And every one of those decisions is more data they'll feed the machine.

There has to be a point where you've got enough edge cases and bizarre situations where you have the data to train the LLM on how to deal with them.

I doubt our dystopian future is completely human-less, but you can probably plot a graph that projects the number of human interventions required for every thousand miles driven and that graph probably has a trend going downwards.

Then you move to another city that has more edge cases and repeat the process. Google can afford the slow rollout, Tesla can't.