Machinery for Generating Insights
Summary of posts 1 to 11: A theory of Artificial General Intelligence (AGI) must first outline the invisible, automatic units of mental behaviour that drive higher-level, complex ones. One such behaviour is “becoming aware of something”, either something external or within one’s own mind. “Becoming aware” is a problem-solving process that results in the formation of concrete memories or thoughts (sights and sounds). These can later be re-experienced as needed. The content you became aware of was a solution to a problem (aka a “tension”). A tension is a negative impulse that learns how to be removed, e.g. pain or unhappiness. Learned tensions are created when other tensions are repeatedly experienced without solution.
There is a feeling that accompanies moments of insight, when the pieces of a problem all seem to click into place. Your mind suddenly clears up; a fog is lifted, and subsequent thoughts run more smoothly. You no longer feel the drag of doubts, discomfort, worry, uncertainties, trepidation, and so on. Rather, you feel like you have “found the answer” to whatever it is you were troubled about. Your subsequent thoughts and decisions will have a different direction due this new insight.
What has happened to your mind? That is what this post will address. It continues the discussion begun in the previous post, by diving into the process by which the mind recognizes solutions as solutions. The above moment of insight represents, subjectively, the moment at which you create new criteria for what counts as a solution, that is, a new trigger for learning. Your mind has found an whole useful “space” of thoughts. As with the creation of tensions, this process is often subtle and its details may be invisible—you can only study it by its effects on other thoughts, and especially on learning.
Say you were asked to identify the first red object you see in front of you. The explicit solution to this task — i.e. the solution to this problem — is whatever red item you happen to notice. To notice it, however, depends on you being able to register that the item is red, or more specifically that a set of red visual inputs signals a solution to this problem. Both the tension and the criteria for the solution were acquired over time—you must at some point learn not only how to sense that there is a problem to be solved (i.e. to find a red object), but also that certain colour experiences solve that problem, and thus should trigger awareness and learning.
Now it’s entirely possible that the first thing you notice, on closer inspection, doesn’t actually solve the problem. If you were looking for your keys, a silvery object may catch your eye, and trigger an identifying image-thought of a key. The thought of the key solves the tension, and registers an “awareness”. But the items details are unknown (a new tension), and it needs to be identified. On looking closely you realize it is in fact a spoon, or perhaps the wrong key. Because it solved the problem (it is no longer unknown), the connection between the unknown silvery item you first saw, and the identified sight of the spoon up close now becomes a new moment of awareness, and consequently creates a new “memory”. As you continue to search through the room, that same silver object catches your eye. The newly connected memory of the spoon reappears, and because the spoon does not solve the problem, the tension persists. The mind continues looking and searching elsewhere. The third time you glimpse that same silver object, the earlier impulse to go towards it has been negated, replaced by the action to look away, to look elsewhere.
This simplified¹ example demonstrates how the learning loop can adapt to exceptional circumstances. Your first observation ended up a disappointment. The underlying problem wasn’t solved as expected by the silver object. What this meant is that on seeing it was a spoon, the thought that it was a key, which originally drove your actions, was negated and replaced with the memory/thought that it was a spoon. This further meant that the prior solution was broken, and the tension — e.g. “I am trying to leave my house and am not prepared” — returned. And so the search continued.
I mentioned the word “negated” above, and this term requires elaboration. “Negation” simply means that a thought or action pattern that was earlier learned, now gets overridden by a new thought or action. The specifics of the current situation demand their own unique resolution. This is akin to learning an irregular verb. Normally your desire to communicate that an action happened in the past tense is solved by adding “ed” at the end of that action’s word. This is driven by a desire to be understood, specifically the tension that you are not getting through to others. However, in the case of irregular verbs, when you attempt to apply that pattern (e.g. “eat-ed”) people look at you quizzically, re-triggering the tension that you have been not been understood. While that second tension is active, they may say the correct answer to you (“ate”), and their tone and context signals that you have heard the right word, thus your problem is solved. The next time you want to indicate that you ate in the past, that new thought/memory would reappear in your mind (“ate”). This will preempt and override the thought of “-ed”. The previous solution is negated in this specific circumstance.
