The Smallest Units of Introspection

How to begin removing abstraction from cognitive theories

From Narrow To General AI
10 min readMay 28, 2023

This post is part three of a series. You can read part 2 here. See a list of all posts here.

Summary of part 1 and 2: There are an overwhelming number and diversity of challenges in the path towards a full theory of Artificial General Intelligence (AGI). To begin the investigation, one should first discover things that the mind does consistently, automatically and unconsciously. These include sensing, introspection, liking/disliking, learning, and paying attention.

The ouroboros, or snake biting its own tail, used as metaphor for the introspective mind.

The mind is fond of imagining “things” i.e. abstract entities whenever it encounters reliable patterns of stimuli. Given a recurring set of coloured shapes, over time it may posit a “tree” or “boat” connected to them, and from then on attach that idea to those inputs. Both parts of that sequence— seeing coloured shapes, and responding as though an object were present — are mental “events” of some kind. They are happening at specific times, and seem to be causally connected.

Similarly, the mind imagines causal or explanatory entities when it looks into itself. These may include “memories”, “plans”, “concepts”, “ideas”, an “ego”, “consciousness”, “truth”, a “soul”, etc. Together, they may comprise a foundation on which to base a detailed understanding of the whole self. Unfortunately, introspection is a notoriously slippery, even frustrating activity. The mind tends to resist and undermine any attempt to pin it down concretely; it appears to be fighting with itself.

Perhaps there is a reason for this difficulty. Perhaps, unlike in the case of “trees” or “boats”, here the mind is getting ahead of itself. All introspection has seen is, at best, a series of mental events. Before we start assuming the existence of something beyond them let’s see if we can, per Occam’s razor, do without such entities.

The goal of this exercise is to trace the patterns of the mental events themselves over time, without inferring any persistent entities like “plans” or “the superego” behind them. The focus in this series will be on explaining each thought; what it is, and how it connects to others. If your aim is to unambiguously understand every granular detail of every mental event, be it a thought or feeling or otherwise, then ultimately what you are looking for is concrete, mechanical details, not generalized abstractions. This can be tricky, since the benefit of using generalized abstractions is that they give you something to anchor your observations to. Without them, it can be like trying to organize the water in an ocean. Fortunately, the above paragraphs have already offered a clue for some first steps.

Say you observe a memory of a Ferris Wheel in your mind. Rather than concluding that a thing called a “memory” was present to be observed, consider instead that all that really happened is you saw an image of a Ferris Wheel in your mind. This latter is an event: it is singular, it happened in a moment in time, and it can be broken down and analyzed regarding its causes and effects on the mind. It makes no claims about the truth or provenance of its content. You may later add to it the hypothesis that this image of a Ferris Wheel is connected to something experienced previously.¹ We’ll get into the details of the nature of thoughts in another post, but for now, the point is to focus on individual thoughts as momentary, sequential units of action.

Now if you’re following along closely, you will also have realized that the act of concluding that you saw a Ferris Wheel in your mind is also, itself, an action. It is a reaction to the first thought. Asserting that what you saw could be called a memory, is also a thought. The act of breaking your thoughts down analytically would also involve new thoughts. Each of these is extending a chain of thoughts about thoughts, one referring to the other, a chain which can go on indefinitely. During the process, your mind appears to be searching for some underlying truth, but even that itself is an additional thought, namely— these thoughts appear to be searching for truth.

All varieties of thoughts can fit this template. They are all reactions to something with connected content, and they were all added in the moment. Perceiving feelings is a thought, e.g. I feel cold. Coming up with an idea is a thought, e.g. I should water the plants. Deciding that something is right or wrong is a thought, e.g. I shouldn’t have said that. Asserting that something is true is a thought, e.g. that seems reasonable. Being convinced is a thought, e.g. I know that is true! A logical inference is a thought, e.g. so that means Socrates is mortal. Even introspective assertions are thoughts, e.g. thoughts can be memories of the past, or plans for the future.

The conscious, introspective self is an ever-progressing series of thoughts reacting to thoughts reacting to thoughts. Like an ever-revolving wheel, it doesn’t “pause” or “consider” or “take a rest”. As long as you are conscious, you are only adding more thoughts to the pile. It is relentlessly active. The mind never “reverses” either — every backwards movement is only a forward movement with different content. Adding a new thought to a chain is the fundamental unit of cognition.

This is what makes the introspection so slippery, because thinking necessarily changes the mind in unforeseen ways. Unlike a tree or a chair, your thoughts don’t stay still when you observe them. To think about a topic is to question it, to doubt it, to add something new to it. To assert that something in the mind is “true” changes the mind in some way; possibly invalidating that truth. Every attempt to grasp and hold onto it in a certain configuration just pushes the whole wheel forward one more foot: it adds one more thought. The mind is eternally chasing its own changing truth, like a snake biting its own tail. Any act of perception, realization, or awareness, either about the self or the world, is bound to change the path and nature of future mental actions. This makes each thought, regardless of its content, not an end in itself, but a means to new thoughts. It allows the chain to continue. There is no concrete footing or stable foundation on which to base your theories of self; no first principles, and no final settled truth.

An simple example of how even recalling a memory alters the mind going forward.

It may seem like a tautology to say that every thought is a thought with some content, but this fact is rarely fully appreciated when we are caught up in the content itself. Any compelling theory you derive from introspection about how the mind works is only a set of momentary thoughts the mind is having about itself. The presence of that theory, how it got there, what it is, all must be explained within its own theory. Inconsistencies are inevitable in any such theory. By framing the entire process as endless chains of thoughts about thoughts², you can overcome any misleading interpretations that may suggest/include static entities.

