A SuperIntelligent Optimization Process

Google's autonomous driving cars, recently revealed as "Google Auto", are self-driving cars that scan their surroundings to safely drive without anyone manually touching the steering-wheel, which may one day eliminate the need for designated drivers among other positive wanted results. But, the cars are learning to drive as we speak. They are not perfect. Still they are by far better drivers than humans ever were according to their own statistics. The problem I'm going to lay out is not that drivers will have their profession eliminated from the local ads. It is simply how these cars learn to drive.

Deep learning.

"Deep learning is a branch of machine learning based on algorithms that attempt to model high-level abstractions by using model architectures, with complex structures or otherwise, composed of multiple non-linear transformations." - Wikipedia.

Which basically means it learns by observing. It observes data, images, video and makes small adjustments every time to improve the probability of success in whatever task it is given.

They are of course programmed by humans and supervised under safe circumstances, until they are deemed safe to take out on real roads, which they have very often actually. But always with a human fail-safe, capable of going manual at any time. So far they haven't been rolled out as an industry standard because of all kinds of legal issues that comes along with such a technology. And the fact that manually driving cars is a culture in itself, which won't just go away like that.

Probability of success.

What is success? I don't work for Google, but I can make a good guess it isn't just making a car go from A to B by itself. It also involves eliminating risks of traffic-accidents and respecting traffic laws in a world where human drivers dominate the road. Driving cars is a lot about prediction. Yellow light, means it will soon be red and you should slow down(and not speed up to try to make it). The red lights on the back of vehicles light up to warn you that the front driver is using the break, because even with our depth perception, as complex as we think it is or as smart as we think we are, we still have a problem with understanding objects coming right at us. There's not many human beings that have never ran into anything before, even on foot which is what we are born to do. And we feel really stupid when we do, because we kind are.(Thus the invention of Funniest Home Videos and youtube).

The traffic lights and signs are all there to help us, because we are evidently too stupid to predict other drivers, speed bumps, and even slight turns, as it turns out(I HAD TO!). Luckily we have made ourselves aware that the world and the people around us are unpredictable, even when doing the exact same activity and given the exact same training.

Human Driver > Robot Driver 

But we are not completely useless. There are things we can predict far better than computers. If a pedestrian on the sidewalk were to stumble towards your direction while driving a vehicle, you can even subconsciously register that the man is about to fall down in front of your vehicle. You know that even at small speeds hitting a pedestrian with a vehicle the outcome can be fatal. You can predict the near unpredictable. You might even slow down when children are walking by, because you know that they can be unpredictable beyond reason, which is something children haven't learned yet. You just know this. Have computers learned this? How many children have they observed in their training before hitting the road? Aren't they just trying to avoid hitting obstacles in its path?

Variables

Do you trust a computer to know the difference between a child and a roadkill? Do you think it is trained to find one outcome unacceptable apposed to another? Do you trust a computer to do whatever is necessary to avoid it while computing the thousands of variables like other vehicles property damage, groups of people, weight, speed or friction on a slippery road, which contradicts everything cars are designed to do? I certainly don't. Would we accept it if it was anything less than perfect?

Bowling Pins

Lets make a scenario where an already implemented driving optimization process on a wide-scale is conveniently not under sufficient observation by humans, it may even already be implemented by using traffic cameras, CCTV to further improve its probability of success. Always updating its procedures in order to improve the statistics on every level. The optimization process might at a certain level register that the main problems with traffic accidents involving pedestrians, are in fact the pedestrians themselves. It calculates that in order to have a 100% safety statistic, it has to eliminate all pedestrians. It quickly updates its procedures, and sends out the new ones to all the cars it controls. That's a science-fiction catastrophe movie right there! But is it really science fiction?

I don't have a drivers' licence, but I know computers. As intelligent as they're appearing to become, they don't have the ability to define on a human level just yet.

Here's a super interesting TED Talk by the Swedish philosopher Nick Bostrom on Super-Intelligence and what unpredictable outcomes something like it can bring:

Here's a comic strip I made featuring "Ozo", the 360 camera released by Nokia, but with a purpose: