(continuing on what I consider a fascinating topic since it's the weekend)
Assuming I understand correctly what you are confirming, I had asked if you you were saying something represented on the display that is understood to be vector-space is evidence of data that was collected by previous vehicles. It seems as if you are saying "yes".
Yet you also just said:
So, I'm unsure if you mean that
some of the vector space data is sourced by previous vehicles in non real-time, and some vectorspace data is collected in real time by the current car. Or if you are saying that some data, such as vector space is collected by previous cars, and other data (non vector space) is collected by the current car.
If the former then that would seem to be at odds with your assertion that vectorspace renderings (such as the creepwall and median box) are necessarily an indicator of having been ground truth sourced by previous vehicles. If the latter that that seems to be different that what was presented at AI day.
At about
16 mins in to AI day Karpathy discusses using multiple cameras to generate a vectorspace map in real time.
And at about
49 mins in to AI day it was said that vector space data collected by the fleet was used only for training labels, and then discarded.
So, it appears that the car does indeed generate real-time vectorspace data, and that data from previous cars is not used by your current car (only that it's NN was trained with that labeled data). Or at least that was the case at AI day.
So I guess I'm confused when I understand you to say that the apparent vectorspace representation of the creep wall and median box are indicators of ground truth sourced by previous cars?