Good point.
As I understand it Douma is being a classic academic here, using a very hard definition of proof and rigor re. FSD. He modifies that by saying that Tesla will most likely solve FSD and recent events are a strong indicator that they are moving fast in the right direction.
AFAIK the hard variant of level 5 self driving, the upmost level, is that the car can handle whatever comes as good or better than expert drivers - come hell or high water.
But when you get down to the details that is maybe not as clear as all that. We know that given sufficiently difficult situations, even very good human drivers give up. But how do they correctly decide that? So, would an FSD/self driven car with the ability to give up in extreme conditions still constitute FSD driving? Perhaps.
The criteria for giving up is a big grey area: If sufficiently weak effort is accepted, then most self-driving could, in the limit, be said to be self-driving: The may 'choose' to give up 50% up the time. That definition is obviously ridiculous.
It is not as obvious when you get closer to 100%. Is 95 OK? Is 99% good enough? Well, giving up one trip per 100 is much better than getting into an accident for sure.
You could argue that 99% is not true self driving: If you use, nominally speaking, self-driving car to transport your underage kids 4 times per day, then off course avoiding accidents is key. But, you would probably be annoyed if once a month or more, at an unpredictable time and place, they get stuck in an unmoving car which has given up.
The central control could solve that. And maybe that is exactly what will happen. But costly. And comes with its own set of legal problems. (When accidents occur are the bad sensors to blame, or the remote operator/driver, or the networking team, or ..?)
Re. Waymo: You could argue that Waymo have solved a subset of self driving which depends on hi-definition maps. Acc. to Douma, lidar is great at location but does nothing for moving object/people so they have to depend on vision here. How good is WAYMOs vision? I don't know.
Re. scale-out, both having and maintaining hi-def maps and having lidar equipment on the cars and having human intervention in control centers are 3 negative scaling factor you may not have if you vision-only software is sufficiently good. Yes, they can scale that, but slowly and expensively.
Re. remote drivers.
If centralized 'remote drivers' are an short-term to intermediate solution until the 'march of nines' is completed or in case the last 1-0.00.1 percent of FSD is really, really, really difficult and will take 20 years to solve then that is in itself raise interesting perspectives for StarLink v Tesla collab.
Note that Elon time and time again has stressed that the StarLink latency/delay should be very low - perhaps lower than all other known communication solutions.
Now, how big a latency would you like a remote driver to have when driving around your underage children, handicapped spouse, old grandmother or good friends? What is low enough except the lowest delay possible limited by only by physics?
Interesting how Elons puzzle pieces come together over time/decades...