The catastrophe of FSD and erosion of trust in Tesla

[Dewg]

(enter the "but you dont understand..the FSD is a different stack and different ML/AI/DOJO than smart summon. Its easier to land a rocket right side up than to not have the code drive up onto a curb" fanbois, most likely from the Bay area or Portland).

Nah, no need for those. What it did was totally wrong. It might be helpful to think of the neural nets like a child learning to drive. A teenager could have 6 months of driving under their belt, and still accidentally hit the accelerator instead of the brake and ram the car into a fence, or back into a closed garage door. In your case the NN's didn't get it right. Hopefully it learns from it and gets better in the future, just like how the teenager hopefully learns not to back into a closed garage door again. Doesn't mean the kid won't back into another garage door, but we can hope.

 

[Daniel in SD]

Nah, no need for those. What it did was totally wrong. It might be helpful to think of the neural nets like a child learning to drive. A teenager could have 6 months of driving under their belt, and still accidentally hit the accelerator instead of the brake and ram the car into a fence, or back into a closed garage door. In your case the NN's didn't get it right. Hopefully it learns from it and gets better in the future, just like how the teenager hopefully learns not to back into a closed garage door again. Doesn't mean the kid won't back into another garage door, but we can hope.

Exactly, this explains why Smart Summon is so much better than it was two and a half years ago.

 

[CyberGus]

Driving with beta does not really train NN. Training is done internally at Tesla. While the training data may include segments recorded by drivers at disengagement points as input, I doubt the beta drivers driving behavior is used for training at all.

Neural networks are trained with input. Lots and lots of input. Tesla's fleet of millions is their greatest advantage over other self-driving products.

Obviously the cars are not streaming all the video to the mothership in realtime, lol. When not engaged, FSD is operating in "Shadow Mode" where it compares it's own decision-making with the actual decisions made by the driver. When the human does something unexpected, the event is saved for upload and analysis.

The same is true for "labeling". The FSD computer categorizes objects in view (car, truck, pedestrian, trash can, etc.) but sometimes will come across a new object that it doesn't understand. That clip will be reviewed and categorized by a human, so that the NN will understand it.

Watch how Tesla trains its neural networks for self-driving in 10 minutes

Andrej Karpathy, Tesla’s head of AI and computer vision, gave an interesting talk to get into how Tesla trains its...

electrek.co

 

[CyberGus]

The short answer is that we don't know. We don't know how Tesla collects and processes data from the FSD testers. When I press the record button when the car does something incorrectly, I know that a short clip and additional data is sent to Tesla (according to the language we all read when we entered Beta), but we don't know what that is, or how much it is. Since the NN's convert visual data into other forms (I love the term "bag of points"), perhaps what's being sent to Tesla is the NN's view of the world and GPS data and not the raw video. They can take those data sets and feed them into a simulator (given GPS data) and see what the car did, then smooth it out in the simulator and feed the NN training data back to the cars. Again, I'm guessing - and everyone here is just guessing.

There are a number of threads here that discuss the enormous amount of data that Teslas upload via wifi. Presumably it is video clips and diagnostic data for NN training.

My car has uploaded 22GB over Wi-Fi in the past few days. Any idea what's up?

I haven't done anything out of the ordinary. Maybe somebody at HQ likes my MP3 collection and ripped themselves a copy?

teslamotorsclub.com

 

[t3sl4drvr]

My guess is

Choosing situations to upload

• When user explicitly reports
• When user disengages
• When a specific situation Tesla team is has broadcasted to be looking for is detected (ex. "record railroad crossings")
• On specific conditions while not engages, when shadow NN would break but human does not

What to upload

• 10-30sec of MPEG buffered from all cameras
• GEO data
• NN outputs
• Driver outputs
• Bunch of other context data

How used

• Manually reviewed and only qualifying situations included in training dataset
• Driver outputs are not used at all (only for detecting that situation is interesting)
• Compression-artifacts are ignored or smoother and an approximation of original camera data is simulated
• In critically important situations, virtual 3D representation of the environment is created (manually) to allow cars to drive many passes in various ways in the situation

This gives a huge advantage to Tesla by surfacing and collecting tons of edge cases in a cost effective manner. That said, this would not mean that driver behavior would be training NN behavior.

 

[Terminator857]

Exactly, this explains why Smart Summon is so much better than it was two and a half years ago.

Yeah, it has blown my mind how bad it is, just like Elon said it would.

 

[DrDabbles]

We don't know how Tesla collects and processes data from the FSD testers.

Why do you keep saying we don't know things that we absolutely do know? Tesla uses campaigns to collect disengagement data, still. The problem is their ability to collect meaningful detail about the disengagement is zero because users can not give feedback. That's what trained testers are used for in other companies. They annotate the scene, detail what went wrong, collect other relevant data, and submit the data for review and training.

This is why we STILL see users like Chuck Cook on the same exact corner having the same exact problem for two years straight. Just like when Elon says AP and FSD are "superhuman", he gave you a marketing message when he said they are training based on user data. They aren't. I'm sorry to be the one to tell you this, but they just flat out are not using customer data the way Elon claimed initially.

 

[Dewg]

Sorry Dabbles, but we don't know. No one from Tesla has said "When you press the record button during FSD Beta here is what happens:" and then says exactly what data from the systems is sent, and how Tesla processes that data and learns from it. t3sl4drvr had a nice summary above that seems logical, but it's still a guess. Tesla's FSD and AP programs are not open-source, so they will likely never tell us how they do things - that would be a massive breach of trade-secrets and let competition duplicate and catch up fast. We may get hints, and piece together information from tweets, news articles, and release notes, to try to form a picture of how everything works, but in the end it's still just a guess.

