I think Teslas approach to FSD is absolutely brilliant and the right way to do it. I'm a semi retired-AI coder, FWIW.
The main reason I support their method, is not for anything technical (I'm not familiar enough with radar vs vision and large scale learning models), but because as an investor it is the ONLY method that makes sense.
Lets ask a simple question: Will FSD EVER be achievable, at any stage, by any company?
I think most people would say yes. Its like asking if man will ever set foot on Mars. Might be 10 years, might be 50, but its unlikely we just 'give up' and make no progress. So once you accept that someone WILL achieve it, you have to ask what an FSD future looks like. Everyone except Tesla seems to be focusing on mapping, multiple sensors, and sensor fusion.
Mapping is absolutely useless. The moment there is an accident that blocks a lane, or the moment someone somewhere forgets to update a database after making a change to some street sign or road marking, all your HD maps become dangerous lies. Something that can NEVER be relied on with safety critical operation.
Multiple sensors sound great, until they conflict and you have to make decisions as to which to trust.
And even if you have weather-dependent, time-dependent, HD-mapped multi-sensor FSD... the minute there is some policeman stood in the road directing traffic to turn around and go back, its all utterly, utterly useless, and you are back to needing vision to process this unlikely event and deal with it.
Waymo, Cruise etc are very good systems for impressing journalists on carefully selected routes, and persuading gullible investors to continue to have faith. FSD is the complete opposite. Its very obviously deliberately seeking out the bad edge cases. The tyre bouncing down the highway (which chuck cook already has proven is detected just fine), the group of dumb 20-something men dressed in chicken costumes for a stag-night bash who suddenly charge across the road. The billboard at the roadside that has an image of a car on it, but its NOT a car....and so on...
Tesla's approach to FSD is brilliant, because its the only system that is desperately trying to avoid a local maxima. They dont always succeed, so there will be sensor suite changes, and already we have seen major algorithmn and architecture changes. But they are the only ones even ATTEMPTING a general solution.
If, in the nightmare situation that waymo/cruise get approval to run a true driverless robotaxi service before Tesla does, and even make a small profit from it, it will be living on borrowed time. Once a GENERAL solution is found, globally, then the dozen or so HD-mapped cities in the US will seem like joke.