Dom, I noticed FSD seemed to hesitate and/or behave questionably almost entirely when you were on shaded/highly shadowed roads. It seems that especially those roads where there was a significant difference in light/dark caused “vision” problems. I noticed hesitation several times on the streets with high contrast light/shadow patterns. Serval times it also had to correct slight mistakes on the heavily tree-lined streets. Maybe you should try the same route on a cloudy bright day with less contrast between light/shadow. Record that test as well and compare side by side. That would make a great video: split screen showing the two videos with you voice over narrating what you observe on the videos and your experiences driving and differences observed. Could be quite interesting. 🤔
To map out the whole route... Enter a destination, then Edit Trip to add another destination, until you got them all in there. You'll have to make sure the order is right, if not, drag them around. As you're driving it will just stop at the first destination and not automatically continue. However, if you tap Edit Trip as you're coming up to the destination, you can easily tap to remove the top destination, and it will automatically continue to the next. (You may also be able to add all the stops via the app and tap Send To Car.) I did my first FSD video a while ago and figured out the hard way how to make it auto advance to the next destination. ua-cam.com/video/lpdoWfnClDs/v-deo.html
Thanks for this. I've tried using edit trip to set up the whole route, but it would navigate first to the last point instead of the first, which was super annoying and unexpected. More recently, the various way points were all near the top of my navi list, so quick and easy to find and tap on, but I've done a lot of road trips since my last video. I like this app suggestion. I'll try that and the edit trip thing again.
It’s trained in two ways, carrot and stick. Cars not using FSD submit videos that are filtered for “good driving” behaviors are the carrots that teach it how to behave “properly.” Then, people using FSD provide the sticks when it misbehaves. Blinker input, accelerator input, and takeovers teach it how NOT to behave. Another factor is “shadow mode” running on cars not using it. If the driver does something different than FSD would have, that’s potentially useful data too, given that the driver’s behavior is considered good, or even more importantly, better behavior than FSD would have done. Both sides are important, and the behavior filtering is the key to the whole “virtuous loop” we’re entering with v12. Improvements should increase exponentially going forward, IMHO.
Since you asked: I believe it’s easiest to understand how Tesla is likely training FSD v12 in the classic idea of carrot and stick. Cars not using FSD submit videos that are filtered for “good driving” behaviors. These videos, driver controls input, vehicle kinematics (including GPS), and annotated route data are fed into the “nothing but nets” which then predict what driver should do next. When it’s correct, the carrot strengthens the network weights in a way that increases the likelihood of predicting correctly given similar input in the future. When it’s incorrect, the stick is applied - through a process known as back propagation of the correct answer - through the network by applying each node’s “loss function.” While I’m not as familiar as I’d like to be with the low level details, its likely that every prediction is, in fact, back propagated and the loss function either strengthens or weakens nodes given how close, or not, the prediction was. People NOT driving with FSD provide the carrot examples: input from the fleet sampled and filtered for “good” behaviors (safe, comfortable, efficient, legal) and provide the backbone of how to drive correctly for the input mentioned above. People USING FSD provide the sticks: what shouldn’t be done. The network is likely NOT trained on intervention and disengagement data, but more likely drop breadcrumbs where the vehicle performed incorrectly, which will trigger the capture of additional samples from the places that aren’t working. Another factor is “shadow mode” running on cars not using it. If the driver does something different than FSD would have, those areas are potentially useful breadcrumbs too. Think of it like this: lots of disengagements near this intersection, capture data to learn how humans drive through that point. Note how this will automatically, and in the same neural network, know the difference of how people drive in every locale Tesla has cars in. California is different than New York and all points in between. And in UK countries, they drive on the other side of the road. And in India and places in Asia that seem impenetrable to western drivers, the rules are WILDLY different. Neural networks are incredibly good at understanding these rules, and all that’s required for this is to input GPS data as input. Both sides are important, and the behavior filtering is the key to the whole “virtuous loop” we’re entering with v12. Improvements should increase exponentially going forward, IMHO and should lead to other locales outside North America coming on line quickly once the data flywheel is smooth and training compute continues to expand.
