Please check me on this. This video appears to confirm what I said about the immense calculation needed to train current AI models, such as FSD. The video then speculates on future, as yet unproven ideas about "liquid" and "spiking" models, which, it suggests might be able to learn after they are trained. Promising, yes, but FSD does not incorporate either of these hypothetical technologies. And the video does not give any clue as to how the liquid model incorporates any experience after training into the parameters of the "core" of the "liquid" model called the "reservoir". I'm not calling BS, but this video sounds a lot like pseudo-science AI. That is to say there may be some reality behind this, but this video does not illuminate.
Did I miss something?
I do hope that some sort of learning in the Tesla cars is eventually implemented, beyond "Home" and "Work" and other favorite destinations and manual settings. Things like how much acceleration after a stop is comfortable for this driver, where the pot holes in my neighborhood are so I don't have to disengage every time I get near home, and which turns should be taken slower so as not to veer over the lane divider line.
So far, Telsa's approach is one style, one set of learning fits all cars. This has never been the reality, where each driver has a different style as well as a different family, pets and friends as critics of their driving style. Safe is necessary, but comfortable for each family is the metric critical for market acceptance, which metric is different for each family. Learning in each car may be necessary, in which case, we certainly still have quite a way to go.