Concepts, Part 1

How A.I. Can Define Its Own Concepts

10 min readApr 5, 2020


To do useful things in the world your mind makes use of a catalogue of concepts. You use concepts when you reason about and solve problems. For example:

“How can I find the right enzyme to destroy this virus?”

Solving a problem like this one already presupposes a handful of concepts like enzyme, destroy, and virus. You also need concepts to think logically:

“This virus can’t survive without Protein A.”

“Enzyme B kills Protein A.”

Therefore: “Enzyme B would destroy this virus.”

If A.I. is to ever solve problems, it must also reason logically using concepts. To do this, we have to understand what concepts are, and how to create them in software.

Discovering the origin of concepts is not as straightforward as it seems. Take a simple concept like chair. Who came up with it? Did you invent this concept then attach the standard English word to it? Or was the concept given to you somehow? If it was given to you, who originally invented it?

As far as we know, our species came up with concepts like virus and enzyme without external help. So how can an A.I. invent concepts without human help, then decide what each concept means? There are an infinite number of concepts it could define, and an infinite number of meanings each could have.

In part 1 of this article, I’ll explain where concepts come from and how they drive your thinking and actions without you being aware of them. In part 2 I’ll explain how you become aware of concepts, and use them to reason logically about problems. Throughout both parts, you’ll see how concepts can be mechanized, that is, systematized to the point where they can be imitated by an A.I.

Let’s start by understanding where concepts come from.

Concepts Are Subjective

You may have heard that the technical definition of fruit includes tomatoes. Despite this, it’s likely your personal, or subjective concept of fruit doesn’t include tomatoes. This distinction is important since your subjective concept is what you use in day-to-day life to organize your actions and thoughts.

The reason you find it difficult to think of tomatoes as fruit is because tomatoes don’t meet your underlying goals behind fruit. Your motivation behind the concept fruit is that you want a sweet, desert-like snack. If you’re putting together a fruit salad, the scientific definition does you no good.

Every one of your subjective concepts is built on an underlying motivation or goal. You use this motivation to decide if a given case should be part of the concept or not.

You can think of concepts in two different ways. The first is your personal, subjective concepts. The second is what others tell you the definition of a concept is. This latter includes dictionary entries and scientific definitions.

Your definitions may not align with the ones you get from others. When someone else tells you how their version of a concept is defined, and it doesn’t match your own motivations, it has little meaning for you, except as a fact to be memorized. You don’t need to understand it in a deeper way. You can learn that “guava” is a type of fruit without even knowing what a guava is or what it looks like. It’s just a sound.

In contrast, you always deeply understand subjective concepts because your mind produced them in the first place. They are each intimately tied to your personal goals and motivations.

Every subjective concept, whether it gets attached to a word or not, starts out in a person’s mind as a motivation. The concept of food is motivated by hunger. The concept of bravery is motivated by fear. The concept of society is motivated by the need for belonging.

These motivations work through your actions long before you consciously realize them and give them names. For instance, you probably didn’t realize how often you felt schadenfreude until someone told you the word for it.


Put simply, a motivation is something you dislike, usually combined with something you aim towards. Concepts exist between these two points. For example, between “hungry” and “full” is food. Anything that gets you from hunger to satiety is part of the concept of food. A concept is the set of actions or thoughts that get you from what you dislike to what you want.

It can often be difficult to see the motivation behind a given concept. As you saw in A.I. that Thinks Creatively, Part 2, the motivation behind a thought or action usually fades once the thought or action is learned. It frequently differs from any colloquial explanation you may hear. On the other hand, once someone tells you the motivation behind a concept, it usually feels obvious.

Let’s go back to the concept of schadenfreude. Schadenfreude is ‘the secret feeling of joy in watching other people fail’. What is the underlying motivation behind it? Your impulse to assume that humans are cruel by nature may readily spring to mind. This is often the proposed explanation, and it’s hard to think of another.

