I respect Thrun. He's a brilliant researcher and a very smart man. I've read his "Probabilistic robotics" book. I can also tell you right now that he's _wrong_. There will not be purely optical FSD on non-modified city roads shared with humans in the foreseeable future. Highways? Maybe. But not city roads. The problem is exponentially more complex there. It is even more so on top of that if you have to _estimate_ rather than _measure_ distances to things. It is true that the LIDAR does not simplify it all that much, but LIDAR at least gives you a fairly good idea of where all the obstacles are so you could avoid them. That's what it's for. Cameras fed into neural nets have to _guess_ based on appearance of things and stereoscopy. They are also subject to all the other limitations of optical sensors: rain, snow, fog, dirt on the camera (anyone who has a Tesla knows the cameras get dirt on them pretty easily), and just plain not recognizing an obstacle if it's positioned in a way that the neural network just happened to not generalize for.
We (humans) can only do this purely optically because we have general intelligence and extensive knowledge about the world. The FSD computer in your Tesla does not.