Which presentation did you watch? Karpathy said specifically "it's not a massive amount of data, it's just very well picked data" when talking about how the cars only send data when one of the configured triggers fires.

There’s a large gap between ‘a few frames’ and a massive amount of data, and the amount sent lies somewhere in the middle. Clearly they can’t send all data (nor would they want to) but it seems it is sufficient for significant learning to take place and the examples shown were good quality over at least a few seconds, so hundreds of frames for each example.

No, it's spot on. It's entirely what I said: the car can only deliver a few raw frames, and only in response to particular triggers.

Notice the cherry-picked examples in the presentation. There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.

They showed video in the presentation which was clearly not ‘a few frames’, unless by a few frames you mean seconds of video.

> There is a whole class of problems the field cars can never help with, since they lack the dead-reckoning sensor setup and precise odometry a development car would have.

Can you give an example? I'm curious what kind of triggers strictly require lab-calibrated hardware.