I need to try some interstate driving with this update. The previous version insisted on always putting itself in the left lane, which is illegal in Florida and some other states and also really annoying. That has really limited my use of it on highways - the place where I really would like it to work the most, so I've just been using cruise control, which works great. I think this week, I'm going to make an FSD video that pits it against certain challenges, and that will include interstate driving.
Love to say this but also not a knock but its completely that testing out and playing around with the FSD to make a video is actually dangerous because the car is so good it doesn't really warrant the demonstration. Also the videos are borning.
I would argue a system that slams on the brakes for no reason and can't follow navigation isn't so good. I understand these may not be the most exciting videos because, as you say, a lot of what the car does is really good. But, it still has key fails, so I'm going to continue to test.
@@DriveElectricWithDomenick Since 12.3 I've had no slamming on brakes anymore...humans don't often follow navigation perfectly and the car still got you there. From the video it looks like it did better, but again the fumbling with the navigation while driving is dangerous to just demonstrate FSD, which is why my comment exists. Also, you can just press the button and tell it where you want to go and not drive distracted, might be cynical but still a reality.
It's a feature demonstration, not an action movie. And people are going to be doing all kinds of things in the car. This will be a thing for years yet while the pretend "FSD" evolves. I'm sure it works reasonably well in places with no weather and very well marked roads, but that isn't all that many places.
@davidmccarthy6061 Looking forward to trying it out in some weather. A robo-taxi needs to be able to function in the rain. And actually, back when I was closer to the taxi business, rainy days were exactly when people needed rides most.
Dom, I noticed FSD seemed to hesitate and/or behave questionably almost entirely when you were on shaded/highly shadowed roads. It seems that especially those roads where there was a significant difference in light/dark caused “vision” problems. I noticed hesitation several times on the streets with high contrast light/shadow patterns. Serval times it also had to correct slight mistakes on the heavily tree-lined streets. Maybe you should try the same route on a cloudy bright day with less contrast between light/shadow. Record that test as well and compare side by side. That would make a great video: split screen showing the two videos with you voice over narrating what you observe on the videos and your experiences driving and differences observed. Could be quite interesting. 🤔
This is looking good. The new parking videos are impressive also.
To map out the whole route... Enter a destination, then Edit Trip to add another destination, until you got them all in there. You'll have to make sure the order is right, if not, drag them around. As you're driving it will just stop at the first destination and not automatically continue. However, if you tap Edit Trip as you're coming up to the destination, you can easily tap to remove the top destination, and it will automatically continue to the next.
(You may also be able to add all the stops via the app and tap Send To Car.)
I did my first FSD video a while ago and figured out the hard way how to make it auto advance to the next destination.
ua-cam.com/video/lpdoWfnClDs/v-deo.html
Thanks for this. I've tried using edit trip to set up the whole route, but it would navigate first to the last point instead of the first, which was super annoying and unexpected. More recently, the various way points were all near the top of my navi list, so quick and easy to find and tap on, but I've done a lot of road trips since my last video.
I like this app suggestion. I'll try that and the edit trip thing again.
It’s trained in two ways, carrot and stick.
Cars not using FSD submit videos that are filtered for “good driving” behaviors are the carrots that teach it how to behave “properly.”
Then, people using FSD provide the sticks when it misbehaves.
Blinker input, accelerator input, and takeovers teach it how NOT to behave.
Another factor is “shadow mode” running on cars not using it. If the driver does something different than FSD would have, that’s potentially useful data too, given that the driver’s behavior is considered good, or even more importantly, better behavior than FSD would have done.
Both sides are important, and the behavior filtering is the key to the whole “virtuous loop” we’re entering with v12.
Improvements should increase exponentially going forward, IMHO.
This is one of the most coherent and cohesive comments I’ve ever read on UA-cam. Well done, Sir! 🫡
@@loriallen67 🙏
Since you asked:
I believe it’s easiest to understand how Tesla is likely training FSD v12 in the classic idea of carrot and stick.
