Gigaom AI Minute – September 7

Transcript

Two areas that AI isn’t very good at are improvising and contextualizing, both of which come naturally to humans. Everyone, at every skill level, can improvise in a way far beyond a machine. If the door handle breaks off in your hand, you don’t just stand there frozen, unable to fathom what to do in a universe you never contemplated, a universe where door knobs don’t just turn, but they literally break off in your hand. No, you try to figure out a way to get the door open.

Consider, for instance, the challenge of building a robotic plumber. Every house is different, there are countless variants of plumbing products, and there are almost limitless things that can go wrong. A plumber doesn’t have to train on every variant of every product. So when the owner of a historic home calls a plumber and says, “I need to have my downstairs bathroom made handicap-accessible, but I want as little changed as possible,” the plumber doesn’t panic and think, “Oh no! I have never trained on that.”

With regard to contextualizing, AI also has a hard time. If you were driving through town and saw a puppy in the road, a toddler running towards it, and a grown woman darting out the front door frantically running towards the toddler, you wouldn’t have trouble piecing that scene together. But to a computer, it is just a series of patterns and vectors. Really, it’s just a bunch of ones and zeros. Think how easy it is for a human to figure out what is going on in a photo: that's a conga line, that one is people hiding for a surprise party, a prom photo taken by a parent, that’s a piano recital, a school play, a christening, and so forth. Every one of those is easy for us because we have the cultural context to decipher it.

These are just a couple of the many challenges AI researchers struggle with every day.