[tarcyanm]

I love the idea of self driving cars but it scares me that no one seems to remember the history of the 5th Generation project (I was a kid in the 80s) and other AI hype-dreams over the years. It has always been the case that approximate solutions can be found for 85% of cases, which has given the illusion of a nearly solved problem.

There is a lot of hype regarding deep learning, but I have struggled to find a concise definition apart from the fact that it is now relatively easy to work with monstrously big networks. Backprop and related algorithms have been around for decades. From what I remember of neural nets, one huge drawback was that they would be close to un-debuggable. The learning contained in the net would be inscrutable to a human, to all intents and purposes. Failure data could be recorded and replayed, but any actual reason for failure would frequently not be found. So, tweak the network, resize some layers and try again... I can think of several reasons why that's fundamentally unsuitable to the problem of driving.

I was in Egypt recently, and the sheer amount of lane crossing, merging, pedestrians ducking through multiple lanes of traffic, roadside obstacles, donkey-drawn vehicles etc would be 100% impervious to even a level 3 solution today. I believe the same would be true in India and many other parts of Africa.

So, I really hope that we are not falling blindly into another 5th Generation sinkhole here. History should have taught us better.