Tesla is innovative as hell, but that doesn't mean they've solved machine learning--camera only AI with an order of magnitude superior safety margins relative to humans yet. I'd say they are less than 10% there. I think they need to amass orders of magnitudes of data more than today before having enough training. Keep in mind as well, they need to have a complete! set of learning algorithms that correctly characterize being human aware/capable to use that data to train with.3-4 years behind on promises and how many years ahead of the competition?
I think other companies will get there sooner (and have registered and utilize geo fenced areas already -Waymo) since they are using active LIDAR feedback to bridge the gap and provide level of safety curtain needed until machine learning has been fully developed and single vision can be solely relied upon.
I developed an autonomous vehicle back in 2007 for the Urban Grand Challenge and felt safer with the level of tech we installed (4 LIDAR's including one flip on its side and actuated to 'sweep' and created 3d point cloud for object detection) then than I do my tesla today. Elon is right that eventually (He promised this would already be done 2020, but really I think 10-15 years), where we won't need anything other than video feeds (as long as all lenses are clear!) to completely self drive (think human learned machines of today).
Tesla needs to hire more computer scientists that can work on the millions of algorithms that will act like a brain by filtered the most relevant (conscious data) to the top of the stack so side cases are far and few between.