I don't have to watch videos, I do believe progress is impressive.
The problem with FSD is that anecdotes do not mean data. If a camera only approach can reconstruct 99.99% of the scenes, that is still not going to be good enough! I have faith that Tesla can aggregate all the key data to ramp up camera-object identification accuracy as much as possible giving the data and deep RL methods, but how could I possibly be confident its going to get to a certain number of .99999 digits? I work in ML and I would be insane to be so confident in the eventual accuracy of my work.
You have to assume probabilistic outcomes. If the camera based approach is only 99.99% reliable, where does that leave Tesla in terms leveraging the technology? I am more optimistic it means people will pay some more money for L2 (maybe L3) that can handle more roads.
It will NOT allow robotaxis (outside of constrained environments). Thus the financial valuation changes.
I am not saying Tesla will not achieve L4/L5 with cameras. I am saying I would give it a 10% chance. The "expected value" of valuation (say $4000/share if successful) given that would still be much higher than the current stock price, so it is a rationale bet.
But enough with the crazy talk, you crazies