FSD Traffic Light Update??

This isn’t good. Highlighting in blue and displaying the colors meant that the car thought those lights were for it, even though that’s incorrect.

?? It responded as though it only noticed the front facing two, so I dk if it was a fluke in the graphic or if it was a fluke in drive response

It’s certainly getting there though! I like noticing these little advances.

With V12 the visualizations come from a separate system that has no bearing on how the car is actually driving, so it's irrelevant anyway.

Oh that explains a lot! 😅 I feel stupid… regardless, it’s reassuring the graphics-side is starting to catch up.

Thanks!

Don't feel stupid. It's logical to think that the visualizations are based on what the system is actually seeing/thinking, and prior to V12 that was actually the case. Unfortunately end-to-end machine learning means that they can't do that anymore. The only part of the visualizations that's "real" now is the projected driving path, since that comes from the acceleration/steering outputs of the neural net.

Regarding visualizations… any word on whether ASS gives you a route it plans on taking? Or does it just show camera feeds… hard to tell with the videos I’ve seen.

Why can’t I just purchase this view on my screen. I loved the more zoomed out, seeing more screen of FSD.

You totally can... that'll be $100/month!! lol

Same, although in practice that might get busy when you’re at speed

Yes, it does give you a planned path on the app, in addition to the camera feeds. ASS also isn't end-to-end.

That’s fantastic news, just what I was hoping for yet doubtful for 😂

Please don’t give them any ideas for further upcharges

This is not true

It is true. You can't generate visualizations like this from an end-to-end system where the only outputs are acceleration, steering, and turn signals. It's not like before when there were a bunch of different neural nets, each outputting things such as vehicle positions, traffic light colors, etc. with hand-written code using those outputs to tell the car how to drive. Now it's just one big neural net that goes all the way from the camera inputs to the control outputs — all the way from one end to the other end. There are no intermediate steps that can be used to render a robust visualization.

But this is not how Tesla’s systems works. Major misconception It is not a neural net takes raw camera images as the only input and outputs a vehicle path to follow.

The map projected driving paths are no longer a great indicator either. There was a case yesterday or a few days ago where the map indicator to go past the crossllights, and make a left turn into the parking lot. It made a left turn at the cross lights, then a right turn into the parking lot completely ignoring what was indicated on the NAV system.

It is with V12. It takes camera frames, map data, turn signal state, velocity state, etc. as input and outputs acceleration, steering angle, and turn signal values projected over time. Those are the only outputs. It's not a "hydranet" like before with many neural nets each outputting something different. It's one huge neural net with only those outputs. That's what end-to-end means. Just one neural net that goes from the inputs all the way to the final outputs.

Except it takes in more inputs than what you mention, it also takes in sensing state from the other networks.

I'm not talking about the map navigation path. I'm talking about the projected driving path on the FSD visualization. It's the thick blue line that comes out from the front of the car model on the screen. The map navigation path is an input to the neural net, not an output. The projected path coming off of the car is a visualization of the acceleration and steering angle outputs of the neural net.

No, there are no other networks used in the system. It's end-to-end now. Just one big neural net. That's the definition of end-to-end.

End to end does not have a rigid definition. Yes I know end to end commonly implies a network from sensing inputs to control, without intermediate networks or steps. It’s not one big end to end network

Further? They just cut the price in half.

It is. They've said this over and over again. It wasn't prior to V12, but it is with V12. That was the massive change of V12.

V12 was a big change that uses a larger neural network yes.

…I paid full price

The subscription price never increased. It started at $200 per month and then *decreased* to $100 per month.

lol… before subscriptions

But thanks for trying to make me feel better 😝

I understand, but the guy you replied to was talking about the subscription. The subscription only decreased in price. The outright purchase increased and then decreased.

OHHHHH apologies I wasn’t on all comments

But yea… still clearly bitter about the subscription thing one week after I picked up 🥴

I wish they would just get rid of the visualization. I wonder if they are still using it for other reasons like a safety harness. I would totally be fine with them getting rid of it.

