Question re. FSDb "learning" local roads

[lzolman]

New FSDb "tester" here. Got the update the night before Turkey Day.
So far my experiences with it on local roads (North Reading, MA) have been horrifying, but this isn't about the details.
I've heard for years about how well AP/FSD do in California, and I was under the impression it's because there are so many Teslas per capita in CA compared to elsewhere, so more feedback and maybe now more training data based on CA to improve the system.
Would I be correct in hypothesizing that FSDb does so poorly around my neighborhood because the neural net hasn't had enough opportunity yet to learn the area? I'm hoping that's the case, but someone in a local Tesla FB group said that it is not...rather, CA roads are just easier for FSDb to grok. Which is probably true in and of itself, the roads around here are much narrower and idiosyncratic (and rarely rectilinear). The question in my mind is, will FSDb now rapidly get lots of local data and improve quickly as a result?
I figure I can get an educated answer here. Thanks!

 

[nvx1977]

I've always maintained that if FSDb can handle New England, it can handle anywhere in the US.

Also, FSDb in LA is pretty bad. So "good in CA" is a pretty bad generalization.

That said, FSDb still fails at some very basic stuff, which I would think plagues anyone in any part of the country.

 

[lzolman]

Thanks. Testing with a round trip to my local grocery store, about 4 miles there and 5 back, required about 6 disengagements, mostly due to extreme slowness and one to keep from scaring the bejeezus out of a bicyclist in front of me on a windy road in Harold Parker National Forest. I didn't want to find out how Nameless would have fared trying to pass that guy on the bicycle. It would probably have given both of us a heart attack.

 

[Tam]

...AP/FSD do in California...

Tesla's position is: It's a generalized self-driving system just like a human who doesn't have to have a map but can read the signs anywhere in the world to drive safely.

 

[lzolman]

Tesla's position is: It's a generalized self-driving system just like a human who doesn't have to have a map but can read the signs anywhere in the world to drive safely.

OK. But humans don't have the benefit of a massive data-collecting neural net that has "driven" every road and can instantaneously recall every twist and turn and fixed hazard that anyone has ever encountered in the past. My hope was that this capability would somehow be leveraged by each Tesla in real time while navigating those roads.
Perhaps I expect too much LOL.

 

[Metralla]

When I first got the beta last October I realized one big hurdle is for people to sit behind the wheel while another "entity" decides what's right.

It gets better as you get used to it-which likely seems like it is learning when in fact it is you that is adapting.

I haven't had a disengagement of any type since I got the latest update. We'll see how it goes but you can be hopeful.

 

[RTPEV]

The system is not going to just automatically "learn" local roads by having you and others in your area drive around on local roads. All the learning takes place in a Tesla datacenter, based on videos that are either computer generated or uploaded from the fleet.

That said, it is possible that they happen to pull video from your vehicle (and others in your area), maybe even as an attempt to get local footage from various regions, but more often than not, I think they probably query the fleet for specific situations (roundabouts, yield signs, etc.) or perhaps they select certain types of disengagements. So in a sense, yes, the more drivers in your area that are driving around, the more likely it will be that footage from your area is part of the training set, but it's certainly not like you drive down Harris Road one day and the next day the car then knows how to drive that road.

 

[BitJam]

North Reading, MA) have been horrifying, but this isn't about the details.
I've heard for years about how well AP/FSD do in California, and I was under the impression it's because there are so many Teslas per capita in CA compared to elsewhere, so more feedback and maybe now more training data based on CA to improve the system.

To echo what RTPEV said, last summer Elon tweeted:

In general, we overfit to SF Bay Area.

So it's not all of California, just the San Francisco Bay Area. This was likely due to the location of the Fremont factory rather than the number of Teslas on the road there.

 

[lzolman]

When I first got the beta last October I realized one big hurdle is for people to sit behind the wheel while another "entity" decides what's right.

It gets better as you get used to it-which likely seems like it is learning when in fact it is you that is adapting.

I haven't had a disengagement of any type since I got the latest update. We'll see how it goes but you can be hopeful.

