What is limiting progress if training isn't to blame? Design? Execution?
It's just a really hard problem that no one knows how to solve, and it's being attempted on extremely compute-limited inference hardware, with quite limited camera resolution, which makes it even harder.
I think they'll figure out ways to get better results, but it's going to take a while. More training resources will probably help them iterate faster.
I am much more interested in the following list since they definitely affect if FSD will get me to the destination safely and many are the law. That is my priority and Tesla's I suspect for 8/8. Except for #1 none of the above would normally prevent you from getting to your destination safely.
Your list seems to be from the perspective of having a product than can handle every situation adequately, so that it can be completely hands off. Kind of the "jack of all trades, master of none" approach. I am more interested in lower intervention rates where it masters the most common tasks and rare occurrences can be mostly dumped on the human. I can understand the alternative philosophy, but I just prefer to have lower intervention rates, and it sounds like this is a Tesla metric as well.
I just don't think many of the things you list would show up at the top of the intervention Pareto.
- Road closure signs so FSD automatically reroutes
Yep, read signs.
- Emergency vehicles- pull over it's the law
- School zones- slow down it's the law
- School busses- stop it's the law
- Hand gestures- stop/go when directed- it's the law
These just are really rare scenarios for me so I have no problem just taking over for them. (School zones - need to read signs, falls in that category which I think is important as it was in my list!)
- Backing up so you can get to your destination
Should never need to do this if it's reading signs, and it seems like just a situation where it will nearly always be better to just take over.
Yeah would be nice but seems really, really hard and I don't want the car waving all over the road dodging imaginary potholes. I'm ok just waiting for the potholes to be repaired and when driving the same route I know where to disengage. Some memory or whatever would be nice.
But just lower priority for lowering intervention rate. Some roads in San Diego are just unusable for FSD, and I can't imagine a pothole-dodging ability would be acceptable. Check
this out, for an example - and this is BEFORE two solid years of rain! I think we just have to accept that something that is essentially a gravel road complete with potholes, except worse, will be impossible for any implementation, ever. It'll just have to go really slowly rather than veering wildly all over the road at 25-30mph like I do, or just hand off control to the driver. (It's possible this specific example has now been fixed, I'd need to go check - but there are always others!)
- Read signs correctly- Example "No left turn" etc.
Yes it should read signs as indicated in my list.
That is my priority and Tesla's I suspect for 8/8.
Nothing is happening on 8/8 for our FSD. I'm not sure why people think it will be significant. This will be like Battery Day or AI Day 1/2 or whatever. Nothing happens. What could possibly happen on 8/8 which would be significant for current vehicles?
It's possible they will reveal FSD Beta, the beta version of FSD. But vehicle support may be limited. We'll see.