Nah, it has a big cool factor, and it's still useful for seeing where the projected path is going in relation to the environment around it (plus systems like automatic emergency braking are still based on it). It is a shame that it's so disconnected from FSD now though. That's a downside of end-to-end, but with how much better it drives, it's totally worth it.

Can’t they train a different net to produce some sort of datastructure that can be used to visualize things? Should be much easier than driving no? Or have the end to end net also produce outputs of what it sees on those camera frames?

They already have nets for that, which they're using for the visualization. It's the V11 perception nets which they're still running concurrently with V12 for visualizations and certain other features such as forward collision warning. The end-to-end net can't output stuff like car positions because car positions aren't in the training data. The training data is basically "when the world looked like this, the driver pressed the pedal this much and turned the steering wheel this much". That's the fundamental thing that makes end-to-end work. The net learns to mimic how humans drive based on readings from the pedals and steering wheel for each frame of the videos it watches. Within that data, there's nothing telling it "there is a car at X:54.7, Y:12.9". Data like that is typically produced with human workers labeling the frames and drawing boxes around cars, but when you have a system that simply mimics human driver steering wheel and pedal movements matched with videos, such labels would be detached from what is actually causing those steering wheel and pedal movements. It's not "when there's a car here, turn the steering wheel this much". It's "when the pixels are these colors, turn the steering wheel this much". The former gives less overall information and therefore produces an inferior result (and requires a ton of human labor, which limits how much training they can do). That's why they train with pixels and not semantic classifications.

If it's no longer what the car sees then I definitely think I should be allowed to at least hide the damn thing. Let me replace it with a full top to bottom music player.

Amazing point, I would also love a split screen music/map setup. I would love that as well. The visualization literally has no use other than the cool factor and marketing.

You are onto something. It may have been doing this internally before, and not being visualized. It may have been doing it "automatically" as an internal NN-feature. Tesla engineers identified the functionality, and solidified it and can rout the output of that network to the visualization layer, and we can see what the car knows. I've always had issues on certain intersections were it thinks some lights meant for side traffic is meant for me. V12 mostly solved it, but the light visualization was still wrong, and the car, is still behaving oddly there. If this is in 12.5 stack, it should completely solve the "odd" behavior on my troubling intersection.

Or full map

Could be worse, I bought my car at the highest price and before the tax credit.

Yeah totally… I mean we could totally all be on fire… so I guess we’re lucky 😂

Or… worse… we could be in the UK where I’m not sure they’re ever seeing FSD

I like being able to confirm it recognizes people

(When not on highways I’m in *populated* areas)

https://preview.redd.it/59ro3vzio5od1.jpeg?width=3024&format=pjpg&auto=webp&s=637d77c8fcd2a37cd4a62e92dc924e667e36ed08 Better example

They could just give it to us, makes no sense that if you don’t subscribe you get a more limited view. It’s just not as safe

This was something I had considered when discussing a different topic but maybe one that might fit here: do you think there is a max load in terms of FSD subscriptions?

So the FSD interface looks so much cooler than the standard. Is there anyway to get this on the normal model 3 without FSD - i.e. don't use FSD but just get the view

I don’t know… i didn’t know they looked different? Can you show us?

This is often repeated and completely false. The visualization comes from the vision stack which is very much still used in v12. V12 is not end to end in the sense that it’s one neural network that takes images and returns acceleration. The only thing v12 changed was it replaced the path finding code with another neural network. That new path finding neural network takes input from the vision stack exactly as before.   @greentheonly who actually digs into the binaries says the same thing.     https://x.com/greentheonly/status/1772763996749230168?s=46 https://x.com/greentheonly/status/1761596692120375396?s=46

Because you can just look out your windshield and see it with your own two eyes.

Because it's a big, stupid advertisement for their FSD. If I could pay to get rid of it, I would. Preferably would like to see more of what's behind me -- than what's in front of me.

Wasting all that screen space to show me what I can see out my windshield is so asinine -- it dives me crazy that they waste all that space.