V13 is much more ‘human’ than V12 in my experience so far. Biggest issue for me is picking the wrong lane, then having to do a last moment change. Yesterday it moved into an exit lane it shouldn’t have, and I had to take over to navigate back to freeway.
I picked up a 2024 AWD MY last month and have been using FSD more and more. I’m pretty impressed with it, leagues above my old 2018 M3 with “enhanced AP” One thing I’ve seen is an inability to define a red light that has a “no right turn” limitation sign, I’ve had to stop it from turning a couple of times. I also don’t know how it’ll handle snow covered roads here in western NY, but I’ll try it when it’s safe to do so. That aside, it’s been pretty awesome and I think it clearly shows what’s possible in the not very distant future. One thing I will say though - keeping the cameras view unobstructed in inclement weather will be a constant challenge. Driving at night on a wet and somewhat slushy edged highway, my B pillars cameras were constantly alerting “obstructed” even when they weren’t. Odd. But overall huge progress any way you slice it.
Yes! Huge upgrade from older hardware vehicles. Unfortunately, FSD still can’t read signs such as “no turn on red”. But I did hear from some of my other viewers that it’s on Tesla’s roadmap!
6:15 FSD can't read any signs, like a human would read text. However, it is trained on real world vision data. So if there's a sign, a road marking or an object attached to typical behavior it can remember this. So think of this more like context cues. If nobody drives through the exit of an intersection with this sign next to it, there may be a high bias towards using the other exits naturally evolving in the neuronal network. You can also enforce this learning behavior by feeding it actively clips of wrong behavior and instead of rewarding, punishing the network for this. This is how "running a red light" or "running a stop sign" is actively trained with negative signals to avoid.
I see! This is interesting. So it is almost as if its taking cues from other drivers on the road and basing it's decisions on that. I hope Tesla releases functionality on "no turn on red" functionality. That one is a critical one to read.
@@DriveElectricToday well, FSD already understands partly that "no turn on red" scenario. But there's guard code on top of FSD, which runs off of that FSD v11 network which also does the visualizations. It will for example disallow FSD to drive straight on red lights ... and also makes sure FSD does a full stop on stop signs. That's why you see FSD sometimes doing a weird double stop. Meaning it does a stop to nearly a standstill, and then FSD tries to proceed, and it suddenly seems to decide otherwise and do a full stop. The second stop is quite uncomfortable compared to the smooth first stop. The reason for this is, that this is the guard code overriding FSD here. Same goes for no turn on red, which is based off of map data currently. So if there's a map data saying that turns are not allowed, FSD may try to proceed, but won't be allowed. If there's no map data saying this FSD may recognize the sign by itself, or it does not. The main issue with no turn on red signs is that it's extremely variable where those signs are placed and how exactly they look like, and there are a lot of similar looking signs on the road - meaning white signs with black text - which FSD learned to ignore. So either they need dedicated "where's a street sign" detection and then recognition what's written on the street sign - which is quite unlikely given that they want it to have just one stack going from the camera to the vehicle controls, or they need to feed it *A LOT* more training data. In training neuronal networks we're usually talking about 2-3 magnitudes more training data on that case, until a certain scenario is properly recognized. So say they're feeding in currently 100'000 cases how to safely turn right on an intersection, they would need additionally 10-100 M to get that scenario right, for both cases: Where it's allowed and where not. That's why they choose the easy route currently, with just reading map data and avoid training FSD for this scenario, as it may lead to a flipped behavior, where it won't turn right when it can.
They need to add a setting for neighborhoods that sets the max speed to the posted speed. And fix lane centering on neighborhood streets. It’s abysmal.
I feel like we're not too far from the possibility of having it drop you off at the door and then parking for you. That would be the future I would love to live in
Thanks for your videos on version 13.2.2. Question for you - Going through neighborhood streets where it is very narrow, parked vehicles on the right side of the street, and an approaching vehicle is coming at you - Has FSD ever realized that it can't make the tight squeeze between approaching vehicle and parked cars and then stopped to avoid side scrapes???? Right now, I don't even trust v 13.2.2 in situations above and purposely disengage FSD and let the other vehicle go through first. To me it just looks impassable for 2 vehicles to meet and pass safely. I don't know if version 13.2.2 has the smarts to realize it cannot do that narrow move and stop. Thanks again. Take care.
That’s a super good point. The safest thing to do in a situation as you described is to absolutely disengage and allow the oncoming vehicles to pass before moving forward. In my situation depicted in the video, I was super curious to see how FSD responded to it, and well, it passed with flying colors. I was very excited to see how well it responded. There have other times where it tried to squeeze through tight spots and I have had to disengage. I must say, those situations happen mostly in the “hurry” profile. @dathyr1 - How are you enjoying your free trial so far of version 13.2.2?
