Tesla Software updates - Australia

[Vostok]

Large chunks of the Hume are not a controlled access motorway... As well as proper on/off ramps there are people's driveways, and lots of intersections with small regional roads. It is more like an 'A' road.

Oh don’t get me started on Victoria calling the Hume Highway the “Hume Freeway”... because it’s not. That label is factually and utterly wrong. A road with at-grade intersections, driveways and other non-controlled access points is, by definition, not a freeway or motorway. Simply being a divided road is not enough to earn the title.

In NSW, the correct definition is strictly applied with naming and road signage. As soon as a divided road with controlled access becomes a divided road with at-grade access, the “End Freeway” sign appears.

 

[TheArtieChoke]

Has anyone without FSD received an update to 20.12 or 20.13? I thought I may have been getting an update as the car was downloading something but it has now pulled 11gb and nothing over two days, no idea what it is doing? Download speed is odd as well, it’s always around 2Mbps, like it’s trickle downloading

 

[TheArtieChoke]

Just an update on above, I decided to try a steering wheel reset which seems to have fixed the issue, downloading 20.12 at the moment so must have been some bug which the fetch for two nights

 

[blasterkings]

Has anyone without FSD received an update to 20.12 or 20.13?

Yep 20.12 on saturday.

 

[Nuclear Fusion]

Switching off traffic control - it’s way too buggy

 

[Hairyman]

Looks like 20.13 has had pretty limited roll out so far. Wonder if we are the test case for this iteration?

 

[Wol747]

I got it last week.

 

[paulp]

This is just wrong. The car wants you to label every prediction of a green light. If it’s a prediction of a red light the only way to over ride is to accelerate through it. This way they can collect high quality labeled edge cases to help train future iterations of TLASSC

The training thing is marketting talk. Every day I come to a fork in Port Road. Every day I have to disenggage and branch left. It always wants to go right. 4 years of training and nothings been learnt. I’m sure you’ll disagree and hurl another insult, but I’m keen to know if anyone out there has managed to see their car change behaviour through training?

 

[ARMARM]

4 years of training and nothings been learnt.

You do know that the cars don't train or learn themselves? The Neural Net is trained by someone reviewing footage from an AP disconnect or an edge case etc., and then making a change in the software to tell the car how to behave.

 

[paulp]

You do know that the cars don't train or learn themselves? The Neural Net is trained by someone reviewing footage from an AP disconnect or an edge case etc., and then making a change in the software to tell the car how to behave.

Oh ok so the system is nothing more than when a user effectively reports a problem (via a disconnect or intervention), a programmer reprogrammes it when the problem reaches the start of the que?

 

[DOMn8]

Oh ok so the system is nothing more than when a user effectively reports a problem (via a disconnect or intervention), a programmer reprogrammes it when the problem reaches the start of the que?

You'd think that your 4 years of disengaging every day would have reached the start of the queue by now, wouldn't you?

 

[TheArtieChoke]

You'd think that your 4 years of disengaging every day would have reached the start of the queue by now, wouldn't you?

What about the people that want to go right? There isn't a different Autopilot for each individual car, it's all the same comands.
FSD mentions that auto driving on city streets is coming soon, meaning it will take on Navigate on Autopilot type of behaviour where if you are going left, it will take the left lane. Right now it is still "dumb" in the sense it will pick whichever lane comes first in the algorithm, probably stick to the middle lane if a branch comes up.

I wouldn't say taking the wrong lane is a problem that the programmers would have looked at. It is more like car doesn't brake here and I had to slam on brakes type of disengagements that would be looked at and modified.

 

[ICUDoc]

I haven seen my car get better from my own inputs, but i have DEFINITELY seen Autopilot get MUCH better over the first few days of a few versions- spectacular changes when it first was released, and quite marked changes a couple of years ago when the first post-Mobileye version was revised.
Clearest example was an exit off a freeway where - everyday- it tried to follow cars up a second branch road then jerked back into the correct lane. Hasn't done that for a year or more. Smoooothly up the road, these days.

 

[GrimRe]

The training thing is marketting talk. Every day I come to a fork in Port Road. Every day I have to disenggage and branch left. It always wants to go right. 4 years of training and nothings been learnt. I’m sure you’ll disagree and hurl another insult, but I’m keen to know if anyone out there has managed to see their car change behaviour through training?

Just because you don't understand something doesn't make it "marketing talk". Talking about your own anecdotal experience as some sort of proof point (because its the only thing you know) is a typical boomer move.

Truth is you have no idea what "training" is the sense of deep learning convolutional neural networks. The confirmation is to label examples of where the classifier disagreed with the user. With this collection they can review oddities in the way stop sign and traffic lights appear and re-train using a tweaked set of training settings. Your parallel to not being able to navigate a fork in the road is akin to those who say climate change doesn't exist because int 1908 their great aunt wrote a letter about how a day was hotter than today.

I don't typically "hurl insults" unless its directed at somebody flaunting ignorance as some sort of enlightenment.

 

[OneSpeed]

Arguing that anything is 'a typical boomer move' is hurling an insult.

 

[paulp]

Just because you don't understand something doesn't make it "marketing talk". Talking about your own anecdotal experience as some sort of proof point (because its the only thing you know) is a typical boomer move.

Truth is you have no idea what "training" is the sense of deep learning convolutional neural networks. The confirmation is to label examples of where the classifier disagreed with the user. With this collection they can review oddities in the way stop sign and traffic lights appear and re-train using a tweaked set of training settings. Your parallel to not being able to navigate a fork in the road is akin to those who say climate change doesn't exist because int 1908 their great aunt wrote a letter about how a day was hotter than today.

I don't typically "hurl insults" unless its directed at somebody flaunting ignorance as some sort of enlightenment.

So anyone without your extraordinary knowledge or doesnt meet with your narrow view is a ‘boomer’
There’s a little trick to success in life. One day you might learn it. It starts with knowing how to have a rational conversation without making yourself feel superior.

 

[paulp]

Arguing that anything is 'a typical boomer move' is hurling an insult.

Its all he knows. Its what those with an inferiority complex do. Makes them feel better.

 

[Wol747]

Back to the subject <g> :

We're all missing the point. I would bet that the present road to FSD (in it's full sense - driverless) is not going to work. AI as we know it now is very sophisticated but it doesn't "know" what's happening. It reacts to events by complex instructions, it "learns" from the events but never actually understands anything. Until it does it will not be truly autonomous, able to go anywhere at any time.

 

[Hairyman]

It is an interesting thought. I think that the AI will need to predict behaviour of other road users from fairly subtle cues rather than just respond to the immediately observable events.

I suppose once the density of AI driven reaches a sufficient threshold, especially if there is vehicle to vehicle communication then that issue will disappear.

Things will be some time away. My main interest is to have some increasingly vigilant copilot with the ability to intervene as needed to help keep me out of trouble.

There are some interesting papers on trust in AI beginning to be written.

 

[TaslaSR+]

Personally I think the subject should change to us just being excited, waiting for the next release to come out and hoping for things like:
a) FSD would autopark my car in parking spots without adjacent cars
b) The entertainment system gets heaps better and reliable - comparable to Carplay/Android Auto - supporting multiple drivers
c) We get an option for changing the scary phone ring tone and steering wheel controls for answering/hanging up!

And then being happy/sad at the next release giving us so much (compared to other car companies), yet so little (compared to what we're use to with big iPad looking screens).

FYI - Another good option when talking about AI/ML is to watch The Good Place and how they handle the trolley problem in the 2nd series (which makes me feel we'll always need the option to intervene). I'm just glad I program software that has little chance of killing people.