Cover art for ‘04 – How Do We Know?’

04 – How Do We Know?

Artificial Creativity

Listen on Apple Podcasts Listen on Google Podcasts Listen on Spotify

References:

David Deutsch, “The Beginning of Infinity”, chapters:

  • 1 “The Reach of Explanations”
  • 4 “Creation”
  • 15 “The Evolution of Culture”
  • 16 “The Evolution of Creativity”

Further reading:

  • Karl Popper, “Back to the Presocratics”

If you have any questions or comments, please leave them here or tweet at me: twitter.com/dchackethal

Subscribe to this podcast via RSS Back to all episodes

Transcript

This transcript was generated automatically and may contain mistakes.

Welcome back to the podcast on Artificial Creativity. This is episode 4, this time about the question, how do we know? In the previous episode, we spoke about how AGI research is currently in a rather sorry state. We explained why we haven’t made much progress toward AGI, largely by invoking the concept of universality. It’s also how we explained that AGI can’t be built merely by adding more and more features, and it allowed us to explain that for neuroscience research to contribute to the field, it would need to violate computational universality. We concluded that building artificial creativity is a philosophical problem in the field of knowledge creation, meaning epistemology, and that due to existing misconceptions in philosophy and AI research, software engineers who have not yet done work on AI or philosophy may have the best chance of making meaningful contributions.

In this episode, we will attempt to answer the question of how we know. This question is also deeply connected with the question of what knowledge is. In order to answer both, we need to take a brief excursion into the world of biology, but I promise that in the end, we will bring it home and relate it to the problem of artificial creativity.

The question of how we know is a very old problem, and it arose in combination with the problem of change. Philosophers in ancient Greece noticed that some changes escape our senses, and yet we know that they occur. For example, we do not see our children growing up, and yet we know that they do. This is a surprisingly simple yet deep realization, and it already hints at how limited a role the senses play in the creation of knowledge, and in fact how they can sometimes hinder it. For more on the problem of change and how it relates to the problem of knowledge, I recommend you read Karl Popper’s “Back to the Presocratics”. This is a really fun and inspiring essay, so I highly recommend it.

Human knowledge takes the form of explanations. Explanations are statements about what is really there, how it works, and why. So knowledge is not just an accumulation of statements or facts. For example, a history book containing only facts about a war and the corresponding dates by no means explains that war. Explanations are created when we solve problems. A problem is not necessarily some tragedy. It’s merely when an existing theory makes a prediction that does not come true. So as a result, there’s a conflict between the interpretation of an observation, which is also a theory, and the original explanation.

For example, a conjurer may perform a trick that makes it look as though a ball is floating in midair. This would violate one of the more basic explanations in your mind about the world. Your best, probably largely intuitive, explanation of gravity says that the ball should fall to the ground. The problem is that you don’t observe it falling. This is something that needs to be explained. There needs to be a new explanation that does not have this conflict and that still explains everything the old theory was able to explain. So the new theory has to contain the old explanation of gravity plus it needs to explain the problematic observation. One viable, though in this case not very detailed, explanation is that the conjurer is performing a trick and that the laws of physics and gravity still hold.

The reason we look for explanations that do not have the conflict is because our theories are accounts of how reality works, and no such conflicts exist in reality. Therefore, any conflict between theories shows that at least one of them is false.

Consider another example. You walk home one day and your car is missing. You remember parking it there the night before, so according to your best theories, it should still be there. How can you explain what happened? All you can do is guess. Maybe it was in the way of something and had to be towed. This is a conjecture theory that makes a prediction. You should be able to call the local towing company to see if they have your car. If you call them and they don’t, the theory is false, and you need to conjecture another theory. Maybe it was stolen. And all this is as true for missing cars as it is for the inner workings of stars, the mechanisms of curing a disease, and how to write a computer program. For all of them, we need explanations. So where did this explanation of the missing car come from just now?

For much of history, it was thought that we derive knowledge from our senses. This is what’s known as empiricism. But empiricism cannot be true, because empiricism itself is not derived from the senses, and so it rules itself out. This is a simple but powerful refutation. Empiricists tend to ignore it and prefer to marvel at the mysteries of how we know anyway, instead of looking for a better theory.

According to our best theories, knowledge is guessed. The origin of all human knowledge is bold conjectures. As David Deutsch writes in chapter one of “The Beginning of Infinity”, we create new knowledge by, quote, rearranging, combining, altering, and adding to existing ideas with the intention of improving upon them, end quote. This is what creativity is. Senses do play a role, but only when we test our theories. Because theories are the result of guesswork, we should only ever adopt them tentatively. And since we as humans make mistakes, we should expect even our best knowledge to contain mistakes in addition to truth. Therefore, we should always expect to find more problems with our theories, and those theories to be superseded by even better explanations. As long as we continue to look for problems, this process will continue forever and this is how unbounded progress can be made.

So what is knowledge? Knowledge is a kind of information, but it’s not just any information, it’s information that is adapted to a particular purpose. What does it mean for something to be adapted to a particular purpose? It means that only few changes would make it better and it would be very hard to change it without making it perform worse at fulfilling that purpose. I want to explain why this matters.