@@DriveElectricToday Thanks for your information and what you have seen with FSD going through neighborhoods or tight streets. My free FSD trial is performing very well. I have one more week before it will be disabled. Version 13.2.2 is much improved over the 12.5.6 version I tried before. Being retired, Still I feel for all my short daily driving trips that FSD is not required.
It's interesting that it can just choose a random route. I wonder if it's a learning experience for it? Is that how the AI is learning how to drive now? Do you clean off the cameras before every trip? Just curious.
Yup! FSD doesn’t require a destination in order to activate. Pretty near feature! Not sure if that is how the software learns. I do know that Tesla leverages a “neural-network” design, in which situations from other Tesla vehicles around the world is used to better the overall software. So, in a sense, everyone contributes to the software. I typically don’t clean the sensors or cameras. I will do a car wash about 1-2 per month which seems to do the trick.
You know, I was thinking about that too! Every time I’ve tried engaging the blinker with no destination the car jerks and ultimately goes straight. Have you had success with that?
It definitely has room for improvement when it comes to speed control. Have you seen any difference when trying the different modes? Chill vs Standard vs Hurry?
It is much safer to use human brain, human eyes, and years of life experience in *Residential* *Areas* . FSD is still just a buggy Level 2 assistant. Unable to observe, think, reason, or analyze like a human.
Be sure to check out 3:30 - 6:00 to see how well FSD does in residential areas!
Wow, Barry! Your FSD commentary is both insightful and engaging-watching it handle speed bumps and snowy roads with confidence is fascinating!
@@SagorAhmEd-SEO-Strategist Thanks for your great feedback!
V13 is much more ‘human’ than V12 in my experience so far. Biggest issue for me is picking the wrong lane, then having to do a last moment change. Yesterday it moved into an exit lane it shouldn’t have, and I had to take over to navigate back to freeway.
it also hits all the potholes at full speed
Same exact problem
Yes I have had incorrect lane choices many times. I’m hoping that is better in 13.2.4
I find it nice how you make your voice sound like it's a voicer-over.
Thank you for the great feedback! I will continue to record in this fashion. Thanks for the support!
@DriveElectricToday Rock on, dude! Great stuff
I picked up a 2024 AWD MY last month and have been using FSD more and more. I’m pretty impressed with it, leagues above my old 2018 M3 with “enhanced AP” One thing I’ve seen is an inability to define a red light that has a “no right turn” limitation sign, I’ve had to stop it from turning a couple of times. I also don’t know how it’ll handle snow covered roads here in western NY, but I’ll try it when it’s safe to do so. That aside, it’s been pretty awesome and I think it clearly shows what’s possible in the not very distant future. One thing I will say though - keeping the cameras view unobstructed in inclement weather will be a constant challenge. Driving at night on a wet and somewhat slushy edged highway, my B pillars cameras were constantly alerting “obstructed” even when they weren’t. Odd. But overall huge progress any way you slice it.
Yes! Huge upgrade from older hardware vehicles. Unfortunately, FSD still can’t read signs such as “no turn on red”. But I did hear from some of my other viewers that it’s on Tesla’s roadmap!
6:15 FSD can't read any signs, like a human would read text. However, it is trained on real world vision data. So if there's a sign, a road marking or an object attached to typical behavior it can remember this. So think of this more like context cues.
If nobody drives through the exit of an intersection with this sign next to it, there may be a high bias towards using the other exits naturally evolving in the neuronal network.
You can also enforce this learning behavior by feeding it actively clips of wrong behavior and instead of rewarding, punishing the network for this.
This is how "running a red light" or "running a stop sign" is actively trained with negative signals to avoid.
I see! This is interesting. So it is almost as if its taking cues from other drivers on the road and basing it's decisions on that. I hope Tesla releases functionality on "no turn on red" functionality. That one is a critical one to read.
@@DriveElectricToday well, FSD already understands partly that "no turn on red" scenario. But there's guard code on top of FSD, which runs off of that FSD v11 network which also does the visualizations.
It will for example disallow FSD to drive straight on red lights ... and also makes sure FSD does a full stop on stop signs.
That's why you see FSD sometimes doing a weird double stop. Meaning it does a stop to nearly a standstill, and then FSD tries to proceed, and it suddenly seems to decide otherwise and do a full stop.
The second stop is quite uncomfortable compared to the smooth first stop. The reason for this is, that this is the guard code overriding FSD here.
