Having tested a modern Tesla car a few weeks ago, I was mostly impressed by how bad the vision algorithm was at understanding the roads. There were plenty of en

Having tested a modern Tesla car a few weeks ago, I was mostly impressed by how bad the vision algorithm was at understanding the roads. There were plenty of entirely fixed curbs, intersections, lane lines, etc that it could not handle correctly, and this was on a clear, dry day with excellent visibility. They’ve certainly pushed the willingness to use so-so technology forward, but I was really not impressed by the underlying technology. The odd thing is that these type of errors really ought to be correctable by training with a small fleet of humans annotating data. AI is very, very good at getting examples from the training set right, and fixed roads a few tens of miles from the old headquarters ought to be in or at least very similar to the training data.