[rogerrogerr]

> AFAIK what Tesla's software does is object detection (cars, people, bikes, road sides) and only uses that. Which of course is not at all safe since there can always be things on the road that are an obstacle but not recognized.

This is not true; in modern FSD the visualization is disconnected from the actual driving model. The visualization runs a lightweight model that labels some stuff and shows it; the actual heavy lifting is now a single model that takes pixels as input and outputs control commands. My car very clearly reacts to stuff that doesn’t appear in the visualization.