FSD v12.x (end to end AI)

[FSDtester#1]

I didn’t think actual smart summon was part of 12.4. For some reason I assumed it was 12.5

You are correct, he just wants ASS earlier..

He may even venture into an Atlanta strip club tonight and ask for 2024.9.5
They won't know what he is saying. They may offer 10.69.420

 

[Mardak]

I wonder if end2end has any ability to make plans and stick to them (unless other circumstances dictate they can't be stuck to) or whether that is an inherent weakness of the approach?

There doesn't necessarily need to be an explicit plan that is shared with subsequent predictions. At 25:10, 12.4 needed to stay in the left of 3 forked lanes to continue straight. (The left lane before the intersection forks to 2 left-turn lanes and the current/right lane forks to a straight and 2 right-turn lanes to leave the San Jose Almaden Costco).

Here the indecision was somewhat acceptable in that if it stuck to the plan in the 1st frame, it would have ended up in the incorrect lane forced to turn right. But the switch back in the 3rd frame was unnecessary especially given that it had an even better view of the painted arrows. With additional training, end-to-end should predict higher action probabilities of taking the left fork for each of these frames and lower probabilities for taking the wrong lanes such that the correct left fork would be the most likely action without indecision of the blue path visualization.

There could be situations that require more persistent plans, but indeed that would need additional care to make sure it doesn't override new inputs such as detecting a red light runner. At least for picking a turn lane, it shouldn't really need this additional concept, and other existing data like recent memory (e.g., another vehicle covers up painted arrows) might be sufficient.

 

[lzolman]

I use the yoke top bar as a reference. Top edge down for "right" button. Lower edge up for "left" button. Like a reverse stalk.

That seems like it might work, I will give it a shot. I've been trying to use the ridge between the directional buttons and go up or down from that, but it takes me too long to find the ridge and be clear that I've found it. It's easier to find those top and bottom edges of that part of the yoke. Thanks for the tip.

 

[Mardak]

And they are going to need to keep their mileage numbers up. TBD how they will accomplish that if v11 is not quality enough vs. AP. Maybe it is worse? Then they have to go v12 which could be a while. Guess mileage requirements may force their hand.

As in the mileage rate will decrease as people stop using FSD free trial? Even without Tesla doing anything special, the mileage numbers will still go up presumably faster than before 12.x as there's more vehicles purchasing or subscribing FSD Capability in addition to the improved capabilities of 12.x.

The slope of FSD cumulative mileage has gone down in the past, so Tesla doesn't seem to be chasing this metric nor forced to do something to keep the rate up. Although hands free nag with Vision attention monitoring might help increase usage of FSD.

 

[Dewg]

As in the mileage rate will decrease as people stop using FSD free trial? Even without Tesla doing anything special, the mileage numbers will still go up presumably faster than before 12.x as there's more vehicles purchasing or subscribing FSD Capability in addition to the improved capabilities of 12.x.

The slope of FSD cumulative mileage has gone down in the past, so Tesla doesn't seem to be chasing this metric nor forced to do something to keep the rate up. Although hands free nag with Vision attention monitoring might help increase usage of FSD.

Is there a reason why they can't poll the entire fleet for video training data, even those with just basic AP?

 

[Twiglett]

Is there a reason why they can't poll the entire fleet for video training data, even those with just basic AP?

I thought they did already, isn't AP running shadow even when it's turned off?

 

[stopcrazypp]

Is there a reason why they can't poll the entire fleet for video training data, even those with just basic AP?

My understanding is they already do if you opted in to data sharing?

 

[fasteddie7]

My understanding is they already do if you opted in to data sharing?

I know mine did. I ordered ultra red day one and when it arrived FsD wasn’t availble yet and my car uploaded a ton after every drive. I can’t imagine the 20-40GB wasn’t video footage.

 

[AlanSubie4Life]

With additional training, end-to-end should predict higher action probabilities of taking the left fork for each of these frames and lower probabilities for taking the wrong lanes such that the correct left fork would be the most likely action without indecision of the blue path visualization.

We’ll see.

