla following distance is 7 is too close for me sometimes and I would like the setting to be able to be extended to 10. I'm assuming that Elon Musk is a tailgater, haha. He would have told the development team to increase the distance by now.

 

[nvx1977]

I can confirm that you can get dinged for unsafe following below 50 mph as I done multiple trips to the grocery store on surface streets without ever hitting 25 MPH AND without encountering another car on the road. How it is possible to get unsafe following in that situation is beyond me.

I have been driving assuming that the unsafe following applied at all speeds. It's too much cognitive load for me to track whether I'm over or under 50mph. Plus it reduces the chance I get dinged for a hard brake if someone ahead does something stupid. Anyway, safe following distance is a pretty credible safety metric, so if you actually care about driving safely, following at a safe distance at any speed should become a muscle memory thing.

 

[nvx1977]

I'm really far from the person in front of me, to the point where people go pass me and then cut me off to let me know that I was driving too far for them. In other words, I have no clue why I got a yellow rating for unsafe following. It's only happened once so far.

I even think that the Tesla following distance is 7 is too close for me sometimes and I would like the setting to be able to be extended to 10. I'm assuming that Elon Musk is a tailgater, haha. He would have told the development team to increase the distance by now.

Could you be getting dinged at the point someone cuts you off closely?

 

[BitJam]

Each person receives a score based on a statistical model, yes you are correct.
However a safety score is based on user data and the statistical model.
An individual's safety score is garbage if there are a lot of mistakes with the head on collision portion. The rest of the inputs are solid but that one is not. Garbage in garbage out. It's that simple.
If individuals are going to be tagged with specific values/labels it needs to be 100% correct. I am amazed that people do not follow that.

Thanks for the clarification. I didn't realize it was the measurements from your own car that you were questioning. But I disagree with you about the need for things to be 100% correct. This is only true for the test to be fair. What they need to do is make sure they err on the side of caution which seems to be what you're reporting. Dinging safe drivers is only an issue if they reject so many that they don't have enough beta testers. I tried to explain all this in detail in my previous post.

 

[Sporty]

I


I can confirm that you can get dinged for unsafe following below 50 mph as I done multiple trips to the grocery store on surface streets without ever hitting 25 MPH AND without encountering another car on the road. How it is possible to get unsafe following in that situation is beyond me.

Same here.

 

[GargantulaKon]

Could you be getting dinged at the point someone cuts you off closely?

Possibly that's what's going on. I've noticed it helps to be on autopilot as much as possible to avoid getting negatively scored in this category and somebody cuts you off.

 

[FalconFour]

I have been driving assuming that the unsafe following applied at all speeds. It's too much cognitive load for me to track whether I'm over or under 50mph. Plus it reduces the chance I get dinged for a hard brake if someone ahead does something stupid. Anyway, safe following distance is a pretty credible safety metric, so if you actually care about driving safely, following at a safe distance at any speed should become a muscle memory thing.

The problem is that it's not dinging for following distance, it's dinging for *ANY* distance - specifically, if someone merges in front of you and you immediately let off the pedal, whoopsie doodle, you're now spending "50%" (or some strangely arbitrary number disconnected from reality) of your time "unsafely following". I can't figure out what it's figuring for that "percent" - literally have no idea where it's getting that percent from.

You're correct in that following distance IS a credible safety metric, and there are far too many people that ride asses in traffic when they should just be passing on the left if they want to go faster... but the way Tesla detects it here is massively flawed. It should be excluding dynamic situations - merging or "cut-ins".

This is my solid case of evidence and I think it trumps anyone that claims there's no flaws with this system:
If a line of Teslas is following each other at the maximum following-distance of "7", smoothly rolling with each one following the next (say: they're set to 70 but the lead vehicle is going 65), in a right lane, and you are trying to enter that freeway by matching their speed and entering a gap perfectly in a textbook manner (say, perfectly spaced in the center of the gap, matching their speed): you would still get an "unsafe following" ding for doing so, if you weren't on Autopilot doing it.

