I think it’s interesting how many of us are skipping past a very competent L4 deployment. I see this as integral to L5. I think we will see L5 soon after L4 but I see it existing as proof of concept at airports and places that hailing a car isn’t required. After that, there will be a city or two that pilots L5. I think almost all of it will be with Tesla vehicles.
I've been saying for a good while I have low confidence of Tesla delivering L5 on current HW.
It's not redundant enough, and it can't handle bad weather reliably enough. Note for example how often NoA turns itself off, dropping back to basic AP, in anything more than moderate rain because one or more side cameras are blinded by the weather.... or how often folks post here about losing certain FSD or EAP features because one of the external side lenses got some mud or dirt on it.
AI can't do much with input from a blind sensor that has no redundancy and no ability to clear itself.
I've even had basic AP turn itself off in HEAVY rain... and that's only using the stays-dry front cameras behind the wiped windshield plus the radar that can see through rain.
(there's a reason Waymos own working but heavily geofenced L4 system first rolled out someplace it basically never rains or snows)
But Tesla L4? I think they could deliver that basically tomorrow on the highway at least if they get vision solved for the stationary objects problem, and define the ODD to be weather no worse than moderate rain or something. If vision is fully solved they could do the same for city driving easily enough.
Now L5? I think the gap between L4-works in anything but bad weather, and L5 works
anytime a human could drive is bigger than some folks think given the current fleets HW. (remember, they didn't even add a heater to the front radar till the Y so if you woke up to snow/ice, features unavailable)
I'd love to be wrong, but haven't seen any reason to think so yet....
(there is some AI work on correcting camera lens vision for rain obscuration, but it's still pretty limited, and the increase in compute required is MASSIVE if you want to do it in real-time as FSD would have to, and that required compute compounds per camera- I guess in theory if that is found to be "good enough" for L5 then maybe they can swap HW4 if it's powerful enough in there and be good but bad weather is a challenge specifically called out at autonomy day and the one I've seen the least said about solutions for since)
I put a fair amount of value on FSD as user features like city streets come out- both in increasing take buy take rate (and driving monthly revenue via subscriptions eventually too)... but there's a reason even ARK only lists Robotaxis as a 50/50 shot by 2025.
