I, For One, Snicker At Our New Robot Overlords

Inigo was chasing his tail last night. Amusement is another perk of dog ownership. He wanted to chew on an itchy spot on his back, but the temptation of the waving, fluffy tail was too much for his hunter instinct.

I don't think Inigo is a stupid dog. A robot capable of all that Inigo could do would be worth a fortune. The state of the art robot can roughly approximate Inigo's ability to walk over rough ground, but it has none of his grace or speed (although Inigo has a hard time on wooden floors.) In addition, Inigo has cognitive abilities we don't understand well enough to teach to a robot. He can smell and eat old dog poop before I even see it is there. He recognizes people who should be greeted enthusiastically. He knows how to sneak food off of the counter without my noticing.

Computers and algorithms are good enough to now beat pretty much anyone at chess; however, Go masters are still unchallenged by computer programs. The robot soccer tournament is interesting, but still mostly silly.

I believe that even when robots become clever and mobile enough to take over, we can still expect them to be stupid in a lot of ways. They will likely be stupid in the same ways that we are.

I have worked on human directed balancing wheelchairs. Every person who sees one wants to know when we were going to add a vision system and give more autonomy to the machine. But before we give autonomy to the machine, we have to figure out how it will detect when the family cat is twining around the wheels (or more importantly, when a grandchild is underfoot.) Vision systems can make balancing more stable, just as your eyes help keep you stable. (If you doubt how important your eyes are to balance, try to stand up strait with your eyes closed.) But if we use vision for balance, we need to figure out that we are next to a moving truck. This unexpected motion would cause the same dizziness in the machine that it does when you are stopped next to a moving truck at a stoplight.

A lot of human (and canine) stupidity is tied to carrying around an incorrect model for too long. I talk to my mother about iced tea, and when she picks up her soda, she is disturbed by the fact it doesn't taste like tea. I turn onto the highway on a Saturday morning, because I pull onto the highway every morning to go to work. I pick up the book I am bringing home from the front seat of the car, and leave my keys in the ignition, because I have already taken one thing from the car and I always take just one thing from the car.

I believe a lot of professions requiring quick, life and death decisions, e.g. doctors and soldiers, mitigate this through constant and extensive training. Young doctors are presented with a ludicrous number of patients during their training so that twenty years later, they have pathways in their brain to recognize symptoms and to apply treatment. Pilots supplement their training with procedures and a controlled environment. When something is wrong in an airplane, it disturbs the meticulous arrangement of the environment and calls attention to itself. Yet pilots, abetted by the automated systems of their planes still make fatal mistakes.

Our robot overlords will also have to face this problem of when does its current model of the world no longer apply. I watched a talk by Marvin Minsky. His thought was that it is stupid to think that computers won't be able to think, just because all they can do is perform billions of logical operations in a second with perfect accuracy. He also thought that it is stupid to believe that there will be a single approach to artificial intelligence. Sometimes the computer will be an expert system, sometimes a neural net, and sometimes a decision tree.

Computers will have to decide to when to switch between approaches and when to alter its model of the world. They will be just as dependent as we are upon the world providing them the clues that things have changed. They will have the advantage that they can more consistently monitor their environment and they can calculate the relative errors in all of the models they are using. But even they will be limited to the amount of information they can process and the number of predictions they can make. They will likely be better than us in almost every way, but we can still expect a fair amount of tail chasing.