The ability to learn to distinguish exceptional cases is necessary to have an adaptable mind. An assumption we’ve made during this series is that thoughts and actions are learned as repeatable responses to similar situations. But generalization brings with it the risk of over-generalization. Reality is nuanced, and rarely is any rule totally universal. A mechanism must exist to learn correct responses to exceptional situations, while allowing the general rule to otherwise continue. To put it simply, if a set of sensory inputs has a certain thought or action attached to it, then a superset or near-superset² of those inputs (which includes other, new inputs) will get its own response, inhibiting the previous one from getting activated. These exceptions may ultimately develop into their own rule. This is the process of “negation”.
What does negation have to do with learning how to recognize solutions to problems? Well, in the above, we presented negation as a means to negate thoughts and action patterns in specific situations. But why stop there? What if you could also negate tensions? What if you could learn that, in some exceptional situations, what you thought was a serious problem wasn’t really a problem?
Consider the following. If you found yourself on a high ledge, unsupported, your mind would likely trigger an immediate tension at the situation. This may have been learned from childhood experiences of falling from less extreme heights. On the other hand, if you were in an airplane looking out the window from the same height, that tension would likely be muted. This is an exceptional case; but rather than an exception to a thought or action, it is an exception to a tension.
In the simplest cases of problem-solving, removing the tension is enough to solve it. The essence of a tension is, indeed, to remove whatever inputs care causing it—and it learns any action or thought that allows it to do so. This is most obvious when it comes to the physical tensions, like pain or hunger. The removal of either is a relief, and the causing actions — or experiences in the case of thoughts — are learned.
However, humans have an added ability to remember and re-experience sights and sounds that are not immediately present. Thus a tension caused by a thought could stick around long after (or even before) it is experienced. Preoccupying concerns about paying bills, failing tests, and imagined heartaches all haunt us even when the immediate circumstances are non-threatening and banal. We do not have the option, like animals, to simply run away from our troubles, because troubled thoughts will come with us. The only other option to remove such tensions is to negate the tension — to learn to identify exceptional situations where the tension is no longer cause for alarm. So when a teacher pulls you up in front of class to answer a question, your degree of stress (tension) will be significantly lessened if you believe you know the answer.
We’ll dig into how you actually solve problems once the solutions are in place in a subsequent post. For the moment, let’s focus on how it comes about that you learn to negate a tension. Let’s start with what we’ve already discussed above, since the general pattern of negation is the same regardless of what you are negating.
The presence of a tension is an indication that the current situation is a “dire” one. You have previously learned that, unless dealt with, this situation will lead you into another, worse tension, and this latter tension will be “inescapable”. Your mind is therefore signalling a useful early warning while there is still time to avoid it. It’s like a sign above an exit lane indicating that this is your last chance to change lanes before the on-ramp, at which point you will have to get on the highway and will not be able to leave it for a while. All this, of course, assumes that the tension (the sign in this case) is only a problem if you actually don’t want to be on the highway. In other words, the child tension serves a parent tension.
Now if there are exceptional situations in which the parent tension could be easily resolved, the child tension need not actually be addressed or cause fear. For example, if you see a sign indicating that the right lane exits to the highway, but then look and notice that you are in the left lane, the tension is already resolved — even without you taking action. In this case, the tension (right lane exits) can be immediately followed by a resolution (I am in the left lane).
Indeed were you in the right lane, any action you take that got you to experience the latter situation, namely moving into the left lane (and therefore experiencing that “I am in the left lane”) would count as a resolution and would therefore be learned. Your goal when faced with a problem (tension) is to find a path — in action or thought — to connect the tension to its negation. Even if you only planned (imagined) moving into the left lane, that would count as a solution-thought and therefore that set of thoughts would be remembered, and would become a thought that served as a plan.
