Is my Model 3 Autopilot learning?

[MikeyC]

There's one stretch of road that I always have to travel to get to the civilized world from my house: a 5 mile stretch of 60 MPH country road that is 2 lanes, 2 way traffic. Been using Autopilot on that road with my hand on the wheel paying close attention. The first two times I went by one crossroad, Autopilot started to veer into the left turn lane to turn into that road. I immediately corrected by turning the wheel right to keep it in the traffic lane and by doing so, got the ding telling me Autopilot was canceled. After correcting it twice, now when I go by that same spot (the third and fourth time), it navigates that part of the road perfectly, ignoring that left turn lane and keeping in the main lane of traffic. Same conditions, about the same time of day. Light traffic (a stray car once in a while) all times.

My question is, did Autopilot "learn" that due to me having to correct it twice? Can it learn that fast? My car is only a week old and has not received any updates but I suppose it is possible that my car uploaded something to the existing AP "database"? Did the fleet learn from those instances? I have a suspicion the answer is no and there must be some other reason because there are a ton of Teslas around here and surely someone has driven that section of road on AP and had to do the same thing?

Just wondering.

Mike

 

[cdmcleod]

I have a similar problem on a stretch of NB Hwy 1 near Moss Landing, CA. I have passed that area at least 5 or 6 times with AP activated and the car always tries to steer into the left turn lane before I cancel by manually steering to cancel AP. So far the car has not "learned" not to do that.

 

[LionXng]

Machines can't learn. Don't let them tell you otherwise even if they have learned to be very convincing!

 

[EVRider-FL]

When you intervene while using AP, data is sent to Tesla for examination. It’s possible Tesla did something to resolve the issue, but it’s hard to know that for certain. The car doesn’t learn on its own.

 

[ScoobyDoo82]

When you intervene while using AP, data is sent to Tesla for examination. It’s possible Tesla did something to resolve the issue, but it’s hard to know that for certain. The car doesn’t learn on its own.

The car doesn’t... but the software running in the car does.

 

[EVRider-FL]

The car doesn’t... but the software running in the car does.

Are you sure? I thought the NN was improved by training, but it doesn't train itself (yet).

 

[ScoobyDoo82]

Are you sure? I thought the NN was improved by training, but it doesn't train itself (yet).

The NN is why your car was better on the 3rd and 4th instance that was mentioned. The Tesla itself didn’t have a “aha” moment but the learning process of the NN did. If a completely different Tesla drove the same road it too would have the corrected course because the previous Tesla provided 3-4 scenarios where it confirmed the improvement and then implemented it system wide.

 

[Silicon Desert]

The NN is why your car was better on the 3rd and 4th instance that was mentioned. The Tesla itself didn’t have a “aha” moment but the learning process of the NN did. If a completely different Tesla drove the same road it too would have the corrected course because the previous Tesla provided 3-4 scenarios where it confirmed the improvement and then implemented it system wide.

I wish my wife could do that. She keeps running over the same big pot hole in her car every day

 

[EVRider-FL]

The NN is why your car was better on the 3rd and 4th instance that was mentioned. The Tesla itself didn’t have a “aha” moment but the learning process of the NN did. If a completely different Tesla drove the same road it too would have the corrected course because the previous Tesla provided 3-4 scenarios where it confirmed the improvement and then implemented it system wide.

I would like to learn more about this. Where are you getting your information? I’ve listened to some of the Andrej Karpathy presentations and read other articles, but it’s never been clear (to me) how much the in-car software can improve without getting new data from the mother ship (not necessarily new software).

 

[gearchruncher]

Safety critical systems do not learn as you use them. It could easily "learn" the wrong thing, and the manufacturer would have no way to control the quality of the system in every vehicle. Many people have looked at the Tesla code, and nothing about the NN is learning as you drive. Plus. the NN is just for vision and identifying things like lane lines and signs, not rules of the road. The rules are all hard coded.

It reacts differently because it has different inputs. Different light angles, different speeds, different cars ahead. The fact it reacts differently tells you the "learning" part still has a lot to learn.

 

[ScoobyDoo82]

I would like to learn more about this. Where are you getting your information? I’ve listened to some of the Andrej Karpathy presentations and read other articles, but it’s never been clear (to me) how much the in-car software can improve without getting new data from the mother ship (not necessarily new software).

I saw it in an interview with the old guy who does vehicle tear downs sitting with Musk. Menro? I believe was his name. Not sure, I’ll look through my history and send you the link when I get back home.

 

[gearchruncher]

Sandy Munro is a manufacturing expert. He tears down cars mechanically. He has no expertise in what is going on inside an AI computer.
People often mistake "The Tesla Autopilot is always learning" to mean "My specific car is learning from what it observes me doing." What Tesla means is that they are always collecting new data, which they feed into their machine learning algorithms as part of AP development, and then they produce a new software package that is rolled out to cars.
Many people have looked at the actual code in the AP computer. No real time "learning" (updates to the Neural Net) is done. It only changes when you get a software update (which Tesla can push in the background if they really want to)

 

[ScoobyDoo82]

Sandy Munro is a manufacturing expert. He tears down cars mechanically. He has no expertise in what is going on inside an AI computer.
People often mistake "The Tesla Autopilot is always learning" to mean "My specific car is learning from what it observes me doing." What Tesla means is that they are always collecting new data, which they feed into their machine learning algorithms as part of AP development, and then they produce a new software package that is rolled out to cars.
Many people have looked at the actual code in the AP computer. No real time "learning" (updates to the Neural Net) is done. It only changes when you get a software update (which Tesla can push in the background if they really want to)

Well, that’s not what him and Elon were discussing. He said the complete opposite.

I understand what you’re saying. But I don’t believe that AP only improves after an update. My wife and I have both noticed problematic situations improve after a half dozen whoopsies. I would chalk it up to coincidence but the car hasn’t made the same mistake again. It literally has behaved better after it “misbehaved” multiple times before. This is all while still on the same software version.
Hackers may have looked into the code but that doesn’t rule out the idea that it doesn’t exist. Maybe the experts overlooked something, or saw it but didn’t understand it.
Haven’t you encountered what I mentioned above? A lot of other people have experienced what I’m saying. Very interesting.

 

[gearchruncher]

Well, that’s not what him and Elon were discussing. He said the complete opposite.

If you trust what Elon says, you would have had FSD in mid 2017. He's a salesman with very little need to stick to facts. He's also often using things he hopes to be true in the future as if they are in existence now.

Here's Greentheonly, a very trusted member that has actually looked in the code and has seen that it does not learn:

https://twitter.com/x/status/1320551443728896000

If you really think about it, learning in the car is pretty pointless. It might learn the wrong thing, which is bad. If it learns the right thing, Tesla should distribute it to everyone. Given the 2-4 week cadence of releases, there isn't much to be gained from individual cars "learning" locally, and there is a lot to be lost (like crashes from local bad learnings).

 

[ScoobyDoo82]

If you trust what Elon says, you would have had FSD in mid 2017. He's a salesman with very little need to stick to facts. He's also often using things he hopes to be true in the future as if they are in existence now.

Here's Greentheonly, a very trusted member that has actually looked in the code and has seen that it does not learn:

https://twitter.com/x/status/1320551443728896000

Lol, I’m no Elon fanboy and don’t get excited by his “time projections”. However I also have no reason to believe everything a hacker says. Using that talent for exploring code is great and all, but in order to get that good means you had to start where most hackers do - doing nefarious things. Not saying he’s a bad guy, but for god’s sake you just compared a billionaire who wants to make the Earth run on sustainable energy to a “highly respected” hacker. But, I see your point of view too and it’s very interesting.

 

[EVRider-FL]

Well, that’s not what him and Elon were discussing. He said the complete opposite.

I understand what you’re saying. But I don’t believe that AP only improves after an update. My wife and I have both noticed problematic situations improve after a half dozen whoopsies. I would chalk it up to coincidence but the car hasn’t made the same mistake again. It literally has behaved better after it “misbehaved” multiple times before. This is all while still on the same software version.
Hackers may have looked into the code but that doesn’t rule out the idea that it doesn’t exist. Maybe the experts overlooked something, or saw it but didn’t understand it.
Haven’t you encountered what I mentioned above? A lot of other people have experienced what I’m saying. Very interesting.

It's possible the car receives (silent) data updates between software updates, which would explain why the car seems to learn from your driving experiences. I don't think anyone outside of Tesla can be certain how the system works if Tesla isn't saying.

 

[gearchruncher]

I don't think anyone outside of Tesla can be certain how the system works if Tesla isn't saying.

Of course they can. In order for the NN to be different, it has to be different. This is stored on disk. It's trivial to prove that the NN did not update as you drove.

Heck, Tesla has debug menus that show the "NN Version" and this doesn't update as you drive.

 

[Microterf]

Lol, I’m no Elon fanboy and don’t get excited by his “time projections”. However I also have no reason to believe everything a hacker says. Using that talent for exploring code is great and all, but in order to get that good means you had to start where most hackers do - doing nefarious things. Not saying he’s a bad guy, but for god’s sake you just compared a billionaire who wants to make the Earth run on sustainable energy to a “highly respected” hacker. But, I see your point of view too and it’s very interesting.

Elon is a hacker.

Btw, they can push updates that you don't know about. So the car can "learn" between updates that you see, but it isn't the computer in the car learning, it's the computer on the car was pushed information thay you didn't know about that taught it.

 

[gearchruncher]

It only changes when you get a software update (which Tesla can push in the background if they really want to)

It's possible the car receives (silent) data updates between software updates, which would explain why the car seems to learn from your driving experiences.

Btw, they can push updates that you don't know about.

Seems like we all agree what is happening here

 

[ScoobyDoo82]

Of course they can. In order for the NN to be different, it has to be different. This is stored on disk. It's trivial to prove that the NN did not update as you drove.

Heck, Tesla has debug menus that show the "NN Version" and this doesn't update as you drive.

What’s your deal man? Give it a break. You might be wrong after all.