You can see why, in the previous post, some everyday mental functions were excluded from the list of fundamental components. For example, the list didn’t include memory. This is not because humans don’t have memories, it’s that an entity-word like “memory” is misleading. It suggests your mind has created a passive, static recording of the world around you, which you can access in a read-only way as needed. But memories involve complex networks of many moment-to-moment behaviours. These include the original “recording” (which is a type of learning), searching for a memory (which is a type of motive), triggering the memory (which is a type of thought reaction), awareness of the content of the memory (which is a type of learning again), and so on.

“Memory” also implies that it holds a reference to some impartial truth; otherwise it’s not a memory but a fantasy. However, during all the above steps you are selecting, editing, embellishing, re-interpreting, and often blatantly fabricating or heavily skewing what you experience. Truth is not inherently forthcoming from memories. To assert any of it as true requires an additional set of thought-actions, to filter out something demonstrably valid from some version of this mess of thoughts.

There are other varieties of thoughts that were left out of the list too, such as planning, predicting, inventing, conceptualizing, and fantasizing. Though these may be convenient labels under which to group your thoughts, clear delineations between them are difficult to establish. For example, on seeing a cow and thinking the word “cow”, is that a prediction of what others would say? A plan for what you should say? A memory of what someone said before? A linguistic reference to the concept of a cow? Or something else? The lines between these are not easy to draw. The best you can do is consider it one of many thought-responses. The only analysis then would be to see if you can find a pattern in the responses themselves. What precedes them? What subsequent effect do they have on other responses?

To explain how and why thoughts appear in the mind requires understanding the nature of learning, which I’ve deferred to a later post. For now, you can note that every thought appears to be a momentary response to something; an experience, another thought, a feeling, an action, etc. Any “truth” referenced in its contents is not a self-sustaining existence that it has lazily surveyed. You are not passively “reflecting something” in the mind, like a mirror. The thought is itself an action, i.e. it is active. “I thought” has the same connotation as “I kicked” or “I ran”. The thought “wants” to have an effect, that’s why it’s there. It is learned in a similar way as physical actions are, i.e. it’s driven by some motive. Thoughts are evaluated and kept around by usefully contributing to some need. This is the link between cognition and desire; between “is” and “want”.

If you were to reduce all of your mental complexity to chains of thought-actions, then the set of things the mind inherently does without learning them first may seem surprisingly small. It senses the outside world, and itself; it has motives and adds new ones (personal growth); it has thoughts and adds new ones (learning); it pays attention to some experiences or thoughts; and it has a certain set of emotions. Every other mental behaviour, like language or logic, can be seen as some interplay of these basic pieces. Perhaps the challenge posed in the first article is not as insurmountable as it originally seemed.

But of course, we have barely even begun. Though these mechanisms may seem simple, many of them were left vague and poorly defined; they may conceal a lot of complexity within them. And even if they were better articulated, there would still be the not insignificant task of applying them to the breadth of human intellectual accomplishments. Only then can we see where they fall short of a full explanatory set of principles.

In the next post, we’ll look into one peculiar limitation of introspection, and see how it gives us clues about an otherwise invisible pattern of mental behaviour.

[Next post: constraints on the formalization of AGI]

¹ The philosophically literate will note that the goal here is to break down any distinction between analytic and synthetic thoughts; more precisely, to completely reject analytic ones. The mind doesn’t have anything “in” it that you can see; each act of seeing is constructive. For example, you don’t “see your plans”, you have new thoughts that are related to previous thoughts which you may decide to think about linguistically as “plans”.

² Yes, the echoes of Hegel’s dialectics are obvious to me too.

Appendix: Examples

What appears to be essential to a mental entity is often a later addition, created during introspection. This can be difficult to wrap one’s mind around. In the interest of clarity, here are are a few examples of this inversion:

  1. Imagine an image of Tom Hanks in your mind. Does the image have ears? When asked this later question, your mind may zoom into the image and see ears there; possibly exaggerated ones. But it was the second thought-action that added ears (or nostrils, or hands), they weren’t there before.
  2. Is a given thought an opinion or a fact? This appears to be essential to the thought itself, but at the time you have a thought it is just that, a thought. Designating the type of thought is a later judgment call when someone else pressures you on it; assuming you haven’t preemptively interrogated yourself on the same topic. The same can be said of theory vs hypothesis: social context determines which it is.
  3. The expression “that is an instance of logical thinking”, where “logical” can also be replaced with “causal”, “utilitarian”, “political”, “irrational”, etc, is a misnomer. It would be more correct to say “that is an instance of thinking, which you later interpret as logical.
  4. Is a particular association one of cause-effect or object-attribute? The thought “ball → red” may seem inherently like an object-attribute connection, but originally it is just a thought-association. It takes time and education to figure out how to designate it. This becomes obvious when you consider more ambiguous instances like “government → organization”.
  5. Any thought which is not connected to other thoughts is unconscious, in the Jungian sense. Unconsciousness is not a property of the thought, but only of whether other thoughts easily connect to it. This example also demonstrates the motivated nature of thoughts; a person may repress a thought they don’t like. (This will be addressed in a later post.)



From Narrow To General AI

The road from Narrow AI to AGI presents both technical and philosophical challenges. This blog explores novel approaches and addresses longstanding questions.