Another scenario demonstrating exactly the same problem:

That is the scenario Tesla needs to work on: dynamic scenarios of changing traffic flow.

Someone that habitually tailgates people on the road desperately needs to be given a low safety score, but unavoidable merging scenarios shouldn't NEED to be on Autopilot to mask what shouldn't have been flagged as a fault in the first place.

 

[nvx1977]

The problem is that it's not dinging for following distance, it's dinging for *ANY* distance - specifically, if someone merges in front of you and you immediately let off the pedal, whoopsie doodle, you're now spending "50%" (or some strangely arbitrary number disconnected from reality) of your time "unsafely following". I can't figure out what it's figuring for that "percent" - literally have no idea where it's getting that percent from.

You're correct in that following distance IS a credible safety metric, and there are far too many people that ride asses in traffic when they should just be passing on the left if they want to go faster... but the way Tesla detects it here is massively flawed. It should be excluding dynamic situations - merging or "cut-ins".

This is my solid case of evidence and I think it trumps anyone that claims there's no flaws with this system:
If a line of Teslas is following each other at the maximum following-distance of "7", smoothly rolling with each one following the next (say: they're set to 70 but the lead vehicle is going 65), in a right lane, and you are trying to enter that freeway by matching their speed and entering a gap perfectly in a textbook manner (say, perfectly spaced in the center of the gap, matching their speed): you would still get an "unsafe following" ding for doing so, if you weren't on Autopilot doing it.

Another scenario demonstrating exactly the same problem:

View attachment 715955

That is the scenario Tesla needs to work on: dynamic scenarios of changing traffic flow.

Someone that habitually tailgates people on the road desperately needs to be given a low safety score, but unavoidable merging scenarios shouldn't NEED to be on Autopilot to mask what shouldn't have been flagged as a fault in the first place.

yeah it seems like there needs to be a grace period for you to adjust to someone cutting in front of you too closely.

 

[FalconFour]

All this is because we're all gunning for a 100 overall score to get FSD Beta earlier than anyone else

We'd all do perfectly well if we didn't have to get a perfect 100 to "win the contest" in the first days of FSD Beta onboardings If it were just a matter of staying over 90, say... then hey, this would be pure cake.

 

[GeoffEV]

A statistical model will be right on average but not right in every case. There are probably some cases where it's way off. The model not working in your case does not imply the 5 billion miles of fleet data they used to build the model are invalid.

People seem to be under the false impression that the safety test is supposed to be fair. It's not about being fair, it's about weeding out bad apples even if that means weeding out some good drivers. As long as Tesla gets enough beta testers, weeding out unlucky good drivers is not an issue for them.

Neural nets like the ones Tesla uses for FSD need a massive amount of input data for training. This is why Tesla is building the world's largest AI computer. It's also why they hired a team of 1,000 people to handle things computers have trouble with. OTOH the NTSB has already blamed Tesla for accidents caused by inattentive drivers and said in no uncertain terms that Tesla should not give the beta to more drivers until the safety issues have been worked out.

Tesla needs a large pool of beta testers to get the data they need to improve FSD. But they can't let the beta process appear unsafe. I thought their safety test was a brilliant way out of this Catch-22. It allows them to safely (I hope) expand the pool of beta testers so they can rapidly improve FSD. If the accident rate is not vastly lower than the overall average then the government will likely shut down Tesla's FSD program which would be disastrous on many levels. Ideally, as Tesla expands the pool of beta testers FSD will get better and safer thus making it safe for them to further expand the pool. This is mind blowing!

For the safety test it appears Tesla did a large linear regression (I imagine they inverted a big matrix using SVD) to find out which measurables correlated most strongly with collisions. Then to make things easier on humans they reduced it to just the five variables that had the highest correlations. This seems like an excellent approach, especially for their first beta. I can't think of a better way. But perhaps you can. If so, please chime up because Elon is often open to good suggestions.

