Intelligent Machines
Many AI systems are built using neural networks; a type of computing system loosely modeled on the organisation of the human brain.
Neuroscience often uses neural networks to try to model tasks performed by the human brain. However, it’s misleading to compare the workings of the dynamic, living biological brain with an artificial neural network, which is a greatly simplified mathematical model.
-
AI isn’t programmed like software. Instead, AI agents, robots, self-driving cars, etc are trained to perform complex tasks using Reinforcement Learning (RL).
The AI interacts with its environment through trial and error and is rewarded when it does something correctly. The hardest part is knowing what to reward in order to get it to do what you want.
This article from OpenAI describes what can happen when developers choose inappropriate reward systems.
-
Current AI models use processing power and massive amounts of data to generate essays about complex topics or outperform humans at games like chess and Go.
But they cannot adapt to new, unfamiliar situations or understand ambiguity like we do. This is because AI lacks intuition.
This article by Josh Bersin discusses the role of emotions and intuition in complex decision-making.
-
AI also lacks the implicit everyday knowledge (common sense) that we take for granted when navigating real world issues.
Tacit understanding of physics, social etiquette, empathy, time, space, etc. comes from a lifetime of direct sensory experience.
By contrast, AI only has access to symbolic representations of knowledge.
In this video, Jensen Huang discusses the significance of this challenge.
-
Because AI is trained on historical data, it’s important not to treat their predictions as fact.
Predictive policing - using historical data to predict future crime is one example of this.
This video by mathematician Kevin Lieber explains how a reliance on historical data can lead us to the wrong conclusions.