Wiki FSD Beta 10.4

[TravelFree]

Tell me you don't understand what a neural network is without saying you don't understand what a neural network is.

This insult session is off topic and ends here.

The ANN is a computer program that tries to replicate the way most experts believe in theory how the human brain thinks.

 

[GranularHail]

... a couple of more-forward side cameras would go a long way here, but even when there was no visibility issue and clearly no other cars it was slower to make a turn than I'd have been- might have to try the aggressive setting next time.

(quoting you not to argue with you, but you sparked a thought I've had several times from numerous other posts relating to forward side cameras)

I think the "we need more-forward side cameras" concept is popular mostly for the wrong reason. I've seen lots of times where the car crept... and crept... and crept before finally deciding to commit to a turn. Yes, some of those involved limited visibility to the side. However, in my personal experience, the vast majority were situations where there was no visual obstruction at all, so an a view from an additional angle wouldn't have helped.

However, one of the points made on AI day was how much better the perception can be when you merge the data from multiple cameras into a singular vector space. I specifically recall visuals from around 20:47 into this video, where the difference in quality is super obvious:

AI Day Single Cam vs. Multi Cam

The glaring problem I see is that if fusing data from overlapping fields of view is required to generate that high quality perception, what happens to the large areas to the sides of the car and narrow area directly behind the car where only one camera can see that area?

Note the large areas left and right of the car where only the light green "Wide Forward Camera" view can see:

Screenshot taken from Autopilot

These areas can't benefit from fusing data from other cameras to improve their accuracy. Sure for fixed objects like curbs, you can remember where you saw something in the other cameras as you approached, but when detecting moving objects like crossing cars, you have to rely on live data you see right now, and there is no other data source that can help you see it.

So, I suspect FSD's perception in those left and right areas is no better than the low-quality "before" images shown in Tesla AI day, and without additional sensor, can never get as good as the "after fusing into vector space" images.

 

[jebinc]

He/she wants to get the latest/regular/standard firmware (example 40.x) updates for the car.

Yes, he wants to trade beta FSD for ARNR for a while. More useful to me right now.

 

[tmoz]

Tell me you don't understand what a neural network is without saying you don't understand what a neural network is.




See above.

NNs are often called black boxes for this exact reason.




Could be either.

Just like "a human wrote the law about stop signs that a human driver may or may not follow depending on other circumstances"

Now, you certainly COULD hard code the computer to NEVER run a stop sign no matter what- that's how the original system worked... but as more decision making moves to NNs that will be less and less the case over time.




How does the programmer know when it can see left and right well enough and when it can't? Specifically?

How does it know there's a bush in the way or not? Specifically?

What is the hard code that tells it that?

(Spoiler- there isn't any- perception is done with neural nets- not hard coding)




FWIW I agree it needs further-forward side/facing cams for this purpose.

But that has nothing whatsoever to do with hard code vs NNs, it has to do with the physics of light and vision.






So funny story to again drive this point home.

Computers designed to play Go, for years and years, using traditional hard code- couldn't beat anybody but low-mid grade tournament players.

Once they switched away from hard code to neural networks they soon began beating world champions.

And much like Chess (which has also moved to using NNs at the top levels of play) the NNs often produce moves unusual and surprising to humans, but which consistently beat said humans.





Again- I agree we need both at least 2 more cams as mentioned, and more computer (Is HW4 enough? I dunno- neither do you, neither frankly does Elon- we won't know it's enough until it's enough)

But the regulator thing is entirely a red herring.

It's already legal to operate an L4 or L5 system in a number of US states. Today. With no additional permission/certification needed from regulators.




That's a question for philosophers.

But what it's definitely NOT, even today, is just hard code running fast.

Artificial AI is a poor choice of words. There is carbon based I and more recently silicon based I - so CI and SI. Right now SI is a new born baby.

 

[Knightshade]

This insult session is off topic and ends here.

It's not an insult- it's pointing out your lack of information is causing you to conflate hard coded programs with an entirely different concept you don't appear to understand.

The ANN is a computer program that tries to replicate the way most experts believe in theory how the human brain thinks.

I'd humbly suggest you do a bit more research on why you're wrong about this.

 

[Ramphex]

Right. I first started with the email, but figured that would take forever, so a seek a faster method to undo this. No change to firmware yet… tick tick tick…. Maybe I’ll be unenrolled in “two weeks,” or “soon”…

Why do you want out? I’ve been away for a couple of days so I might’ve missed it

 

[jebinc]

Why do you want out? I’ve been away for a couple of days so I might’ve missed it

@Ramphex Hey, bro! Reasons include, but are not limited to, the following;

1) Change/improvement has been so small for me the last three releases;
2) Would enjoy production firmware (2021.40.5.1) features more then beta FSD right now, such as active road noise reduction ("ARNR")- more useful (to me, anyways)
3) 10.x series releases are hard on my Plaid's steering and braking components.... And I just had my half shafts replaced due to that vibration issue
4) Large efficiency hit using beta FSD, when combined with the cold and rain here
5) I can always opt back in, at any time, per the reply I received from Tesla betafsd team (see below)

"Hello Xxxx,

FSD Beta 10.4 will be removed from your vehicle shortly. To ensure you do not receive future Beta updates, please go to the Autopilot UI and select Controls > Autopilot > Request Full Self-Driving Beta, which will trigger a pop-up which will then present you with the Opt-Out button. You can opt back into the FSD Beta queue at any time by selecting the Request Full Self-Driving button.