That's saying a lot as you're also in NE! Good to hear.

 

[Ruffles]

I've been testing it for over a year now. I've never had as good of results in my small town as I see in YouTube videos from CA but in the time I've had it, my local roads have improved. With the latest update, I finally got a 0 disengagement drive.

 

[y2khd]

I am also a new tester and I have to say I am disappointed in the latest version of the software. One thing I don't understand is why is the FSD so aggressive. I mean on occasions it accelerates like someone is after you. Coming to a red light it starts slowing down at the very last moment. And how it does not detect speed bumps, I mean the car just flies over the bumps....

 

[Jay S]

I'm sorry to report that FSD Autopilot does not learn anything. It can't, partly because it seldom knows that it has done something wrong. It is not able to even detect many nearby dangers at all. The "neural net" is only in your car. It doesn't connect to other cars or road systems. My 2018 Medel 3 Long Range (now only 266 miles on a full charge) cannot safely drive the car under any circumstances, City streets or highway. It is a pathetic advertising gimmick that should be banned from the roads. My 2019 Model S is even worse! It drives itself as if it just consumed a 5th of Jim Beam. Horrible device installed in an otherwise fine car. I'm the fool, I believed the hype and paid my money for FSD. For both my cars, I've renamed Auto Pilot to be Auto Crash! If you use the features, drive the car as if Auto Pilot is turned off. You may live longer.

 

[spacecoin]

I used to be very bullish on FSD until I did a lot of research into machine learning and AI. It seems to me that many have a belief that this is "intelligence" or some black magic and aren't understanding of the current state of FSD and AI in general. I found this a good, informative read. Perhaps someone else will appreciate it.

Take self-driving, for instance. We’ve invested about $100B in the field over the past 10 years — roughly half of the inflation-adjusted cost of the Apollo program. And we’re now just starting to see fully driverless cars able to handle a controlled subset of all possible driving situations. You can ride in one in SF from Cruise (in private-access beta) or in SF or Phoenix from Waymo (in public access). Other players tend to be far behind on the generality curve. Crucially, these results were not achieved via some kind of “just add more data and scale up the deep learning model” near-free lunch. Deep learning only represents a subset of the system. These results come from years of work-intensive engineering that went into crafting complex systems that encompass millions of lines of human-written code. And we’re still very, very far from generalizing to all locales or all situations.

AI is cognitive automation, not cognitive autonomy

Like the rest of computer science, AI is about making computers do more, not replacing humans.

fchollet.substack.com

 

[lzolman]

I'm sorry to report that FSD Autopilot does not learn anything. It can't, partly because it seldom knows that it has done something wrong. It is not able to even detect many nearby dangers at all. The "neural net" is only in your car. It doesn't connect to other cars or road systems. My 2018 Medel 3 Long Range (now only 266 miles on a full charge) cannot safely drive the car under any circumstances, City streets or highway. It is a pathetic advertising gimmick that should be banned from the roads. My 2019 Model S is even worse! It drives itself as if it just consumed a 5th of Jim Beam. Horrible device installed in an otherwise fine car. I'm the fool, I believed the hype and paid my money for FSD. For both my cars, I've renamed Auto Pilot to be Auto Crash! If you use the features, drive the car as if Auto Pilot is turned off. You may live longer.

Well, YMMV... I used AP for most of my drive from CA back to Boston in May, and to/form Mt. Rushmore in August (the only reason I didn't use it going Boston to CA is that I was enjoying driving Nameless myself too much). There's one day I had to beat a storm bearing down on Wyoming/CO as I made my way through Yellowstone, and ended up driving two days' worth in one day as a result. AP allowed me to remain alert/awake, and the last part was in white-out conditions on I80 approaching Cheyenne...as I learned later, some of the most treacherous driving in the country during a storm (no, I didn't use AP during the storm LOL). Point being, AP worked great on the highway and greatly reduced my fatigue. I only had to supervise. My wife and I did Boston <-> Chicago and Boston <-> Minnesota in her Model 3 four years ago, same success.