Childhood occurs before adulthood: Awards and misguided hype aside, AI is on the wrong side of the tracks
It’s very simple …
Suppose someone uttered the following sentence:
(1) Carlos has a Greek statue in every room in his house.
A 4-year old understands the above sentence as “In every room in his house, Carlos has a Greek statue” — that is, the 4-year old somehow knows that “a Greek statue” does not refer here to a single statue but to many — that is, they somehow know that the scope of the quantifiers has to be reversed (from A-in-Every to in-Every-A) and precisely because they know a physical object cannot be in more than one place at a specific point in time. In other words, they know the following:
Consider also another simple example. A 4-year old hearing the sentence
(2) Tom broke his leg in the car accident his family had near Toronto.
knows, when asked “where did Tom break his leg” that it happened near Toronto, and precisely because they somehow know that a sub-event (e.g., Tom’s breaking his leg) has the same location of the bigger event (the family’s accident). More specifically, they are somehow aware of the following Universally Valid Cognitive Template:
Consider now the following exchange between David and his mother:
David: Mom, don’t forget my baseball glove.
Mother: Where did you put it?
David: I put in the blue back bag?
Mother: Oh then it’s already in the car.
David would know from his mom’s last reply that the blue back bag where he put his glove was already placed in the car and that all is OK, since he knows the logic of the “contained-in” relationship:
And finally, consider the following example:
(3) The trophy did not fit in the suitcase because it was too small.
Again, a 4-year old understands that “it” in (3) refers to the suitcase (if “small” was replaced by “big” then “it” would have been a reference to the trophy). The 4-year old would arrive at this understanding because they know the following:
The universally valid cognitive templates (UVCTs) 1 through 4 are, as the name suggests, universal — meaning they are individual, culture, age, etc. agnostic. They are not a function of any individual experience. They are also not “approximate” nor “probabilistic” — they simply are true, in-spite of us and our individual experiences and observations. And because we are not allowed to learn them differently, they are not (and they cannot be) learned. Admittedly, how we come to respect such universal laws/rules and how such logic was encoded in our DNA is not a question we know how to answer. Regardless of such a big question, however, it is clear that these commonsense cognitive templates are the basics building blocks that allow us to function intelligently in the world we live in (while the above examples involved language understanding, the same applies to planning, reasoning, etc.)
So what kind of AI are we doing (now)?
The four universally valid cognitive templates (UVCTs) we discussed above are clearly needed to do basic language understanding. In fact, children achieve language competency only after they have experienced (and instantiated) these cognitive templates — that is, after they become aware of the basic universal laws of how the world they live in works.
in Summary,
The universally valid cognitive templates (UVCTs) are agnostic to individual experiences/observations, and are culture, age, location, , etc. agnostic (the logic of the universally valid cognitive templates discussed above holds true in Bangladesh as much as it holds true in Norway). Thus these UVCTs are not learned, but are there and we simply become aware of them. Moreover, the only way these UVCTs can be modelled is using a symbolic logic that allows us to quantify over symbols of specific ontological types.
Before we master adulthood by finding complex patterns in massive amounts of data, therefore, it would seem that AI needs to first master childhood — acquire the basic commonsense psychology and naïve metaphysics that allows us to function as intelligent agents. In fact, I argue that it is this type of knowledge (that is acquired, and not learned from experience/observation) constitutes most of the knowledge that matters for building an intelligent agent. In my view, therefore, we are still not tackling the most important aspect of building intelligent machines — we are still on the wrong side of the tracks, so to speak.
A concluding remark
The proof below proceeds in logical steps. To refute any claim (including the last concluding claim), you have to refute any claim above it. If you cannot, then you must accept the conclusion.
- Whatever we learn from observation/experience we learn incrementally and can learn differently because we have different observations/experiences.
- Most of the knowledge that matters to our functioning as intelligent agents is not learned but acquired because (i) it is not acquired incrementally and (ii) we are not allowed to acquire it differently based on our individual observations (see note 1).
- From (1) and (2): Most of the knowledge that matters to our functioning as intelligent agents is not learned incrementally and is not learned from observation/experience.
From (3) we conclude that extreme empiricism — the extreme bottom-up “learning” paradigm that currently dominates AI — is misguided, to say the least.
An immediate U-turn is in order — and specifically, a logical turn!
Note 1
Informally, that the universally valid cognitive templates (UCVTs) are acquired and are not (and cannot be) learned is due to two important facts: (i) learning happens differently by different individuals because of their different individual observations and experiences — since these UCVTs cannot be acquired differently they are therefore not learned but acquired; and (ii) learning is approximate and probabilistic; while UCVTs are not, and thus — again, UCVTs cannot be incrementally and “approximately” learned, but are acquired.
