FSD v12.x (end to end AI)

[TK211X]

Is it possible the trigger for the next wave is a % of the previous pending installs to download coupled with a date trigger so it doesn’t drag on too long?

If 80% download then next wave.
Or 24 hours then next wave.
Unless manual intervention to stop?

I can’t imagine there’s a Tesla employee manually pushing these waves.

Would be fun to try and figure out the triggers.

 

[Iain]

I was lucky enough to find a fresh download waiting for me this morning. Will this be the version my wife won’t be looking for the brake peddle in the passenger seat?

 

[uscbucsfan]

I was lucky enough to find a fresh download waiting for me this morning. Will this be the version my wife won’t be looking for the brake peddle in the passenger seat?

I'd say it passes the wife test, yes.

 

[rlsd]

My v12.3 drive over Rancho Santa Fe San Diego yesterday. Speed limit mostly 55 mph. Smooth like silk.
I din't have crossing over double yellow line problem like PianoAI had. - See left repeater video.

16950 Via De Santa Fe to 417 S Sierra Ave

www.google.com

 

[lzolman]

I'd say it passes the wife test, yes.

Hey, the only way it passes "the wife test" is when it passes the "MY wife test"

 

[uscbucsfan]

Hey, the only way it passes "the wife test" is when it passes the "MY wife test"

My wife doesn't like FSD and today she said, "This is pretty natural"...then she got annoyed I tried a turn 3 times,

 

[Dennisis]

Updated overnight and went for a short drive this a.m. First impressions: 12.3 is markedly better and yes I'm liking the autospeed setting so far. Overall drives very smoothly and confidently and lane selection is much improved. I think this version is dressed to impress!!

 

[fasteddie7]

Everything I see seems to indicate the cars seem to ride and turn much closer to curbs than earlier versions

 

[yerEVan]

Am I reading teslafi correctly? 481 installed and 197 pending, overall? I know it is a small sample of actual NA installs.

 

[kabin]

My v12.3 drive over Rancho Santa Fe San Diego yesterday. Speed limit mostly 55 mph. Smooth like silk.
I din't have crossing over double yellow line problem like PianoAI had. - See left repeater video.

16950 Via De Santa Fe to 417 S Sierra Ave

www.google.com

Good to know. His roads might have been a little tighter blind turns. In any event it'll be years before we can drop our guard given the lack of redundancy and box of chocolate effects.

 

[Iain]

My first drive was zero intervention, zero peddle push… perfect. I had a U-turn and had a message, please assist unable to complete maneuver, but it did it like a pro

 

[Cherry Wine]

Do you have a legacy vehicle? Looks like Tesla pushed legacy cars last week to this version. Mine included. My speculation is because no interior camera, maybe they need more validation first.

Yes, it's a 2020 MX.

 

[Ben W]

Can someone explain this:

On the old system of hard-coded instructions, if something isn't working, say the car is driving too slowly, it can be fixed by changing some programming.

With this new neural net system, where the car is basing its behavior on examples of others driving is similar situations, how does one fix a problem like driving too slowly?

In general, by gathering a whole bunch of video examples of the "correct" behavior, and using that in the next training set for the neural net (which tries to mimic the training set), and rolling it out in the next version. It is sort of a black art, though.

 

[rlsd]

Everything I see seems to indicate the cars seem to ride and turn much closer to curbs than earlier versions

I saw that too but it's not too close. It seems FSD takes the approach: be defensive as much as possible.

 

[Ben W]

I'd settle for it being able to just drive me (yes, drive me) for a half-hour back roads jaunt without my having to be on shpilkes the entire time. My good friend in Cupertino just got it, so Boston can't be TOO far behind....

Until FSD reaches true L4 reliability (less than one safety mistake per million miles), the shpilkes (love the word) will regrettably be necessary. There is going to be a vast uncanny valley where the reliability is enough that you REALLY want to just relax and stop paying attention, but not enough to be actually safe. We are not to this uncanny valley yet for city streets (currently about one mistake per mile, in my experience), but are approaching it for highway (currently about one mistake per 100 miles).

 

[rlsd]

In general, by gathering a whole bunch of video examples of the "correct" behavior, and using that in the next training set for the neural net (which tries to mimic the training set), and rolling it out in the next version. It is sort of a black art, though.

I have the impression that they twisted some parameters to make the car behaves a little bit differently, not always retraining FSD. For example: v12.3 is different than v12.2.1: in v12.3 the car is closer to the curb in most situations, not in the center of lane lane like v12.2.1. Did they retrain to have this new behavior?

 

[FSDtester#1]

Getting ready to take it out for the first time for the Beaver test.

I guess when they built the houses on my street, they had a fascination with Beavers.

So they decided to put one in the middle of the entrance to my neighborhood in the middle of the street in the ground, protruding about 3 inches up above the asphalt.
You can easily drive over it, just a couple of bumps.

Fsd has always when pulling onto my street on a left turn entrance, been calm until it sees the beaver, and then yanks the yoke hard left and emergency brakes. This has been such a pain point, I disengage now all the time driving into my neighborhood.

So V12, if you can avoid throwing my yoke suddenly near the beaver , I will be impressed.

 

[Ben W]

I have the impression that they twisted some parameters to make the car behaves a little bit differently, not always retraining FSD. For example: v12.3 is different than v12.2.1: in v12.3 the car is closer to the curb in most situations, not in the center of lane lane like v12.2.1. Did they retrain to have this new behavior?

Yes, I believe each dot-release is fully retrained. The amount of C++ code and explicit control logic is rapidly diminishing; I'm not sure what sort of controls and parameters are still explicitly coded, other than the user-facing settings.

 

[rlsd]

Yes, I believe each dot-release is fully retrained. The amount of C++ code and explicit control logic is rapidly diminishing; I'm not sure what sort of controls and parameters are still explicitly coded, other than the user-facing settings.

I am not talking about C++ parameters.
See training vs parameters here:

Training computation vs. parameters in notable AI systems, by domain

Computation is measured in total petaFLOP, which is 10¹ⵠfloating-point operations estimated from AI literature, albeit with some uncertainty. Parameters are variables in an AI system whose values are adjusted during training to establish how input data gets transformed into the desired output.

ourworldindata.org

Parameters in notable artificial intelligence systems

Parameters are variables in an AI system whose values are adjusted during training to establish how input data gets transformed into the desired output; for example, the connection weights in an artificial neural network.

ourworldindata.org

 

[sleepydoc]

Can we just push 12.3 out to everyone then... please? PLEASE. heh

No. We’ll stick with all the cool people on v11!