But you feel schadenfreude more strongly at some people’s failures than at others, so this explanation seems deficient. Rather, begin by realizing that you, like everyone, are competitive, and that you are competing socially all the time — to spread your own beliefs and values over others, for popularity, for money and career advancement, for love, etc. In a competitive situation, others’ failure often helps your own success. A co-worker not getting the coveted promotion leaves it open for you. You feel schadenfreude strongest when the failing person or their ideas most strongly competes with what you want.

Seems obvious in hindsight, doesn’t it? When you understand the real motivation behind a particular concept you have understood it as deeply as possible.

Now let’s see how this plays itself out in practice.

  1. You see someone competing against your ideas or your personal advantage.
  2. You are having difficulty competing (something you dislike).
  3. You see that person fail.
  4. You spot some advantage to yourself, even if only a vague moral advantage, by this failure (something you like).
  5. The thought of that person failing now becomes a “solution” to your problem, and you gleefully promote it to others.

You may forget all about steps 1, 2 and 4, and only remember that there was someone who failed, and the glee you felt about it. You might even experience remorse over how this whole affair makes you look — but that is a different problem, the problem of shame.

The above pattern may happen many times and in many different circumstances over the course of your life. The details of the person and their failure will change, but the underlying motives remain consistent. Looking through your memories you might group these together as moments where “I felt joy in someone else’s failure”. Then one day, someone gives you a word to attach to it.

For A.I. to invent concepts it needs to develop its own motivations, and to solve problems related to those. You can see in detail how both happen here and here, respectively.

Concepts Vary by Person

Since concepts are driven by individual motives, the content of each subjective concept, and the number of concepts varies greatly from person to person. They are influenced by your goals, experiences, and circumstances. Furthermore, a single word may have many overlapping motivations underneath it.

For instance, the scientific definition of “animal” includes humans as well as insects and spiders. But if a friend told you there was an animal in your house, despite the fact that the majority of the time it would be an insect, spider or human, those would likely not be the first “animals” that come to mind. Except in special circumstances, few people use the word “animal” to refer to humans. Your motivations related to people are different from those you have in relation to animals.

Even non-human animals tend to be grouped into domestic animals (motivation: nurturing and companionship), farm-animals (motivation: food), and all remaining animals except insects and spiders. Insects and spiders are grouped under bugs, and their name says it all: they bug you, and you’d like to get rid of them.

However, if you were raised on a farm, and grew attached to a particular lamb, you might consider sheep as domestic animals alongside cats and dogs. I’ve known people who approach tarantulas and snakes in a nurturing way. A cattle farmer may have a separate concept just for cattle, which are special to her.

The ever-shifting boundaries between subjective concepts means it’s impossible to create a complete list of all of them. An A.I. must define its own concepts out of its experiences and unique motivations. Its circumstances determine what is a useful motivation for it to have. A word only gets attached to a concept in order to speed up communication, by finding commonalities in people’s subjective concepts.

The challenge of clearly defining concepts compounds further when you consider abstract concepts like love, justice, good, or bad. Life is messy and complicated. Sometimes you’re confused as to which of these concepts a given experience fits into. You might change your opinion from moment to moment, and your opinions may contradict each other in confusing ways. You see this with endless discussions over whether a hotdog is a sandwich.

The reason for this confusion is because your motivations solve problems on a case by case basis, without worrying if they are perfectly consistent with other decisions you’ve made. Fortunately, even without clear definitions, most of us do alright. It turns out you don’t need to clearly define concepts in order to act on the motivations behind them. Attaching a word or label to a set of experiences, or clarifying the boundaries of what experiences should belong to that label, are, in every case, optional. Concepts, at least in your mind, can exist just fine without clarity, awareness, or definition.


Many useful concepts are abstract. Concepts like profit, the future, society, trust, and possess are crucial for business and social interactions. If A.I. is to be truly useful it must have the ability to define and work with these too.

