I am not surprised if V12 ends up being better in every day, common driving. End-to-end is a very efficient way of training NN. And NN are very good at imitating patterns. Tesla says that they used millions of clips for V12. So if you plug in millions of clips of good driving, you can except V12 to be good at imitating what it sees in those millions of clips. And I am not surprised if V12's driving is more natural, more human-like. Tesla has trained it on millions of clips, presumably taken from good human drivers, so we would expect V12 to imitate that human behavior and therefore drive in a more natural, human way. We also saw this when Waymo changed to a ML-first planner. They also saw a big improvement in the driving being less robotic, more natural, more confident.
My 2 big questions right now (which I am sure will get answered in time):
1) We know in the past, Tesla was overfitting the data, causing FSD beta to drive better in places like SF and not as well in other parts of the country. Will we see a similar pattern with V12? I look forward to trying V12 in my area where FSD Beta has struggled for a long time.
2) Will end-to-end get to 99.99999% reliability where Tesla can remove the driver supervision requirement and if yes, how quickly will that happen?