No. Perhaps between models, I don't know if they train all model's neural nets on the same training set. But definitely not on individual cars. AI works by training a neural network on a huge data set (millions if not billions of kilometres of video). It takes a supercomputer to run this training, not something you can do in each vehicle (also, each vehicle doesn't have enough training data). That NN (i.e. the result of this huge training exercise) is then deployed to the fleet. So all we get in the cars are the hidden layers of the NN.
Tesla have in fact given up on catering to special cases and have in their latest iteration for US FSD removed 300k lines of code that was handling special conditions to now base the entire behaviour on "quality checked and good" data from the fleet. This means for example that on the latest FSD stack in the US, the car will no longer adhere strictly to speed limits, but will drive the way a "reasonable" driver would, and to match the traffic flow. While that is (in my view) the way it should be done, the law very clearly states otherwise.
This approach is extremely worrisome because now there are no sanity checks left and Tesla will not be able to tell you why the car did what it did. Whatever it does is simply the output of the NN.
In Australia, we have not had any significant improvements in autopilot functionality since 2020. Autosteer still malfunctions at the same Delhi Rd / Epping Rd intersection, it still mistakes 40 signs on buses for the speed limit on occasion, it always still mistakes lane open indicators in tunnels and on bridges for traffic lights and wants to stop, and it often misreads speed signs that are not applicable to the lane I'm driving in.
It sucks to be in bad wiper land. Mine have never performed satisfactorily, but then again I can switch them off with a physical twist on the left stalk, and press on the outside of the stack when I need a wipe - don't know how that works in the stalkless cars, but I can imagine it's a huge pita.
Last edited:
