Are FSD updates giving diminishing returns?

[cyberos]

This is not a FUD or troll post. I’m a happily committed Cybertruck owner, for better or worse.

I’ve noticed that the latest FSD updates are less noticeable. The first few last year fixed curb hopping, roundabouts and self parking, to name a few.

A credible source believes the lack of Cybertruck purchases provides less data. The Model S/X/Y have millions of cars on the road. That’s plenty of fresh training data for AI updates, especially rare events.

Basically I’m starting to think Cybertruck FSD may be as good as it gets.

Sponsored

 

[Jhodgesatmb]

When we got our Model 3 we bought FSD with the promise of fully autonomous driving just around the corner. Five years later and 2 Model Ys and a Cybertruck later, we still don’t have autonomous driving but we have been happy with FSD and it continues to improve though sometimes it’s a one step forward and two steps back kind of thing.

what I really wanted to say, though, is that when we got our Model 3 there weren’t many Model 3s on FSD but updates came periodically (not regularly).

Much of what the Cybertruck does is the same for every vehicle in the fleet; the difference being the vehicle model. Much of the vehicle model is geometry and camera locations, with the added complexity of SBW and 4WS. All of this can be addressed without a vanguard of Cybertrucks. What we need is a new and improved model.

Hang in there!

 

[hemiarch]

This is not a FUD or troll post. I’m a happily committed Cybertruck owner, for better or worse.

I think that’s the first truck monogamy statement I have ever read. ?

 

[Willinak]

Basically I’m starting to think Cybertruck FSD may be as good as it gets.

Man, I hope not!

 

[SCTesla]

They haven’t had a new major release for any vehicle since thr CT got FSD. All work is going to the Austin FSDU.

This isn't the norm. They will begin sending out meaningful releases again after Austin launches.

 

[JimBuck333]

FSD 13.2.9 came out not even a month ago. Based on the release notes, it should significantly improve the hardware/software package’s ability to capture and process data. That should pay off going forward, when the driving model is configured to take advantage of the increased processing.

 

[JimBuck333]

FSD 13.2.9 came out not even a month ago. Based on the release notes, it should significantly improve the hardware/software package’s ability to capture and process data. That should pay off going forward, when the driving model is configured to take advantage of the increased processing.

 

[riggster]

This is not a FUD or troll post. I’m a happily committed Cybertruck owner, for better or worse.

I’ve noticed that the latest FSD updates are less noticeable. The first few last year fixed curb hopping, roundabouts and self parking, to name a few.

A credible source believes the lack of Cybertruck purchases provides less data. The Model S/X/Y have millions of cars on the road. That’s plenty of fresh training data for AI updates, especially rare events.

Basically I’m starting to think Cybertruck FSD may be as good as it gets.

The lack of cybertruck purchases indeed provides less data than what is available on the Model Y. The amount of time the cybertruck has been out also leads to having less data. That said, why would you think Cybertruck FSD is as good as it is going to get when you argue there is not as much data at this time? They aren't going to stop collecting data for the cybertruck, and more Cybertrucks are being sold all the time. When they have enough data to conduct a native training session with Cybertruck data, of course, it will work much better than tuning Model Y's deep neural networks to work for the Cybertruck. As someone who programs DNNs, I have a lot of confidence that Cybertruck FSD will improve as more data is collected.

 

[cyberos]

The lack of cybertruck purchases indeed provides less data than what is available on the Model Y. The amount of time the cybertruck has been out also leads to having less data. That said, why would you think Cybertruck FSD is as good as it is going to get when you argue there is not as much data at this time? They aren't going to stop collecting data for the cybertruck, and more Cybertrucks are being sold all the time. When they have enough data to conduct a native training session with Cybertruck data, of course, it will work much better than tuning Model Y's deep neural networks to work for the Cybertruck. As someone who programs DNNs, I have a lot of confidence that Cybertruck FSD will improve as more data is collected.

Good point and very reassuring.

My concern is that the volume of data may not be enough, even when collected over the next few years. Cybertruck sales numbers are a lot smaller than the other models. The amount of data needed to train FSD is massive, especially for those weird life-or-death edge cases.