-FSD Team"

@WilliamG

 

[sdmack]

5) I can always opt back in, at any time, per the reply I received from Tesla betafsd team (see below)

As that email is written, you can opt back into the "FSD Beta queue" which probably means driving your Plaid in granny mode for a week or so to requalify....

 

[jebinc]

As that email is written, you can opt back into the "FSD Beta queue" which probably means driving your Plaid in granny mode for a week or so to requalify....

@sdmack I believe it means, worse case, the same as a person who resets their safety score - 100 miles and a safety score of 99 or 100 (Current requirement as of today).

 

[miracj]

I don't know the exact details of my state laws, but when turning left into a road with a middle turn lane, I turn into the middle turn lane and then merge, never stopping, unless there is another vehicle in the middle turn lane that I might hit. If the middle turn lane was clear, I might use it to speed up to affect the merge.

 

[ACPark]

In Alberta, what I believe is being referred to here are 2-way left turn lanes. They are not intended for turning into, only out of across traffic as a safe position to stop.

 

[greenmore]

It only works on highway driving but not in city or off highway turns. The whole idea of FSD beta on city streets is it makes turns automatically. If you don't want to test FSD beta on city streets then just disengage FSD beta / including AP while driving off the highway.

I suggest you set up a separate profile- call it "FSD-Beta" then select the appropriate menu settings. Go into your own profile and disable FSD-beta.

Turns and lane changes are different. Here we were mostly discussing lane changes for getting into faster lanes. Those could be made optional (with confirmation), like in NoA. Although the car often goes into the wrong lane even in preparation for turns too, which could also be improved by human confirmation.

What is very useful is that FSD also allows the driver to request lane changes by taping the turn signal, similarly to NoA. Without that, just letting the car choose its own lane, in the city it would be almost constantly in the wrong lane on multi lane roads.
However, unfortunately FSD often fights my lane change request (wants to go back to its own suggested lane which is usually wrong), and sometimes just simply ignores them.

 

[SilverString]

I don't know the exact details of my state laws, but when turning left into a road with a middle turn lane, I turn into the middle turn lane and then merge, never stopping, unless there is another vehicle in the middle turn lane that I might hit. If the middle turn lane was clear, I might use it to speed up to affect the merge.

Middle turn lanes are almost all of my main highway routes and neither 10.3.1 or 10.4 handles them correctly, often it is scary.

 

[greenmore]

I'm sure many people mentioned it here, but I don't have time to read all the comments, is that FSD only works when there is good GPS lock on the car.

I park in a garage and overnight the GPS location often drifts away by up to a couple hundred feet, and when I drive out of the garage the car thinks I'm a half block away from my real location. If I start using FSD when the GPS is inaccurate, it's goes nuts and wants to go the wrong way all the time. I guess it can't use vision to correct for the drift and can't deal with the mismatch between what the car sees and where it thinks it is on the map.

The only way around this is to wait until the GPS gets an accurate lock, which sometimes takes several minutes. This morning it was worse than usual, I had to drive about 3-4 miles before I got a good GPS location and I could enable FSD.

 

[miracj]

In Alberta, what I believe is being referred to here are 2-way left turn lanes. They are not intended for turning into, only out of across traffic as a safe position to stop.

View attachment 732661

That covers the turning left from the middle turn lane. It doesn't mention turning into the middle turn lane. Maybe that's not legal and you are supposed to turn into the "fast lane" of the road. However, in practice, if there is a lot of traffic, you use the middle turn lane as an onramp to the fast lane.

 

[JulienW]

....Refers to pedestrians and bicyclists who enter moving traffic on the roads placing themselves at risk of being mowed down as "road kill."....

Streets are NOT just for cars and driving a car doesn't make the humans that are on foot or bikes some type dumb animals that are "...placing themselves at risk of being mowed down..." Wrap a human in a 2 ton powered metal container and suddenly they are superior beings with entitled rights that supersede anyone NOT in a car. This American attitude needs to change.

Now the CORRECT definition: A Vulnerable Road User is a person using the road that is a pedestrian, biker or scooter or not in a car or motorcycle.

 

[EVNow]

Just show me how AI is not just a very fast very robust calculator that is silicone hardware using software to decide what to do.

Obviously FSD computer is just silicon that executes instructions very fast - same way the brain propagates electrical impulses very fast.

But software has come a long way since the days of FORTRAN to implement certain well known calculations and algorithms. The point of NN is that you can teach it how to do stuff without explicitly telling how - just by examples / training data. This is so that we can solve problems like image recognition that could not be solved using traditional algorithm based approach.

 

[jebinc]

Obviously FSD computer is just silicon that executed instructions very fast - same was the brain propagates electrical impulses very fast.

But software has come a long way since the days of FORTRAN to implement certain well known calculations and algorithms. The point of NN is that you can teach it how to do stuff without explicitly telling how - just by examples / training data. This is so that we can solve problems like image recognition that could not be solved using traditional algorithm based approach.

Ah, FORTRAN 77.... Fond memories.....

 

[EVNow]

I had to drive about 3-4 miles before I got a good GPS location and I could enable FSD.

Is there a way to tell when the lock is good or is it just by looking at the map and making sure it shows the car in the right place.

 

[JulienW]

Is there a way to tell when the lock is good or is it just by looking at the map and making sure it shows the car in the right place.

Also the UI is displaying Google Maps but the car is reading and using Open Street Maps "under the hood". So there is a possibility that what is shown in the UI is just incorrect.