Actually, autonomous driving stacks can get better doing millions of miles in simulation. The teen, not so much.
Have they not already completed many millions of miles in simulation, back in 2018?
To me, it looks like self-diving capabilities have reached a maxima. The resources poured into it have barely been needle movers in objective metrics.
Yeah, but videos aren't objective metrics. For me, an objective metric is "went $FOO miles in $BAR conditions without needing any intervention".
What would you consider an objective metric?
But who decides whether intervention was necessary? This is still too vague.
Irrelevant - until the car has actually gone without an intervention for whatever the median is for human drivers[1], we can't very well claim that it is safe for use.
[1] The average human driver goes hundreds of thousands of kilometers without crashing.