The pre 12.5 stack uses traditional C++ conditionals together with computer vision for highway driving. 12.5 and after is the a single stack setup (at least that's what Musk claims).
The 12.3.6 (or the 12.4.1 that's being trialed right now) is pretty much the current production state of the art for Tesla street level (not highway) single stack self driving where the steering commands are directly issued by the neural network.
It’s about as shit as the 300k lines of c code. Added with drink spilling acceleration in 12.3.6 for no reason at all.
There are major improvements in taking a curve and not scaring the shit out of the driver, much faster basic turn right on stop sign, lane changes when it does it, it’s hard to describe just how smooth they are it’s a little nuts the car just floats over it’s not human, but it really should not have changed lane most of the time… so decision logic is shit. Etc.
And that's the whole problem with it - there isn't really any 'decision logic', it's a giant low level neural network trained on the outputs of human drivers. It understands driving less than GPT understands semantic concepts in language. There's no 'executive' high level control, it's just a big stupid animal brain that reacts to stimuli like 'turn left ahead' or 'car ahead slowing down'.
The 12.3.6 (or the 12.4.1 that's being trialed right now) is pretty much the current production state of the art for Tesla street level (not highway) single stack self driving where the steering commands are directly issued by the neural network.