FSD Subscription

[Knightshade]

I thought user disengagements during AP use were within the realm of campaign triggers. A user without AP won't be able to trigger that.

He explains disengagement reports in the thread-

Another one of this kind is a "disengagement report". It sounds exciting, but is actually rather boring. The data there is coordinates, how did you disengage, speed, heading and time

no pictures in those fixed disengagement reports so they are really small (under 1kb).

That's not uniquely useful since they don't know WHY the person disengaged, and it sends no visual data back to Tesla at all.

Most data collection are campaigns to collect visual data to send back and have humans manually label, then use to train the NNs.

For example if the system were bad at recognizing bikes it would send a campaign out "Send me back pictures of anything you remotely suspect could be a bike"- and it will... then humans label the objects and the NN trains on them to get better at correctly recognizing bikes.

The cars can all do that regardless of having FSD or not.

And to be honest, I think it's fair to assume that Tesla's data collection methodology is going to change quite a bit once they incorporate dojo.

Dojo AFAIK is to replace the humans for labelling the same data they collect today- just do it a lot faster.

So they might increase the AMOUNT of data they collect... but I don't know why it'd change what or how they collect it?

(and Dojo's at least another year away)

 

[Fernand]

Reports of GIGabytes of data being uploaded from
beta FSD cars. Then today they got another download,
2020.40.8.11. It's happening faster and faster.

The protocol, including how much data of what type
is retrieved, can be changed at any point, so all of
the "normal behavior" descriptions are obsoleted on
the fly. And short of sitting with the developers 24x7,
there's no way to know what data is flowing up and
down with what car and when.

It's logical to suppose that at some point the servers
will be requesting more and more data in response to
automatic detection of edge cases, not just "anomalies".

It's at the point where the neural networks won't be
spoon fed, they will "go out" and hunt with their own
massive appetites. Exponential. And not necessarily
only for the data that the programmers might expect.

A cautionary tale comes to mind. About how social
networks' algorithms "discovered" that spreading conflict
is the best way to increase the humans' engagement,
that being what they were tasked with. That in turn got
us to where we are, with any seeded insane conspiracy
theories being pushed up the selection ladder, to elicit
clicks, and a more profitable platform for advertising.
The moral of the the story might be "be careful what
you ask for".
.

 

[Phlier]

....The moral of the the story might be "be careful what
you ask for".

I'm going to hell, anyway. Might as well have my car drive me there.

 

[Knightshade]

Reports of GIGabytes of data being uploaded from
beta FSD cars. Then today they got another download,
2020.40.8.11. It's happening faster and faster.

That's not surprising- they suddenly need to train the NNs against a LOT more situations.

Simply sending out a ton more campaigns of exactly the type they always have would explain that very easily.

"Send us video of every time you go through a roundabout" for example- something they'd not have done originally since the system wasn't intended to go through those at all.

Apply for a ton of other similar, now-new, cases.

The other reason the amount would be much larger is the car used to only think in single-frame single-camera pictures.

Now it's combining video from all 8 cameras in real time to get that much more advanced view of what is going on- so I'd absolutely expect the outputs to be a lot larger dataset to upload even if they're triggering in exactly the same kinds of ways.

 

[CertLive]

That's not surprising- they suddenly need to train the NNs against a LOT more situations.

Simply sending out a ton more campaigns of exactly the type they always have would explain that very easily.

"Send us video of every time you go through a roundabout" for example- something they'd not have done originally since the system wasn't intended to go through those at all.

Apply for a ton of other similar, now-new, cases.

The other reason the amount would be much larger is the car used to only think in single-frame single-camera pictures.

Now it's combining video from all 8 cameras in real time to get that much more advanced view of what is going on- so I'd absolutely expect the outputs to be a lot larger dataset to upload even if they're triggering in exactly the same kinds of ways.

Something about that post was exciting! 8 Camera's in real-time... maybe the turning point in this whole FSD world.

 

[Phlier]

Again you appear to think there's some functionality that does not exist.

The car has no idea when it makes a "mistake"

I'd encourage you to read the link I posted- it explains how, why, when, and what data Tesla collects- and why the driver having FSD or not makes basically 0 difference to their ability to collect it.

I think what he's referring to is a situation like the car saying (if the car could talk), "I was in FSD. Driver intervened. Not sure what I did wrong. Sending video." Which, of course, would require the car to be equipped with FSD in order for the trigger to exist.

Edit: I should learn to read the entire thread before posting. I see this has already been discussed. Nothing to see here... move along....