Same goes for no turn on red, which is based off of map data currently. So if there's a map data saying that turns are not allowed, FSD may try to proceed, but won't be allowed. If there's no map data saying this FSD may recognize the sign by itself, or it does not.
The main issue with no turn on red signs is that it's extremely variable where those signs are placed and how exactly they look like, and there are a lot of similar looking signs on the road - meaning white signs with black text - which FSD learned to ignore.
So either they need dedicated "where's a street sign" detection and then recognition what's written on the street sign - which is quite unlikely given that they want it to have just one stack going from the camera to the vehicle controls, or they need to feed it *A LOT* more training data.
In training neuronal networks we're usually talking about 2-3 magnitudes more training data on that case, until a certain scenario is properly recognized.
So say they're feeding in currently 100'000 cases how to safely turn right on an intersection, they would need additionally 10-100 M to get that scenario right, for both cases: Where it's allowed and where not.
That's why they choose the easy route currently, with just reading map data and avoid training FSD for this scenario, as it may lead to a flipped behavior, where it won't turn right when it can.
They need to add a setting for neighborhoods that sets the max speed to the posted speed. And fix lane centering on neighborhood streets. It’s abysmal.
Definitely agreed here
I feel like we're not too far from the possibility of having it drop you off at the door and then parking for you. That would be the future I would love to live in
Yup! We aren’t too far from there. I’d say within the next 2 years or so!
Can you check out some tight roads, like in mountains with serpentines?
That would be a great test! Yes I will absolutely plan this in a future video.
@@DriveElectricToday thanks!
Thanks for your videos on version 13.2.2.
Question for you - Going through neighborhood streets where it is very narrow, parked vehicles on the right side of the street, and an approaching vehicle is coming at you - Has FSD ever realized that it can't make the tight squeeze between approaching vehicle and parked cars and then stopped to avoid side scrapes????
Right now, I don't even trust v 13.2.2 in situations above and purposely disengage FSD and let the other vehicle go through first. To me it just looks impassable for 2 vehicles to meet and pass safely. I don't know if version 13.2.2 has the smarts to realize it cannot do that narrow move and stop.
Thanks again. Take care.
That’s a super good point. The safest thing to do in a situation as you described is to absolutely disengage and allow the oncoming vehicles to pass before moving forward. In my situation depicted in the video, I was super curious to see how FSD responded to it, and well, it passed with flying colors. I was very excited to see how well it responded.
There have other times where it tried to squeeze through tight spots and I have had to disengage. I must say, those situations happen mostly in the “hurry” profile.
@dathyr1 - How are you enjoying your free trial so far of version 13.2.2?
@@DriveElectricToday Thanks for your information and what you have seen with FSD going through neighborhoods or tight streets.
My free FSD trial is performing very well. I have one more week before it will be disabled. Version 13.2.2 is much improved over the 12.5.6 version I tried before.
Being retired, Still I feel for all my short daily driving trips that FSD is not required.
It's interesting that it can just choose a random route. I wonder if it's a learning experience for it? Is that how the AI is learning how to drive now? Do you clean off the cameras before every trip? Just curious.
Yup! FSD doesn’t require a destination in order to activate. Pretty near feature!
Not sure if that is how the software learns. I do know that Tesla leverages a “neural-network” design, in which situations from other Tesla vehicles around the world is used to better the overall software. So, in a sense, everyone contributes to the software.
I typically don’t clean the sensors or cameras. I will do a car wash about 1-2 per month which seems to do the trick.
you don't need to takeover to make those turns, just give the indicator and watch the maic!
You know, I was thinking about that too! Every time I’ve tried engaging the blinker with no destination the car jerks and ultimately goes straight. Have you had success with that?
@DriveElectricToday yes, everytime. Since v12. 5
Can I ask you what FSD costs??
FSD cost is based on either a subscription or outright cost. The subscription is $99 per month, and the outright cost is around $8k.
🥳❤💯💢💥💫👍👍👍
Interesting. I've stopped using FSD in residential areas because Tesla has horrible speed control.
It definitely has room for improvement when it comes to speed control.
Have you seen any difference when trying the different modes? Chill vs Standard vs Hurry?
I haven't had that experience and it's set at 60% above speed limits, but I will change it to 25%
It is much safer to use human brain, human eyes, and years of life experience in *Residential* *Areas* . FSD is still just a buggy Level 2 assistant. Unable to observe, think, reason, or analyze like a human.
Yes and No
I cant believe people drive these things, look at that bland ass interior with an iPad slapped in it 😂
You can't. Just as there are people who can't use smartphones.