Todd Burch
14-Year Member
While conceptually you are correct that different data could be used to train different networks in the different modes, doing this with just two toggle settings means you have to train and test 4 separate networks. Then multiply by the regions that need their own networks and this quickly becomes an unscalable problem.The "preferences"(like Chill mode etc) typically become inputs to the NN, with appropriate training data for each mode, though this can be expensive if there are many such settings (one reason Tesla keep them simple).
Dont forget the NN doesnt just get camera data .. it also get map info, accelerometer, GPS, compass etc.
No way is Tesla doing this. To maintain the chill/average/assertive modes + minimal lane change toggle and 3 regions (NA/Europe/China) means you now have at least 18 separate networks to train. Then those all have to be tested. And they all need vast amounts of data. The compute time and extra storage required alone would be prohibitive.
I think the settings were maintained to support the continued v11 highway stack. For the next few years, I think the only settings we’ll see are ones that can be applied to the output of the neural network by traditional C++ guardrail code.