I've had to explain this aspect of Level 3 to so many people before, and in my experience few people get it. I think there's a bit more nuance to it though. H

I've had to explain this aspect of Level 3 to so many people before, and in my experience few people get it. I think there's a bit more nuance to it though. How far it has to predict something into the future is however long it takes for a person to safely take over - let's say 10 seconds. The thing is though, a Level 4 system also has to be able to predict things into the future in terms of erratic drivers and pedestrians - so the problem is surprisingly similar. If you want to know what's really going on with self-driving, talk to the prediction teams at any of the companies. It's pure research, because how can you predict the future? And it's not a hypothetical problem, because pedestrians do weird shit and you need to predict if they're going to walk in front of you 5 seconds from now all the time (at least for urban driving). Humans do this instinctively - we slow down in uncertain situations and are constantly monitoring it for changes. When the robots try to do this, they slow down and then just stop because they can't really guess what's going to happen next. For computers, I'd argue it's roughly equivalent to solving hard AI.