No matter how much sense it makes to programme human logic into AI systems, doing so ends up limiting how smart they can actually become.
We’ve seen examples of this across all major areas of AI
– From computer chess and Go games
– To speech recognition
– And computer vision.
Researchers start off by trying to hard-code all their knowledge about the problem into the algorithms.
And at first it works okay, which makes them confident they’re on the right path. But, then something else comes along that changes their approach.
Eventually, a new AI system emerges that simply uses brute force computing power and learns entirely by crunching massive datasets.
What Sutton argues we need to do instead is develop general techniques like search and learning – meta-skills that can rediscover the intricate patterns and solutions that we could never conceive of on our own.
By
By Dean Hardy-White
AI/Tech Ethics Expert
Comments are closed.