I was very encouraged by this Tesla FSD decision given how challenging thin horizontal objects still are for convolutional nets. I was sure I was going to have to intervene here, but FSD recognized the problem, reversed, and found another way out of the lot $TSLA

fsd 14.3.8 on ai4

I probably would have hit the brakes right after that first little lurch forward. One of the many reason why I won't be able to get a FSD streak worth bragging about.

my curiosity was too strong to disengage lol

@ChuckCook Paint the chain yellow..

Encouraging @ChuckCook

I have that at a school here. V13 failed, has worked on v14 ever since. Sees it and understands it.

@JCChristopher Oh, FSD passed @ChuckCook chain test in this scenario. That’s definitely encouraging

@JCChristopher Chuck will be pleased if you can reproduce this

@JoeTegtmeyer That’s terrific! It’s one area where I’m told improvement is needed.

@ChuckCook wen Chuck's chain though?

But there were multiple videos saying it wasn’t possible to solve.

this def doesnt prove its solved, just one encouraging anecdote also approaching head on is different than alongside/backing up

@JoeTegtmeyer For me V 14.3.8 is incredibly good. So close to sleeping in the back at this point.

Oh nice!

Yea. As signal improves they can increase the number of smaller things to look at…

It's done really well there!

This is very impressive, and I would say thin horizontal objects are probably harder for Tesla's model than traditional CNNs. Very impressive!

dude. you don’t understand this. they are not using cnns. there has not been convolutions in Teslas pipeline going back for at least 3 years. probably 5

thats not true. elon did say tesla doesnt use cnns much but that doesnt mean they dont use them at all fsd has shifted largely to transformers and end to end control (larger multimodal models) but afawk one of the first stages of the vision encoding uses a cnn to process high res images

if you know anything about modern ai, cnns are no longer part of the stack. i don’t know exactly what they are using but there is exactly zero chance they are using cnns. if @jamesdouma is around he can confirm for us.

Incredible

Hopefully it will continue to improve

Wonder if the chain's color is the same as the road, would it see it?

Who said that these are difficult for nets? Was it a Tesla engineer? I feel that there is a false narrative going around. Am I wrong?

@JoeTegtmeyer That’s great to see 👌

@ChuckCook TCT

@JoeTegtmeyer Nice!! I’ve been waiting for this.

@JoeTegtmeyer Good start 💯👍👏

So the car only wasted 5 minutes doing what you would have done to begin with...

Whoa 😮

Awesome. This was my last nagging concern about FSD (I come back to states each summer to demo the karst build and list of concerns was down to mainly this). Bravo Tesla

This is exactly what makes FSD so impressive. It didn’t just detect the obstacle—it understood the situation, reversed, and found another way out. 🚗⚡️ That kind of real-time decision-making is the future of autonomous driving.

Hope for Chuck Cook.

Saw that on your show. Interesting that the car needed to be so close before acting. Happy that it did. Curious if the speed of approach makes a difference.

Hope floats!