To describe the process summarily: a child tension is created when a parent tension appears and persists as though inescapable. However, if the child tension activates, but the parent tension either doesn’t appear, or appears but is easily solved, this should count as a “false alarm” signal — the situation is not as dire as it was expected to be. If, as a beginner driver, you see a “right lane exists” sign, you may begin to panic that you are about to go on the highway. If then, you do not find yourself on the highway (since you were in the left lane), the presence of the sign plus the sight of yourself in the left lane negates the tension. You now have a template for a solution.
This itself is not enough, though, because the negating experience may have been a one-off, a rare exception, and unlikely to repeat. A true solution template is one that could be confidently generalized to many cases—or as confidently as is possible in a chaotic world. Some confirmation is needed that what you have found is general.
Aside: In case some readers are just joining this series, all these mechanisms must be automatic. No conscious choice must be involved, and in no cases must the mind deviate from the pattern described. We are not looking at what the mind should do—e.g. if the mind were smart or rational — but what the mind will do, in every case, regardless of the specifics of the content. Only then will we have discovered truly general rules for AGI.
The content of the matter can’t be brought into consideration at any point in the process. If the brain were distinguishing between cases of, say, writing English vs driving a car, that means it is parsing and sorting the content in a way that the biology of the brain couldn’t possibly account for innately — cars and writing are evolutionarily recent inventions. And even if the brain did make a distinction at a biological level, it would do so by building in circuitry to differentiate between writing and driving. This would qualify them, by definition, as different “types” of signals in the mind. All that the brain can consider in its fundamental, unconscious processes is the type of the mental events and their timing.
So we are looking for criteria that would be sufficient to count as confirmation of a general solution, that the mind can use to signal that the problem is negated. If all that we can use is the type and timing, it is likely something to do with the timing of the parent and child tensions, as well as the candidate solution, since there are no other signals relevant to the present tension that could be used to make the decision.
Going back to the example of the exit sign, say you learned that if you are in the left lane at that particular on-ramp, you need only stay in that lane. If this situation were to repeat again precisely, at the same location and in the same general surroundings, the negating solution would be elicited, and you would skip over both the child and parent tensions. This, however, means you would have no way to confirm its generality. The situation must be different enough so that it counts as new, but the solution, namely being in the left lane, must still be the same or similar. This is what confirms a solution as general — it is not that it has been applied more than once, it is that it has worked on the same problem, in multiple different situations. The specifics of the context can then be removed from the equation, and only the solution remains — reconfirmed for that problem. This process is the foundation for that “moment of insight” with which we started the post. Through it, you have discovered something new, something important, something deeply connected to the way the world works outside of specific spaces and times— a new rule or law of living.
There are still a few questions remaining before we can wrap up the topic of motivations. So far the examples have been about concrete problems, but have not yet addressed more abstract problem solving. In addition, we’ve used words like “activate”, “elicited”, “similar”, and “superset” without much elaboration. For a model of AGI to be truly comprehensive these still require precise, technical definitions. There are also some gaps which you perhaps haven’t noticed, such as how the mind can focus on one topic or domain and all its related issues, or how learning from one domain can be transferred to another. These, and more, will be addressed in upcoming posts.
Next post: Free Will and Agency
¹ These are a select few tension-solution pairs that get activated during a thorough search. Others would include, say the desire to lift a pillow and look under it, a thought of where you last saw it followed by a search through memories, even unconnected thoughts like the movie you saw yesterday, etc. It also says nothing of the intervening actions that carry out the search. Adult lives are full thousands, possibly tens of thousands of such tensions and solutions, which flit in and out of your mind as they are triggered and solved.
² The term near-superset is meant to indicate that the inputs are not a strict superset. Because pattern matching of inputs between the time that they are learned and when they are re-experienced will never be exact, a close-enough match will still activate a solution, tension, etc. This also means that when you learn to negate it, not all inputs of the original pattern may be present, and thus the negating set may be missing some of the originals —and adding some others. The details of this math will be addressed in a later post.