Copernican Shifts

In which we learn how adult development theory might help uncover the reasons for the limits of artificial mean-making systems.

Ade and Dimitri independently pointed me to Robert Kegan’s Adult Development Theory (ADT, aka Vertical Development). This caused my pandemic-strained mind to undergo an irreversible phase transition.

TL;DR: according to Kegan, human cognitive development is a series of transitions from “subject” (what we are, what we lack agency on) to “object” (what we have, what we have agency on).

From birth to death, human acquire sophistication in their ability to make meaning out of their sensory perception by creating “layers of indirection” between what they feel they are and is acted upon and what they feel they have and they can act on.

These transitions are often painful, traumatic, confusing and destabilizing. Infants, toddlers, children and teens all go thru such transitions at roughly predictable times. Jean Piaget studied all this with fantastic brilliance and foresight.

Kegan takes over Piaget’s theory and extends it to adults, consolidating the difference between “horizontal development” (the ability to expand one’s mind with more information along the plane of one’s current cognitive capacity) and “vertical development” (the ability to expand one’s mind by moving vertically away from the plane, gaining another dimension to with to observe the previous plane of cognition at a distance).

This is akin to a “copernican shift” of one’s own mind: what was complicated to understand when the center of the world was the ego becomes simpler to reason about when the center of the world is something else.

Basically going from this

to this

I recognize this will make exactly zero sense to you at such an abstract level, so I’m going to use a different model to explain how insightful this theory is.

Most structured knowledge representation systems use the “triple” as the unit of information representation. The components of such triple are (subject, predicate, object). Subject is the actor, the agent, the origin. Predicate is the type of action and Object is the entity the subject is acting on.

Knowledge representation systems can “make meaning” out of data by stitching together as triples. The set of all the triples is called a “knowledge graph”. Horizontal development is effectively equivalent to adding more triples to such graph. We could start with things like “Stefano -(born in)-> Italy” and then augment it to “Stefano -(born in)-> Padova -(part of)-> Italy”. Such augmentation is what we mean by “horizontal development”.

One simplistic view of “mean making” with triples appears to be whispering to us that the limit of our knowledge graph are effectively missing triples. All we need to do it to a) add more triples and b) make sure stitch them together tightly (by avoiding duplicating entities).

But what if I disagree with the assertion that I was born in Padova (which, in fact, is false)?

We need our first copernican shift. We need a way to treat an entire statement as an entity we can make assertions about.

Once our mean making machinery is expanded vertically in such a way, it is possible to model statements such as “Stefano -(disagrees with)-> [Stefano -(born in)-> Padova]” but also “Stefano -(believes that)-> [[Stefano -(born in)-> Padova] -(is)-> False]”.

Why is this vertical development? It is an expansion in the modeling capacity of our mean making system. Even if our knowledge graph had no limits to the amount of triples it can be composed of, it would have never reached the ability to make sense of belief or disagreement. This is similar to how a 2D plane is effectively infinite, but it would never be able to contain even the smallest 3D cube.

Piaget identified four “stages of development” which are the cognitive copernican shifts that most humans undergo during their development. For example, infants learn to recognize that “hunger” is something they have (and can change), not something they are (and can’t). Not coincidentally, this coincides with their ability to understand “object permanence”: that things don’t disappear from the world simply because they were hidden from their view. Peek-a-boo goes from being “wow” to being “duh”.

Kegan expands Piaget’s model all the way to adulthood and adds more stages of development. What’s intriguing to me about the parallel between stages of cognitive development and modeling expansion via subject->object shifting is the notion that it might explain why all of our current methods for artificial “mean making” being symbolic (ie. knowledge graphs) or sub-symbolic (semi-supervised neural language models like GPT-3 or LaMDA) both feel “shallow” compared to a human being even when expanded to vast richness and sophistication.

Just like the largest 2D plane can’t fit the smallest 3D cube, it feels to me that the limits of “mean making” reside in such “vertical development”, in the richness the ability to make abstraction and increase the dimensionality of the mean making machinery.

These thoughts tickle me for two different directions: one for what they mean for artificial general intelligence and one for what they mean for large-scale boosting of vertical human development to facilitate the ability for humanity to tackle problems that require coordination and collection action at dramatically larger scale.