When it is perfectly trained, what does it do when it has two exactly equally correct scenarios? That is what I am curious about.

to make sure it doesn't override new inputs such as detecting a red light runner

Yes of course, as addressed in my post. Need to be able to change.

As in the mileage rate will decrease as people stop using FSD free trial?

Yes.

The slope of FSD cumulative mileage has gone down in the past

Really? Significantly? As much as it will likely do without further extension of trial?

It's possible of course that they don't care but the initial datapoints looks like a big jump in usage. If they just don't publish the stats it'll probably be fine, that will likely be the approach if they don't like the way it looks.

 

[Ben W]

I didn’t think actual smart summon was part of 12.4. For some reason I assumed it was 12.5

Prediction: 12.4 will have it, but limited to driving in the forward direction only. 12.5 will be ASS backwards.

 

[kabin]

To judge v12.4 after 1 video is silly. You should know better. We have no idea how reliable the post is to begin with and we have no idea if the version is one the public will even see. Patience Grasshopper.

Come on now. We know the trend hasn't been friendly for young FSD grasshopper. There isn't gonna be a quantum jump with v12.4.

 

[Dewg]

When it is perfectly trained, what does it do when it has two exactly equally correct scenarios? That is what I am curious about.

Out of random curiosity, what would a human do in that scenario?

 

[AlanSubie4Life]

Come on now. We know the trend hasn't been friendly for young FSD grasshopper. There isn't gonna be a quantum jump with v12.4.

The dried remains of the young grasshopper are probably causing the intermittent wipes!

Out of random curiosity, what would a human do in that scenario?

I think they'd pick one or the other most of the time and stick to it! You just decide and then go. Obviously there are situations where you'd wobble, but that's usually because the inputs change (not what we're discussing here). Humans are pretty good at deciding when things make little difference and you just need to go with a plan, any plan, but exactly one plan.

 

[KLam]

That seems like it might work, I will give it a shot. I've been trying to use the ridge between the directional buttons and go up or down from that, but it takes me too long to find the ridge and be clear that I've found it. It's easier to find those top and bottom edges of that part of the yoke. Thanks for the tip.

You might already found out that you can steer by hooking your index fingers on the top edge notches. It's almost like just two fingers driving. If you do that, your left thumb will naturally fall on the "right" button.

Another thing I do when trying to touch the screen buttons, I hold the bottom/top edge with my fingers and use the thumb.

 

[PianoAl]

This is off topic but been meaning to ask. Has anyone else noticed the shortcut for navigate to Work via swiping on the navigation search bar no longer works? Since about 2 months, BOTH vertical and horizontal swipes set the destination to Home. It used to be that one of those would set the destination to Work.

The logic seems to be:

If Home, set destination to Work, else set destination to Home.

Perhaps the car is not realizing it is at Home. That is, GPS is off, map data is off, or you inadvertently reset the home address.

Another possibility, you have selected a different profile unintentionally.

 

[SidetrackedSue]

Wait, FSD changed lanes and it was halfway onto the shoulder before you "regained control"????
I've been using FSD since v10 and Autopilot before that and never once has my car made it even over the line without me stopping it. Never once have I needed to "regain control" - because I understand that I must always be in control.

For range anxiety just follow Tesla's advice and switch to percentage instead of range. The range indicator is a guess based on a moderately informed list of criteria. Just like a stock picks and nobody expects them to be accurate.

Love how you interpret my "Thankfully we were already in the right lane and so moved partly onto the shoulder before control was regained." into "FSD changed lanes and it was halfway onto the shoulder"

The torque the driver provides was enough to pull the car partly onto the shoulder before steering the car back into its lane. Had the driver been driving without hands on the wheel, the move towards the shoulder would not have happened. Which actually seems like a flaw in the "driver attention monitoring" system since had the left hand been providing the torque the car would have pulled towards or into the left lane. How far the car moved towards the other lane or shoulder depends on the mood of FSD that day and whether it is hugging the centre line or the right line of the lane. Lately I've been pleased FSD has been traveling to the right side of the lane but I didn't notice if that is still true for 12.3.6. If it was, that would explain how quickly the car pulled onto the shoulder.

