My understanding is that, while there's no doubt that the "rules of the road" are specified in the system, the diving policy/decision layer is not "hard coded" in that there is a procedural code layer implementing these decisions... at least not any more.I think that doc is a good watch and can inform thinking but it's healthy to keep in mind they are not very comparable.
AlphaGo and the subsequent software is an AI that plays a game and has a very clear identifier for declaring an improvement, it wins.
AlphaGo can play itself, or its predecessor in a virtual environment to improve.
FSD is - at least as I understand it - two things. An AI developed to interpret visual input - Tesla Vision-, And hard coded rules for driving.
Tesla Vision can't really play against itself as I understand. I guess its possible that they could feed two models through a video clip and then whichever model had a higher degree of confidence about what it saw wins. But to validate it it needs to roll out to a test fleet and have users say "Hey this is way better, but it tried to drive through the concrete pillar under the monorail".
That process is much slower than a computer playing itself 24/7.
AI Day could prove me very wrong though.
At Autonomy day, Karpathy talked at some length about moving the policy/decision making process in to a neural net... and the "Software 2.0" model. As such. there's a NN for vision, and one or more NN's for the decision making process(es). And those decision-making NN's would take in to account the specific rules of the road.
I think we see evidence of this in the FSD videos where the car will cross the dividing line when no oncoming traffic is present to allow pedestrians more space, or similar.