FSD discussion

[Knightshade]

The two main (new) problems I've seen in 12.3.3 so far are:

Wobbly/indecisive lane changes.... numerous times it'll put on turn signal, start to lane change, then either wobble between lanes for several seconds, or bounce back aborting the change then immediately start it again, sometimes taking 2-3 attempts to complete. I've seen this both with other cars around (confusing and annoying those cars) and even when there was nobody next to or behind me.

Addendum to above- it also seems to wait longer than I would to initiate changing into a correct lane for an upcoming turn (like if I'm on a 2-lane each way road and I know there's a left coming up in 1 mile, I wouldn't wait till the last 1/4 mile to get to the correct side.... 11.x usually didn't wait that long either, but 12 seems to.)

Other thing is weird yellow light behavior-- thrice now I've seen it when a light turns yellow briefly accelerate, then hard brake- one of those times it stopped me halfway across a crosswalk- not QUITE far enough forward to be dangerous but close- and AFAIK illegal too.

It also seems to still, at at least a couple intersections, act as if it's a 4-way stop when it doesn't have a stop sign (but the cross-lane does)-- but 11.x did that too.

The below-set-speed issue seems improved from 12.3 to 12.3.3 but definitely still not solved--- previously it'd drive like 10-15 below and if you tweak with accelerator it'd drop back down shortly after you lift off.... Now it'll drive like 5-7 below and if you accelerate up it seems to mostly stay there... scroll-wheeling set speed up 5-10 mph seems to be another alternative (it'll drive 5-7 below but that's about where you wanted it net).

The big 11>12 improvements overall so far I've seen are:
Better/smoother handling of roundabouts and non-90 degree intersections
Better acceleration after turns
Ability to do U-turns at all
11s tendency to dive into every new lane, even though it's often a turn lane turning/ending before needed or when I need to go straight, seems mostly fixed

That last being what was the most annoying thing about 11, just a shame that lane selection behavior is replaced by another, different, annoying one in 12.

 

[drtimhill]

Wobbly/indecisive lane changes.... numerous times it'll put on turn signal, start to lane change, then either wobble between lanes for several seconds, or bounce back aborting the change then immediately start it again, sometimes taking 2-3 attempts to complete. I've seen this both with other cars around (confusing and annoying those cars) and even when there was nobody next to or behind me.

Yeah II've observed this too. Also, when given two lanes to choose from that both go try same way (say for a two lane left turn) the car dithers a lot before it makes up its mind which lane to take.

 

[fasteddie7]

Here’s a new one. Both Hw4 model s vehicles drive differently than our Hw4 model x. There are 2 stop signs in my town that the x actually rolls through while both s vehicles stop. And there is another that it stops at and goes out immediately every time, where both s’s stutter pulling out every time.

 

[tivoboy]

Here’s a new one. Both Hw4 model s vehicles drive differently than our Hw4 model x. There are 2 stop signs in my town that the x actually rolls through while both s vehicles stop. And there is another that it stops at and goes out immediately every time, where both s’s stutter pulling out every time.

This is the thing… there should be NO difference AT ALL.

 

[elasalle]

Not sure if these were posted here (worth repeating)

the 1st tweet can mean only one thing.

https://twitter.com/x/status/1774673326583210476

https://twitter.com/x/status/1776134718406508548

 

[drtimhill]

This is the thing… there should be NO difference AT ALL.

Not true. NNs are deterministic. IF the input data is the same, the output data will be the same. But that's never the case. The cameras have different positions between models. The time of day will vary, as will the weather. The GPS reception will differ and so the car position will dither a little. Many of the weights in the network will be very similar, making small changes in input cause forks in the decision process. So behavior will be similar but not identical.

 

[tivoboy]

Not true. NNs are deterministic. IF the input data is the same, the output data will be the same. But that's never the case. The cameras have different positions between models. The time of day will vary, as will the weather. The GPS reception will differ and so the car position will dither a little. Many of the weights in the network will be very similar, making small changes in input cause forks in the decision process. So behavior will be similar but not identical.

In some ways I agree with you, sure.. but all things being equal in the ENVIRONMENT, whatever the different cameras are interpreting, it’s ACTION should be the same. It should come to a stop at a stop sign the same each time. I don’t really know how much GPS should affect VISION for FSD and inpreting a stop sign, or the proximate location to another car, or the objective positioning on the same road between the median line and the side line. Technically, certainly same MODEL CAR, should be the same. It shouldn’t be 6 inches more to the right than the same car doing the same route on the same day at the same time. And one interpretation of a flashing yellow light, coming up to it at the same place in time should react the same way. These aren’t individual NN’s that should be thinking differently on their own.

Time of day, light vs. dark, rain vs sun blindness. Sure, interpretations are going to be different and ACTIONS will be different. but, assuming those things are equal - for the most part, we still see variance and that is something that IMO should not occur.

Will they be identical. No, that’s just the nature of moving a 5000 lb object around without actual rails, and variances in movement through space. But many of these differences that ppl discuss are much farther from identical than I at least would presume.

 

[drtimhill]

In some ways I agree with you, sure.. but all things being equal in the ENVIRONMENT, whatever the different cameras are interpreting, it’s ACTION should be the same. It should come to a stop at a stop sign the same each time. I don’t really know how much GPS should affect VISION for FSD and inpreting a stop sign, or the proximate location to another car, or the objective positioning on the same road between the median line and the side line. Technically, certainly same MODEL CAR, should be the same. It shouldn’t be 6 inches more to the right than the same car doing the same route on the same day at the same time. And one interpretation of a flashing yellow light, coming up to it at the same place in time should react the same way. These aren’t individual NN’s that should be thinking differently on their own.

But to your very point this IS the same NN. So do what you attribute the difference? Goblins? One thing to understand about NNs is that though they are deterministic in the sense same input = same output, they are non-linear, so you dont know what even a small input perturbation will do to the output (well you know probabilistically via training, but that's not quite the same).And remember the car is NOT just relying on vision, it also has map data input, route input, accelerometer, compass, etc etc.

 

[heltok]

Good luck getting all the photon counts from all the cameras to be identical over a few hundred frames from two different time points of a vehicle moving. Even getting the car to move two identical paths is hard enough even without worrying about the light.

It's called probabilistic robotics for a reason.

 

[drtimhill]

Will they be identical. No, that’s just the nature of moving a 5000 lb object around without actual rails, and variances in movement through space. But many of these differences that ppl discuss are much farther from identical than I at least would presume.

The basic problem here is you are assuming a linear relationship, which is NOT how NNs work (nor would they be useful if they were). You ass assuming "small different in input" equals "small difference in output". but NNs dont work like that.. sure, you can train it to closely simulate a relationship like that, but there may be very specific inputs that differ only very slightly but generate significantly different output (mathematically this is embiodied in catastrophe theory). There are of course already many NNs that operate in a similar way already driving quite well, we call them "humans".

 

[willow_hiller]

I don't think I've seen this posted yet. But it's nice to hear a story of FSD actually making a difference in someone's life.

https://twitter.com/x/status/1777083263535976484

 

[JupiterMan]

This is a FSD 12.3.3 video with a difference - from the perspective of someone who has worked on autonomy (not Tesla) for 20 years and is now trying FSD for the first time, including some history of the projects he's worked on. good stuff!

 

[Electroman]

FSD is getting kudos from a lot of my friends who own Tesla, who haven't even tried Autopilot all these days.

It is indeed a ChatGPT moment

 

[Knightshade]

FSD is getting kudos from a lot of my friends who own Tesla, who haven't even tried Autopilot all these days.

It is indeed a ChatGPT moment

That's the tricky thing to me.... Someone who's never even used AP (which is weird since it's free on all the cars for years now) trying FSD the first time is going to be impressed.... but as someone who's been in the beta for years now, 12.3.3 is more a "Yes, this is definitely smoother, but it's still just L2 and still needs to improve before it's anything more"

Is "chatGPT moment" people realizing something decent exists at all--- or is it "People with no previous knowledge of a thing thinking it's better than it actually is based on suddenly getting a tiny taste of it but not enough for the flaws to be obvious yet?"?

That second one is how ChatGPT turned out... all the folks who eventually discovered it just makes up imaginary sources and whatnot and realizing it can't be relied on 100%...but it took a while.

 

[Electroman]

I think the comparisons with ChatGPT is quite apt here.

People were pleasantly shocked that it could even do what it can, but they also soon found out the first iteration out of the block has its limitations, but has tremendous promise - true for both

 

[EVCollies]

My morning commute has a very busy light a block after a railroad track, although FSD did not visualize that hazard on the screen, it did not stop the car on the tracks waiting for the green light, Bravo TSLA.

 

[jerry33]

Cat on 12.3.3.

https://twitter.com/x/status/1777676078758195416

 

[Electroman]

I am impressed at how it slows down for rail tracks and speed bumps even within an apartment parking lot. It either sees them visually and takes actions, or even those bumps are mapped out.

So far my only challenges are unprotected 'Chuck Cook' lefts. It misses good opportunities and waits for traffic to clear on both directions, annoying cars behind me.

Also there was one instance it took a left turn on yellow very confidently, when there were cars on the left turn lane on the opposite direction waiting to turn, and that obstructed the visibility to determine if the coast is clear or not. Alteast from my viewpoint I couldn't figure out if there are any oncoming cars or not, but car went right ahead and took the turn confidently. It was a three lane (each direction) and a left turn lane at that position. There were no cars oncoming though.

Was it just lucky, or it has a better visibility than me?

 

[elasalle]

https://twitter.com/x/status/1778067634430328835

 

[fasteddie7]

Looks like more than just employees.