I have a feeling that AGI is a requirement for FSD. There are enough tail events while driving where you can only make "the right decision" if you actually understand what's going on.
Human drivers fail reliably, in ways that are understandable. Non-AGI artificial intelligence based on a mix of heuristics and dumb statistics (aka 'deep learning') will fail in highly unpredictable and to a human eye completely foolish ways. The day these cars are let on the road in my area is the day I start taking the train.
You can predict the driving patterns of drunk human drivers, or humans that have seizures, cardiac, or other critical health events or just fall asleep at the wheel? Quite impressive.
> I have the feeling this bar disqualifies most human drivers, too.
There are plenty of people who wait for road-side assistance for a flat tire. It's arguable that a FSD system may be totally viable without having to solve every eventuality on the road. If it can stop someplace safe and call for help that's the most a lot of humans can manage.
To my mind there are certainly a lot of unknowns around interactions with people outside the vehicle. A simple scenario like someone directing traffic around an accident seems like a nightmare for FSD.
That's impressive, although a dramatically easier problem than I was picturing.
Imagine nighttime, having to pass an accident, which means driving on the wrong side of the street with other cars beyond the accident waiting, and someone (quite possibly a civilian) waving you on.
Heck, I've been flagged down by a prostitute who wanted to get away from her John, running down the middle lane of an interstate at 2am. That was, admittedly, not a common scenario for a self-driving vehicle to have to deal with.