Cars not using FSD submit videos that are filtered for “good driving” behaviors. These videos, driver controls input, vehicle kinematics (including GPS), and annotated route data are fed into the “nothing but nets” which then predict what driver should do next.
When it’s correct, the carrot strengthens the network weights in a way that increases the likelihood of predicting correctly given similar input in the future.
When it’s incorrect, the stick is applied - through a process known as back propagation of the correct answer - through the network by applying each node’s “loss function.”
While I’m not as familiar as I’d like to be with the low level details, its likely that every prediction is, in fact, back propagated and the loss function either strengthens or weakens nodes given how close, or not, the prediction was.
People NOT driving with FSD provide the carrot examples: input from the fleet sampled and filtered for “good” behaviors (safe, comfortable, efficient, legal) and provide the backbone of how to drive correctly for the input mentioned above.
People USING FSD provide the sticks: what shouldn’t be done. The network is likely NOT trained on intervention and disengagement data, but more likely drop breadcrumbs where the vehicle performed incorrectly, which will trigger the capture of additional samples from the places that aren’t working.
Another factor is “shadow mode” running on cars not using it. If the driver does something different than FSD would have, those areas are potentially useful breadcrumbs too.
Think of it like this: lots of disengagements near this intersection, capture data to learn how humans drive through that point.
Note how this will automatically, and in the same neural network, know the difference of how people drive in every locale Tesla has cars in.
California is different than New York and all points in between.
And in UK countries, they drive on the other side of the road.
And in India and places in Asia that seem impenetrable to western drivers, the rules are WILDLY different.
Neural networks are incredibly good at understanding these rules, and all that’s required for this is to input GPS data as input.
Both sides are important, and the behavior filtering is the key to the whole “virtuous loop” we’re entering with v12.
Improvements should increase exponentially going forward, IMHO and should lead to other locales outside North America coming on line quickly once the data flywheel is smooth and training compute continues to expand.
Also, please turn on and demonstrate auto max speed option, it’ll fix a lot of the speed issues you’re experiencing.
Is this beta available for older model S/X?
Dudes doing DEWD loop
How does this perform on the interstate? I imagine highway driving would be my typical use case if I had FSD.
I need to try some interstate driving with this update. The previous version insisted on always putting itself in the left lane, which is illegal in Florida and some other states and also really annoying.
That has really limited my use of it on highways - the place where I really would like it to work the most, so I've just been using cruise control, which works great.
I think this week, I'm going to make an FSD video that pits it against certain challenges, and that will include interstate driving.
FSD v12 is city streets and some highways only.
Freeways are v11.4.9 ish.
The difference is palpable.
I'm wondering if it will be on version 99.1 and still in Beta?
🤣🤣🤣
Is this the free trial FSD?
No, I bought this car used a year ago and it was already on it.
Love to say this but also not a knock but its completely that testing out and playing around with the FSD to make a video is actually dangerous because the car is so good it doesn't really warrant the demonstration. Also the videos are borning.
I would argue a system that slams on the brakes for no reason and can't follow navigation isn't so good.
I understand these may not be the most exciting videos because, as you say, a lot of what the car does is really good.
But, it still has key fails, so I'm going to continue to test.
@@DriveElectricWithDomenick Since 12.3 I've had no slamming on brakes anymore...humans don't often follow navigation perfectly and the car still got you there. From the video it looks like it did better, but again the fumbling with the navigation while driving is dangerous to just demonstrate FSD, which is why my comment exists. Also, you can just press the button and tell it where you want to go and not drive distracted, might be cynical but still a reality.
It's a feature demonstration, not an action movie. And people are going to be doing all kinds of things in the car. This will be a thing for years yet while the pretend "FSD" evolves. I'm sure it works reasonably well in places with no weather and very well marked roads, but that isn't all that many places.
@davidmccarthy6061 Looking forward to trying it out in some weather. A robo-taxi needs to be able to function in the rain.
And actually, back when I was closer to the taxi business, rainy days were exactly when people needed rides most.
So don’t watch. Duh 🤦🏼♀️ 🙄