From your title, I thought you were gonna bash the Safety Algorithm - so my bad knee jerk made me skip it. I left the thread to get some work done and luckily it landed back on yours as I reopened TMC. Yours is the best information I've found on "why" this Safety Test is what it is - and great insight on Tesla's strategy to move FSD forward is very encouraging. Absolutely brilliant!

 

[BitJam]

I can confirm that you can get dinged for unsafe following below 50 mph as I done multiple trips to the grocery store on surface streets without ever hitting 25 MPH AND without encountering another car on the road. How it is possible to get unsafe following in that situation is beyond me.

Interesting! My guess is the safety test was designed and/or tested by people working on and using 10.x. If 10.x handles/reports these situations better then this might explain why Tesla says you aren't dinged when using AP yet some people are finding that they are.

[...] If it were just a matter of staying over 90, say... then hey, this would be pure cake.

H'mm. I doubt driving over 90 MPH would help with your score.

 

[TIppy]

I can't figure out what it's figuring for that "percent" - literally have no idea where it's getting that percent from.

Once you are closer than 3 seconds a timer starts. Then once you get closer than 1 second it starts timing how long you are there. The percent is the ratio of the time you are closer than 1 second over the time you are closer that 3 seconds times 100, of course. It probably accumulates all occurrences of the less than 3 second and less than 1 second times for the entire trip for when your speed is greater than 50 mph.

Since it's a ratio of these two times, the anomalies people are seeing may be because they never encounter another car, never get closer than 3 seconds or never reach a speed greater than 50 mph. If the two times aren't initialized properly, their ratio at the end of the trip may be non-zero.

 

[dckiwi]

Maybe we should close this thread. The button is here. Just sayin’

 

[FalconFour]

The percent is the ratio of the time you are closer than 1 second over the time you are closer that 3 seconds times 100, of course.

That's something that was written in the docs as well, but it looked like some kind of weird doc typo. It makes about as much sentence-flow logical sense as the phrase "the ratio of the time in 1 second compared to the time in 3 seconds times the square root of six", which makes perfect sense if two measurements of the same period of time are different I don't at all understand what it's referring to.

Maybe we should close this thread. The button is here. Just sayin’

Hesitantly agree, but this is a fun thread Stuff that happened more than 2 pages ago is just totally lost to the sands of time...

 

[Atari2600]

Thanks for the clarification. I didn't realize it was the measurements from your own car that you were questioning. But I disagree with you about the need for things to be 100% correct. This is only true for the test to be fair. What they need to do is make sure they err on the side of caution which seems to be what you're reporting. Dinging safe drivers is only an issue if they reject so many that they don't have enough beta testers. I tried to explain all this in detail in my previous post.

I am under the impression this will be used by Tesla insurance in the future or possibly is used now. Sure, not getting into a beta is one thing but I assume this score is sticking around with us folks forever. If this data only exists for this beta and is gone later then it's not a very big deal.
I raised my score from a 3 to a 96 but I had to take some weird actions to do so like stop driving on my normal routes due to the false collisions.

 

[TIppy]

That's something that was written in the docs as well, but it looked like some kind of weird doc typo. It makes about as much sentence-flow logical sense as the phrase "the ratio of the time in 1 second compared to the time in 3 seconds times the square root of six", which makes perfect sense if two measurements of the same period of time are different I don't at all understand what it's referring to.

Well, if you notice when the car ahead of you passes a sign and you start counting seconds until you pass the sign then you know how many seconds you are behind it. If it's more than 3 seconds your car ignores it. If you counted the time and it was 3 seconds or less between you and the car ahead, your car starts a timer to see how long this remains true. Let's say this is 10 minutes. Now let's say that during this same 10 minutes you continue the process of counting as you pass landmarks, and sometimes you are only 1 second behind the car ahead of you. Let's say this happens for a total of 2 minutes. Keep in mind that it's still less than 3 seconds so that time continues to increase for the full 10 minutes. So, you were within 1 second of the car ahead of you for 20 percent of the 10 minutes, 2 minutes out of 10.

Every time you get within 3 seconds of a car ahead of you this process is repeated.