Just a Thought - On Narrow AI

Some people are concerned about killer machines. Although great progress has been made in machine learning and neural networks in the past twenty years, such progress has been realized in narrow AI. While some think that general AI is on the horizon, there is quite a large jump in capability that our artificial agents need to make to match that of their biological counterparts.

At the risk of oversimplification, narrow AI focuses on solving a well-defined specific problem while a general AI would be capable of handling more nebulous and varied tasks. With the current stage of artificial intelligence, we define some sort of boundary in a defined space where we train the agent to discover that boundary by learning from a set of example data. This is the fundamental premise of machine learning.

But what happens when we don't have a boundary? What happens if we don't know the boundary? What if we change the boundary given changing circumstances. While there are a lot of cases that this boundary is static there are still scenarios where being adaptable, flexible, and acting on intuition is important in decision making that machines have yet to show a capacity to demonstrate.

Let's say we define a game with a core set of rules. Given enough training data, a model should be able to find some winning strategies if not the optimal strategy. But let's say that we add an additional rule. Any person would be able to adjust to the rule and develop new strategies, but the machine would struggle without experience with this new rule. Without prior examples, it lacks context to adapt to the change.

We could easily reprogram the model to adapt to this rule, but this takes time. It requires an outside agent to gather data and feed it to the machine. People simply are able to adjust their strategy based on prior experience and logical projection. People can reprogram themselves while the narrow AI still needs assistance.

While machine learning has opened the world to new ways of automation, there is still a lack of adaptability built in machines that prevents us from reaching general AI or even things we could consider sentient. That being said, algorithmic processes used malevolently or carelessly could still lead to catastrophe. While folks fear superhuman monsters capable of destroying humanity, humanity itself is plenty capable.