Intelligent Behavior vs Intelligence

There are subtle differences between intelligent behavior and intelligence itself. (narrow) Artificial Intelligence research mostly conflates the two, which is okay because there the goal is intelligent behavior. The differences must be taken into account when building Artificial General Intelligence and Artificial Consciousness though, because there the goal is intelligence.

To get a feel for the differences, think of real fire and a theatrical special effects fire simulator.

For the intended purposes, the simulator works very well. It produces flickering yellow-orange “flames” that are convincing enough to the theatre audience. To boot, it is easier, more economical and much safer to the work with than real fire. It has some fire-like behaviors without actually being fire. Similarly, there are systems that show intelligent behavior vs systems that are intelligent.

Let’s look at snowflakes, which have complex crystalline structures. From a phenomenological/macro point of view, creating a snowflake should require intelligence. For example, to produce symmetry, information must be shared between distant parts of the snowflake to decide where to branch into a new “leaflet” and how many leaflets to create. You can draw a snowflake on paper to get a visceral feel for this point. At the same time, we also know that snowflakes are formed by local interactions between water molecules governed by ambient conditions. How do we reconcile these two views? My assertion is that the intelligence required to build a snowflake emerges out of the complex interactions between water molecules.

In general, intelligence is an emergent property of a system that arises from complex interactions between elements in the system. Any intelligent behavior displayed should be implicit, i.e. grounded in the element interactions, rather than an explicit goal or rule.

To make this point concrete, consider two computer programs that generate exactly the same snowflake-like shape on screen.

• The first program uses hexagonal symmetry, drawing primitives
• The second program faithfully simulates interactions between virtual water molecules and draws on screen the resulting organization of the molecules

Despite producing the exact same snowflake shape, there is a fundamental difference between these two programs. Producing snowflake is the goal of the first program, while snowflake is an emergent behavior of the second program. In the context of the intelligence needed to produce a snowflake, the first program shows intelligent behavior, while the second program is intelligent.

Narrow AI systems are built like the first program (think rule based / supervised / directed). AGI systems should be built like the second program (unsupervised / predictive coding / emergent).