Interestingly, my FSD did a 4-wheel-drift when it blew through an offramp, and then struggled making a left-hand-turn at an intersection. I doubt more people are sliding through turns than driving through intersections. Intersections are just a mess of edge cases that require overwhelming amounts of training days

 

[JDin NFLA]

Why do you think Tesla gave CT owner's free FSD for a year? They need miles to help improve. My FSD has been very good. I don’t have any serious issues. It’s getting better.

 

[riggster]

Good point and very reassuring.

My concern is that the volume of data may not be enough, even when collected over the next few years. Cybertruck sales numbers are a lot smaller than the other models. The amount of data needed to train FSD is massive, especially for those weird life-or-death edge cases.

Interestingly, my FSD did a 4-wheel-drift when it blew through an offramp, and then struggled making a left-hand-turn at an intersection. I doubt more people are sliding through turns than driving through intersections. Intersections are just a mess of edge cases that require overwhelming amounts of training days

Let's do some calculations and make a few assumptions. Let's say the average distance driven by Cybertruck owners is 12,000 miles per year. Let's also assume that the number of Cybertrucks sold every year is 50,000, which is a significant assumption, but it's being made to estimate the number of Cybertruck miles driven annually. You can adjust the numbers to be as pessimistic or optimistic as you care, but the point is that over time, there will be a lot more data available with which to train.

Total miles in year n=50,000×n×12,000=600,000,000×n\text{Total miles in year } n = 50,000 \times n \times 12,000 = 600,000,000 \times n, so
Total miles in year n=50,000×n×12,000=600,000,000×n

 

[Outdoors]

Since AP1 there have always been several steps forward then an occasional step back. Sometimes the steps back seem to take a bit longer to "solve" if ya want to call it. Sometimes it can be something added that many loved that just wasn't right. So it goes back a step and people groan.

I will take literally hundreds of updates and some of those steps back vs. an industry that is just figuring out over the air updates. I also think it keeps people attached to the vehicle/brand. Updates do keep the car/truck up to date and feeling newer.

 

[SCTesla]

FSD 13.2.9 came out not even a month ago. Based on the release notes, it should significantly improve the hardware/software package’s ability to capture and process data. That should pay off going forward, when the driving model is configured to take advantage of the increased processing.

13.2.9 is a bug fixes release. 13.2.X was the first version for thr CT and no one has had a major or minor update since. Only bug fixes.

 

[cyberos]

Let's do some calculations and make a few assumptions. Let's say the average distance driven by Cybertruck owners is 12,000 miles per year. Let's also assume that the number of Cybertrucks sold every year is 50,000, which is a significant assumption, but it's being made to estimate the number of Cybertruck miles driven annually. You can adjust the numbers to be as pessimistic or optimistic as you care, but the point is that over time, there will be a lot more data available with which to train.

Total miles in year n=50,000×n×12,000=600,000,000×n\text{Total miles in year } n = 50,000 \times n \times 12,000 = 600,000,000 \times n, so
Total miles in year n=50,000×n×12,000=600,000,000×n

Woah, I see what you’re saying. With enough time there will be plenty of data. It’ll be wait, but not hopeless

 

[cyberos]

Since AP1 there have always been several steps forward then an occasional step back. Sometimes the steps back seem to take a bit longer to "solve" if ya want to call it. Sometimes it can be something added that many loved that just wasn't right. So it goes back a step and people groan.

I will take literally hundreds of updates and some of those steps back vs. an industry that is just figuring out over the air updates. I also think it keeps people attached to the vehicle/brand. Updates do keep the car/truck up to date and feeling newer.

This was the other thing I was wondering. The state of the auto industry is kind of a joke, when compared to self-driving and Tesla.

I see ads for competitors offering “traffic aware” and “pedestrian aware” features, “AI defined vehicles”, and surfaces coated in screens. None of it even approaches FSD.

I say this as a huge champion of Waymo’s use of LIDAR. Waymo is the dream-made-real, but I can’t buy one (and it’s limited to cities with pre-programmed mappings).

That and I love fully manual ICE cars, passionately. I moved on to EVs for “reasons”, and FSD was a HUGE one

Sponsored