Remember the car was simultaneously slowing dramatically (since the system aborting FSD also engaged the regenerative braking.) You'll have to forgive the driver for not responding fast enough to meet your definition of "in control" when faced with a car pulling towards the shoulder while dramatically losing speed while a loud alarm and red flashing warning is happening. Steering straight came first, then returning to speed of traffic flow while moving fully back into the lane, and finally listening to me read the message to him from the screen.

If your definition of control is always both hands on the wheel, looking straight ahead and foot fully on the go-pedal to react to any sudden drop of speed, then you'll have to tell me how to avoid the problem I faced for the first versions of FSD where my two hands on the wheel and eyes watching the road ahead kept triggering the "apply torque" warning (and me missing it until the audio warning because I wasn't staring at the screen.) It was this forum that taught me to apply the torque with only one hand, a level of 'supervision' or 'control' I'm not comfortable with but am required to do because tesla was too cheap to put a sensor in the steering wheel.

As for the recommendation of only using percentage for range, that is what we do and have done from the first year of ownership. Which introduces a new problem, leaving for a long drive with only 80 or 90% battery no longer offers me the potential of 400 - 450km of range because the range the battery offers is now roughly 450km. And that's theoretical range, not the 80% efficiency that my Teslafi account shows the car gets on summer highway road trips.

The tesla is supposed to be taking weather, road speeds, topography AND the state of the battery into its calculations for how much I'll have at the end of the trip. Yet somehow, when I take the route it sets out for me, a route that has been taken a dozen times before, the car is incapable of accurately estimating the range I'll have left when I arrive at my destination. And since the only way to charge the car once I'm there is L1, the car is incapable of calculating accurately if I could even make it to an L2 or L3 charger should my L1 not work. I actually know it can in warm weather because Teslafi has tracked that part of the trip for me multiple times and it averages 15% of my charge and at the cottage the car still has 25% battery, which is not the 40% it told me it would have but enough that I know I shouldn't be stranded should my charging cable fail. But without Teslafi, I would be clueless as to if the car was going to leave me stranded at the side of the road or not it the cable wasn't working. And I can do these calculations for myself with this drive, but put me into a different "charging desert" I would have super range anxiety based on my experience up until now.

 

[Dewg]

The dried remains of the young grasshopper are probably causing the intermittent wipes!


I think they'd pick one or the other most of the time and stick to it! You just decide and then go. Obviously there are situations where you'd wobble, but that's usually because the inputs change (not what we're discussing here). Humans are pretty good at deciding when things make little difference and you just need to go with a plan, any plan, but exactly one plan.

Humans have advanced generalized intelligence, where Tesla has highly specialized, task specific intelligence. Not something that can be compared, obviously.

The closest comparison might be to an infant who can't make up its mind on selecting a toy. But that's even a far stretch.

The planner should smooth out with more training and perhaps tweaks to map/navigation data weights.

 

[tmoz]

I can’t believe you made me give you a thumbs down. Colonoscopies are magical!1!! Granted the cleansing takes patience, but the drugs are heavenly and the encouraged methane fart fest after is absolutely glorious. Nowhere else in this world can a man carry on a conversation while simultaneously spewing out audible metric sht tons of air from his sphincter, and have the person you’re conversations with be totally normal like nothing is happening.

I’d do it weekly if they let me

You could just buy one of those devices they use to pump you full of air. That can take a day to get rid of all of that air. I've got to do a virtual one because the doc had trouble making the last turn with the camera. This time I'll be awake and see what I've been missing.

 

[tmoz]

Supposed 12.4 leak from an employee:

Or perhaps an ex-employee by now.

 

[tmoz]

I also wish they would only compare FSD usage or Autopilot usage with miles driven on the same road (or same difficulty level).
The bias we worry about is that FSD or Autopilot is only enabled on the easy roads, while when it is off, the easy roads plus the hard roads are included.
I'd like to see that data to see if they still draw the same conclusion.