Fortunately, thinking of concepts in terms of their underlying motivations makes it easier to understand where abstract concepts come from. Take for instance the concept window.

Although window might seem like a concrete concept, there’s a lot of nuance hidden in it. You can use window in an abstract way such as “the eyes are the windows to the soul” or “a window of opportunity”. To understand why, look at the underlying motivation behind the concept.

A physical window, like one in a building, has a purpose. It brings much-desired light and air into a closed room. In the same way, a “window of opportunity” is a small gap in an otherwise uninterrupted obstacle, during which you can achieve your goals.

In both cases, there is a persistent, long-lasting obstacle that doesn’t let something you want through. This is the wall. Within this wall, there is a finite, narrow gap that does let it through. So the underlying motivation behind window is to let things through that are otherwise blocked.

Connecting the definition of a concept to its underlying motivation helps explain why some things that would otherwise be called windows are not. For instance, in a solarium where all the walls are made of glass, you may hesitate to call them windows, even though materially-speaking they are. You are more likely to refer to them as “glass walls”. In fact, I did that two sentences ago, and you didn’t mind. This is because for something to be called a window it needs an impenetrable wall around it to give it meaning. If one of those glass panes swung open to let in fresh air while the others didn’t, you might call that the window.

To a contractor who is installing the solarium the glass walls may all be windows. This is because his motivations are different from yours. He cares more about the materials and shape, not whether the windows are a means of something getting in or out. That is what his job requires him to care about. The same word can be used in different ways, even by the same person, if the circumstances and motivations are different.

Understanding the concept of window in terms of your motivations also explains where the expression “the eyes are the windows of the soul” comes from. Whereas the rest of the human body doesn’t give much insight into a person’s character and thoughts (the wall), the eyes may let some signals through (the opening). It goes without saying that the person who came up with this expression was trying to see into people’s souls through their external features, but was blocked at every turn, at the elbow, the ankles, the nostrils. Then he or she found that the eyes let something special through.

Once you realize that every subjective concept is centred around a motivation, you can apply it to new instances as well. When you are in heavy traffic trying to merge onto the highway, is a gap in the cars a window? At a crowded concert, where it’s difficult to see, there may be a space between the heads of the people in front of you that offers an unobstructed view of the stage. Is that a window?

Because all concepts are centred around motivations, all concepts are inherently abstract concepts. You saw above how window was used in both concrete and abstract ways. The same is true of any physical object or concrete idea. This is why you can use concrete concepts as metaphors, such as in the Indian phrase “he is a tiger among men”.

And because subjective concepts are not defined by their physical appearance but by your motivations, they are as flexible and dynamic as your motivations are. In this sense, you are always defining your world according to your needs.

Concepts in A.I.

Until now, every major attempt to get A.I. to understand concepts has assumed that the A.I.’s concepts should match our own. Expert or logical systems can reason and think about ideas and their connections, but those ideas ultimately come from us. Classifiers that learn to automatically catalogue images of birthdays and beaches assume that those two concepts are the important ones to look for, rather than say, carbon, or palindrome, or something else for which we currently have no name or concept.

Without including motivations and problem solving into A.I., the result is at best an academic exercise of equating sounds with other sounds, images with images. It is a colour-by-numbers approach to understanding the world. The result is predetermined before the exercise begins.

As humans, our creativity has its roots in our subjective concepts, because these are the ones that drive us to act and learn in new and useful ways. The same must be true for A.I.

In part 2 I’ll talk about how we apply concepts when in logic and reasoning.

Are you also working on applying human creativity, human understanding, even human values to Artificial Intelligence? I’m looking to connect with others who have a similarly ambitious vision of the future of A.I., who want to tap the full creative potential of human intelligence, in software.




I lead an applied A.I. team at a robotics company. On the side I develop A.I. that uses motivated reasoning and invents abstract concepts from scratch.