#5 The Ladder of Causation

Why machines are so far away from General Intelligence?

We are obsessed with the project of digital intelligence. However as Douglas Hofstadter has said, “each new step towards AI, rather than producing something which everyone agrees is real intelligence, merely reveals what real intelligence is not”.

Speaking about real and human-like intelligence, I found this concept of Ladder of Causation quite insightful. Judea Pearl writes in his book, The Book of Why, that we humans have 3 levels of cognitive ability: Seeing, Doing and Imagining.

  1. First level is “Seeing” ability. This ability is about detection of patterns in the environment, the ability we share with most animals. The cognition is about seeing the association between things we observe, e.g. What does symptom tell about the disease? What does survey tell us about election results? According to Judea, the present day AI/ML is still mostly on this level. This ability concerns with the world we can see.

  2. Second level is “Doing” ability. This ability is about predicting effects of alterations in our environment. Only a handful of species demonstrate this ability, e.g. how many species have an intentional use of tools? This level is higher than just association because we are making interventions. AI/ML can’t reliably predict results of interventions because an intervention breaks the rules of training environment/data. This ability concerns with the world we don’t but can see.

  3. Third level is “Imagining” ability. The ability concerns with our ability to create theories and observe counterfactuals. Counterfactuals are like What-if scenarios. Counterfactuals don’t rely on data. Counterfactuals give us the ability to think, reflect and improve. We can imagine and believe things that aren’t real. Much of our culture relies on our ability to imagine, e.g. Movies, Novels, New Scientific theories, Sci-fi. And when we choose, we can test those beliefs to create new knowledge. This is highest level of cognition, only humans demonstrate this. It concerns with the world that can’t be seen.

Our machines have become really good at “seeing” and giving us “doing” suggestions. But until we figure out models to impart them “imagining”, we can say that they are nowhere close to general intelligence.


Quick Notes on Abstraction: Whether it is making snow art, or writing software - Abstraction makes representation flexible and constrains expression.


Quote: “The machine itself makes no demands and holds out no promises: it is the human spirit that makes demands and keeps promises.” - Lewis Mumford