If you, for some reason, know things for certain that the rest of us don't, please let us know how the system works, where the flaws are, and how to fix them. I'm sure Tesla would pay you 6 figures to walk in and solve the problems.

 

[drtimhill]

6 THOUSAND views for this thread
1 thousand for this one:

The success of FSD and my increase for trust in Tesla

I paid 10 grand and now my already great Model Y now drives itself through my town. It slows down for speed bumps, handles four ways on its own, swerves around bikers and parked cars who's doors are open. It handles the 5 and the 405 quite well too! Obviously it's not perfect and can do the...

teslamotorsclub.com

AI/ML analysis by DOJO (without LIDAR/RADAR) shows that the stats would be reversed if FSD actually worked as promised by the CEO.

Yeah lots of people stop to gawk at car crashes and train wrecks too .. doesnt mean they approve of them.

 

[_Redshift_]

The auto wipers on my Model 3 are freakin fantastic, all thanks to the NN that senses much more than rain..

For example, in my garage. They start wiping furiously sometimes. Now, I am not clever enough to know why, but I trust that the NN does. And I trust that it has learnt things no normal human could have learnt.

 

[nvx1977]

6 THOUSAND views for this thread
1 thousand for this one:

So basically this proves people love drama. Thanks for contributing, i guess

 

[Dewg]

The auto wipers on my Model 3 are freakin fantastic, all thanks to the NN that senses much more than rain..

For example, in my garage. They start wiping furiously sometimes. Now, I am not clever enough to know why, but I trust that the NN does. And I trust that it has learnt things no normal human could have learnt.

My favorite is when I wash the car, and painstakingly dry it to a showroom shine, then I get into the car and engage FSD, and it says "Cleaning Windshield" and then sprays fluid and engages the wipers. LOL - that window was spotless. Oh well. I started to learn that I don't engage FSD right after a wash. I let it sit for a few hours before engaging FSD. I have no idea why it happens - I can guess logically that there might be a sensor somewhere that senses moisture and that part is wet after a wash?

 

[S4WRXTTCS]

The auto wipers on my Model 3 are freakin fantastic,

I'm glad they worked during your one rain event this year.

 

[daktari]

Sorry Dabbles, but we don't know. No one from Tesla has said "When you press the record button during FSD Beta here is what happens:" and then says exactly what data from the systems is sent, and how Tesla processes that data and learns from it. t3sl4drvr had a nice summary above that seems logical, but it's still a guess. Tesla's FSD and AP programs are not open-source, so they will likely never tell us how they do things - that would be a massive breach of trade-secrets and let competition duplicate and catch up fast. We may get hints, and piece together information from tweets, news articles, and release notes, to try to form a picture of how everything works, but in the end it's still just a guess.

If you, for some reason, know things for certain that the rest of us don't, please let us know how the system works, where the flaws are, and how to fix them. I'm sure Tesla would pay you 6 figures to walk in and solve the problems.

Each press of the "FSD report button" will increase the sucker counter at a big screen in Theslanos hq. Every time the sucker counter passes a new 100, it pings the gongong and a robotic voice says "superhuman", "blow your mind", "biggest assert appreciation in history" and similar quotes
randomly and everyone laughs out loud. In yhe lunch they play bingo with these words.

 

[S4WRXTTCS]

So basically this proves people love drama. Thanks for contributing, i guess

This is why I think its impossible for social media companies like facebook, IG, etc to really provide a healthy platform.

The only way for a human to win is not to play the game. That a lot of times its really only ourselves that we have to blame.

If Moderation is heavy handed then people don't bother having any tough conversation. Like today on a FB group I commented about the great Rivian Giveaway of 2022 where Rivian is selling vehicles below the true market value. Not by just a little bit, but around $20K or more per vehicle. I spent like 20 minutes writing the comment, and when I went to post it the original post was deleted. Maybe the mods deleted it or maybe the poster deleted it. It doesn't really matter as that was clearly a big waste of time despite how interesting I thought the situation was.

If Moderation is light then people attack each other and its just a blood bath.

If Moderation is just right like on TMC then people get addicted to the site like me. Haha

 

[Dewg]

I never understood the blood-bath concept, or why people attack others, sometimes viciously, over issues they are powerless to resolve. Society would be so much better if people worked together to the betterment of society and themselves. Instead of saying "I'm right and you're wrong", people should find common ground and try to stay positive. Also, it's helpful to step back and look at the big picture. Look at trends over time instead of focusing on the moment. My mom used to tell me, when I was a kid, "there are people starving in [insert region here - usually Africa or China]". It made me think that no matter what problems I'm having, my life is so much better off than many people on Earth. I still struggle sometimes with that, and have to remind myself that lesson.

 

[_Redshift_]

I'm glad they worked during your one rain event this year.

Yeah, we haven’t had too much rain this year otherwise I would have experienced and appreciated the awesome wipers so much more.

Now, the traditional sensor based wipers on my S are… well, boring. I mean they work only when rain is falling. Not like my 3.

I am thinking of asking Tesla to convert the sensor based wipers on my S to NN based ones if they can.

 

[Sporty]

The auto wipers on my Model 3 are freakin fantastic, all thanks to the NN that senses much more than rain..

For example, in my garage. They start wiping furiously sometimes. Now, I am not clever enough to know why, but I trust that the NN does. And I trust that it has learnt things no normal human could have learnt.

And certainly there are fully zero ways to create a test case for this..

 

[EVNow]

60k testers. Probably 300k miles per day, 9 million miles per month. Hardly any reports of accident.

What a catastrophe FSD has been !

 

[Terminator857]

60k testers. Probably 300k miles per day, 9 million miles per month. Hardly any reports of accident.

What a catastrophe FSD has been !

This isn't about how well people can supervise a car.