Presumably this is referring to 12.5.x where 12.4.x "only" increased the training data and training time while the newer one also increased neural network size to take advantage of additional training compute for the potential for it to learn more nuanced differences in expected behaviors.
12.4.x was trained on 12.3.x interventions where drivers needed to "do something." Most of driving is just staying in lane continuing straight at the same speed, but if the neural networks are overfit believing a "just keep driving" situation is actually "do something," this makes it less smooth because it does things unnecessarily such as switching lanes or slowing down.
I did tell everyone you are a reliable source.
It's a waste to give it to Chuck Cook. He won't touch it.
I’d like to know how the speed issue is that’s a huge problem for me
On a trip from Palm Bay to Cape Coral instead of turning on Fl70 from I95, it wanted to go 2 miles further, take a two lane road that eventually connected to 70. Would have added miles and time to the trip. Then it saw 30 miles of 70 as 55 not 65, which is the limit. In Okeechobee it wanted to do two unprotected left turns in heavy traffic rather than go one block to a stoplight with two turn lanes. Is that enough?
I am hoping we will see more improvement in the FSD route & lane planning to be more human like. It seems as though it needs to learn common sense.
You posted after people downloaded v12.4.2. I hope this is not v12.4.2 behaviors.
I'll just say that most of that is not FSD problems, that is navigation system issues. (The exception is getting the speed limit wrong, which it should be able to read the signs correctly.)
