A.I. That Thinks Creatively, Part 1

From Narrow To General AI
7 min readApr 5, 2020

Part 2, Part 3, Part 4, Part 5

As you read these words, you are also hearing a voice speak them in your head. This voice isn’t coming from your ears. You are adding sounds to the written words, sounds you create in your thoughts.

Try this experiment. Below is the muddled name of a famous historical figure. Say it over and over in your head until you can decipher it:


You’ve likely never seen that spelling before. So how did you figure it out? It was simple. You sounded it out in your mind, listened to what it sounded like, and in your mind you heard something meaningful.

Take a moment to realize how extraordinary that is. I managed to get you to think about this person without ever writing any word that could possibly be associated with him or her. To use terminology from Machine Learning, you could never have been trained from samples to make that association. You created it anew for yourself.

How did you do this?


Thoughts are sights and sounds that you create for yourself. You do this when you daydream, or when you make plans. Although many everyday experiences come through your eyes and ears, many of your experiences come not from the world, but from your own thoughts.

Where do these thoughts come from? Let’s dig into their causes, since doing so will help us understand more about the human imagination.

While reading this article, a voice in your head speaks the words to you in English. This is because at some point in the past, you learned to read English, and you remember how each written word sounds. Thoughts that you produce in your head you generally learned from some previous experience. You sometimes call these thoughts “memories”, especially when they are memories of experiences you had through your eyes and ears.

When you want to remember something you first store an experience in your mind, and then recreate it later as a thought. This lets you experience it a second or a third time, this time as a self-generated experience.

Where does imagination come into play here? Well, if thoughts are experiences you generate for yourself, and you can remember your previous experiences and recreate them as new experiences (thoughts) later, what’s to stop you from remembering your own thoughts in the same way you remember experiences from the world?

This is in fact what you do. For example, without looking above, can you remember the strangely spelled name of the historical figure you read earlier in this article, the one that starts with “L”? Since you never heard the words spoken out loud, any memory you have of the sound of his name must be a memory of sounds you created in your head. It is a memory of your own thoughts.

So here’s the whole loop: You first have experiences through your eyes and ears, and you remember them. You later recreate those experiences in your mind as thoughts. Finally you remember those thoughts as if they had happened to you in real life. These new memories can also be remembered and recreated, and so on, without end.


What if you got more creative with this ability?

For example, if you try to remember a real boat that you once saw, you could only remember the boat as floating on water, since that’s what boats do.

But if you think of the sky, and you think of a boat at the same time, in your mind you’ll see the sight of a boat floating across the sky. That’s what you’ll remember. That’s a brand new image, a new thought, a new memory.

When you have two thoughts at about the same time, you experience them together, and remember them together.

You can then bring up this new memory anytime you like as a thought. Your ability to put together and remember your own thoughts extends the limits of your experiences to things that are impossible in the world. You call this your “imagination”, and it is the canvas on which you practice your creativity. It allows you to experience things that you could never experience in real life.

When you think creatively, you create new memories out of the raw material of your thoughts. You put thoughts together in haphazard ways (like a boat floating across the sky), and occasionally some of these prove interesting.

Since this new image is also a new thought, you can continue adding it together with other thoughts. For instance, you can add a racoon in the boat.

This works for moving images (‘video’ memories) as well. In fact, any experience, sight or sound, including ones you create yourself can be remembered and thought up again. The sum-total of all these self-generated experiences makes up your thoughts. In fact it makes up your whole conscious experience.

It may seem strange to say that all thoughts must be memories of something you previously experienced. But when you understand that in addition to remembering what you see through your eyes and ears, you can also remember your own thoughts themselves, that hesitation greatly fades.

Imagination in A.I.

If you’ve followed along so far, you might have an idea of how to add imagination to an A.I. A computer could remember what it saw, then recreate it later in its imagination, almost like a hallucination. It could then remember those new, combined thoughts as well. All that is required is to connect the output of its mind right back to the input. This lets a computer come up with and explore new ideas.

The more technical-minded reader may notice the above diagram has some similarities to Recurrent Neural Nets, with the additional ability to combine multiple, distinct thoughts together into one thought.

The A.I. would separately learn what “boat” and “sky” look like, and can imagine them separately. Then, when it reads a sentence that combines the two, its thoughts also combine.

Thoughts have to be useful. Enabling an A.I. to imagine sights and sounds is only the first step to a full creative imagination. In order for new thoughts to be called creative they must also solve real and imagined problems. As you already saw in the article on awareness, memories are only formed when they solve your immediate problem. They must answer a question or solve a problem.

Thoughts are also best when they are generally applicable. As you can read about in Day 10: Generalizing From Specific Examples, and A.I. can select what is common across its experiences, and generalize those if they prove useful. It keeps the parts of the experience that are common, and prunes away the rest.


Such an A.I. only stores memories when they solve problems, and generalizes common patterns throughout its memories into knowledge. In so doing, it makes sure its thoughts are useful, grounded in reality, and significant.

The thoughts that such an A.I., and you, have follow patterns. They are combinations of previous experiences. This means that, to a degree, they are connected to what is real. You saw this in the case of the boat above. Even in the last stage, where a boat was floating in the sky with a racoon, the images were still connected to reality, in that boats, the sky, and racoons truly exist.

As you add more levels of thought on top of one another, combining thoughts into new thoughts, the resulting similarity to reality fades. Somewhere in the middle is the sweet spot — where there is just enough connection to reality to be practicable, and just enough freedom to be creative.

In this article you’ve seen how memories and thoughts are the same thing, self -generated experiences. The question arises: how do you learn to associate certain situations with certain imagined thoughts? In the next section you’ll see how memories are formed, and how they can be made useful to an A.I.

In this series:

Part 1: Imagination

Part 2: Solving Useful Problems

Part 3: Motivation

Part 4: Defining Skills

Part 5: Thoughts: Knowledge and Intents

Are you also interested in applying Artificial Intelligence to human creativity, human understanding, even human values? Do you feel that our current goals with A.I. are limited? I’m looking to connect with others who have a similarly ambitious vision of the future of A.I., whose goal is to tap the full creative potential of human intelligence through software.



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.