This FSD clip highlights two things that remain the biggest question marks to me with self driving: taking cues from the tiny details, and lack of memory. Humans are so good at this. We can see the faintest “only left turn” paint in the distance and know how to stage the car. Or a tiny little overhead sign. I’m less concerned about this for FSD and think it’s probably a matter of increasing parameter count, but not sure if they’re going to be able to get proper resolution with HW4 In addition to this, we also have memory. When you’re pulling out of a parking spot in some small private lot that is one-way, you’re usually angling the car out correctly because you remember which way you’re supposed to pull out. If you have no memory, oftentimes the only possible cue is a small arrow painted on the tarmac. Sometimes this isn’t even available. In scenarios like this, I don’t see how even the most intelligent AI can perform correctly without memory. There are definitely other moments where human memory is 100% relied on for correct vehicle operation. Maybe the system just needs to be intelligent enough to hold context from one drive to the next when it recognizes scenarios like this? Idfk. There’s probably a way to do it without human-level memory, but seems hard. I think in terms of safe driving, Tesla has already passed human capability and I have no concerns here. But when it comes to autonomous fleets at massive scale without being disruptive, I wonder whether they’ll be able to do it with HW4