I've always wondered if Tesla quietly uses safety score or similar to find the "good driver" cohorts. The other possibility is that good drivers stand out in the training set because good drivers all do very similar things, while bad drivers are bad in a wide variety of ways. The good drivers will all be clustered together, while the bad drivers will be scattered all over the place.
With their existing/previous rule-based driving system there is some sort of underlying optimization problem and scoring of preferred paths and it simulated these and had rules to choose/filter them.
They could run that off-line back in the lab (with higher fidelity as well) and find segments & hence people who scored well on this.
Pretty soon the problem is not finding these clips, but finding enough diversity. The 'natural measure' (speaking mathematically) from a uniform selection of the existing fleet is going to have lots of easy driving in California which it does fine at---the challenge is finding and properly training the infinite variety of other stuff. The train set will have to have all the weird stuff heavily overrepresented compared to their natural occurrence probability.
If they're able to fully close the loop and can effectively use self-supervised training and curate the distribution, they will have finally achieved something significant. And this will finally scale well with more data. They'll have better and better filters for pulling out useful train data in both policy and perception and a fleet size which will let them get enough and I hope a training pipeline which can ingest huge datasets with little human labelling needed.
On the chatGPT type of system the amount of unlabelled data (basic text pretraining) is enormous vs the expensive human labelled (for them it's the instruction tuning and reinforcement learning specific examples which are hand crafted by intentional thought and human authorship). The biggest discovery of the LLMs which made them a big deal is how well the relatively small amount of human labelled training can leverage the huge dataset's previous training for good representations, that they can do interesting things with such small instruction datasets. By the way---that was an unexpected empirical discovery---not something that was intentionally engineered. They really shouldn't be that smart, being quite low level and stupid down below.
I'm hoping the same would apply for driving. The 1st phase training is on a giant set of reasonably well-driven human scenes taken automatically as above, then on top of that a fairly small amount of 'instruction tuning' which is examples which make the car do what the nav controller tells it should be doing to achieve a given goal (which the pre-training didn't have as people were just doing their own thing). I assume this is the architecture or where they're going to.
I'm finally hopeful about the FSD, as the giant quantity of raw data theoretically available could finally be useful in a scalable way where limits is only train capacity. Right now they say "trained on 1 million video clips". Why not trained on a billion video clips? Perhaps 5,000 human created instruction tuning they could manage (i.e. a human verifies/programs in the exact relationship between the nav destination, internal state and selects the best preferred driving outcome) but taking advantage of tremendous base datasize which can refine the representations to be whatever is able to video predict scenes and naturally observed driving choices forward in time.