Popper suggested that for a theory to be considered scientific, it needs to make testable predictions. But as David explains, testability can’t be sufficient. Consider the following example. If somebody told you that thunder was caused by the Greek god Zeus angrily throwing bolts of lightning down to earth from up on Mount Olympus, would you consider it a scientific theory? Sure enough, it explains the cause of lightning, albeit poorly. And it is testable. For example, it predicts that if we climbed up Mount Olympus, we could see and meet Zeus there. Presumably, we could even talk to him. But I don’t think anyone in their right mind would actually send an expedition.

I think Popper understood this. For example, he said that origin myths, which often invoke gods and the like, as whimsical as they may be, were fine starting points to explain the origin of the world, as long as we are willing to ask some awkward questions. What is the purpose of awkward questions? It is criticism. And criticism can precede tests. And we can criticize the Zeus theory of lightning by saying that it is too easy to change some of its details without destroying the entire theory. For example, why Zeus and not a wizard? Why is he angry? And why on Mount Olympus and not another mountain? Greek myths may have placed him there, but the theory would still work if he threw lightning off of other mountains.

This kind of explanation is what David calls easy to vary. We can easily change its components without hurting its explanatory power. This means, by definition, that this explanation is not adapted to the purpose of explaining lightning. Therefore, we shouldn’t consider it scientific, because it is a bad explanation.

Our best explanation of lightning is hard to vary. Roughly, it explains lightning as an electric current that is the result of electrical charge building up in a cloud due to friction between ice crystals. These crystals form because some clouds are high up enough that it gets sufficiently cold. Change any part of this theory, and it falls apart. Remove the ice crystals, and you don’t have friction. And without friction, you don’t get the necessary electrical charge. And without connected theories of weather and our atmosphere, that are all hard to vary in their own right, you don’t get clouds at all. So in a somewhat eerie way, this theory resists change. This is also why, when it is told from one person to the next, it is transmitted faithfully, even though neither party necessarily cares about its faithful replication.

Also, this explanation of lightning has what David calls reach. It transcends our own atmosphere and predicts that we should find lightning on other planets with similar atmospheres. And we do. On Jupiter as well as Venus. Although interestingly, on Venus, the crystals are made not of water, but sulfuric acid. The explanation of Zeus does not have reach. If he is on Mount Olympus, presumably he can only throw lightning far enough to cause it in Greece, unless he has some superhuman throwing abilities to cause lightning on the other side of the planet, which would be another bad explanation.

Again, bad explanations are easy to vary and good explanations are hard to vary, and being hard to vary is a desirable feature of explanations because it means we’re getting closer to the truth. Reality is impossible to vary. On the topic of lightning, Popper was struck by lightning once while hiking in the Austrian Alps, but that’s just a random factoid, but luckily he survived.

To find out more about the concept of hard-to-vary explanations and the reach of explanations, I recommend you read chapter one of “The Beginning of Infinity” called “The Reach of Explanations”. I know I’ve recommended this chapter a lot by now, but I should mention it again because invoking conjuring earlier and Greek myths just now are both devices I borrowed from that chapter for this episode, and David explains it much better and in more detail.

Now, there is a peculiar thing about the creation of human knowledge in that it is a continuation of a much older biological process. Genes also contain information. And since genes evolve, meaning there are random mutations and selections, genetic information is also adapted. Therefore, it also constitutes knowledge. Contrary to human knowledge, however, genetic information is not explanatory. From “The Beginning of Infinity”, chapter 4, “Creation”, quote, the knowledge embodied in genes is knowledge of how to get themselves replicated at the expense of their rivals, end quote.

Now sometimes, in order to do this, genes may impart useful functionality to their organisms and so may contain regularities of the environment and even approximations to the laws of physics. But really they are there because they help genes suppress rival genes and replicate more than them. That’s why knowledge in genes is not explanatory, and when it happens to contain explanations, it is not adapted to the purpose of explaining, but only to the purpose of spreading through the population. For example, a bird knows how to fly, but the knowledge in the genes of how to fly is not adapted to the purpose of flying or explaining how to fly, but only to the purpose of spreading through the population.

Or perhaps we could phrase it this way, though I don’t know if David would agree with me here. It is adapted to explaining only one thing, and that’s always the same. It’s adapted to explaining how to replicate itself and spread through the population in the particular niche of the environment the gene finds itself in. But it doesn’t explain what is really out there and why.

Genes encounter problems much like we do. Changes in the environment can make it harder to spread through the population and suppress rival genes. A genetic mutation can be an attempt at solving that problem. By using the word attempt, I don’t mean to attribute any agency here. A mutation occurs randomly and without regard for the problem situation. It always only happens to solve a problem. Now I think that what’s interesting from the perspective of the gene is that problems are not just about replication per se, they are about how the gene can create and turn into the thing that will make it spread through the population.

Since theories, and as you may have noticed I use the terms theory, knowledge, and explanation interchangeably. Since theories are created in response to problems, and problems are born out of mistakes in existing theories, it’s interesting to think about what the very first problem and the very first theory may have been. They must have arisen sort of simultaneously. My guess is that the first problem had to do with the question of how to replicate faithfully, and that the first theory was a theory about faithful replication, which was created by the precursors to life as we know today, but I don’t really know.

Another difference between genetic knowledge and human knowledge is that human knowledge is exosomatic. Popper liked to explain this by considering the difference between Einstein and an amoeba. When an amoeba procreates and the knowledge in its genes is altered through random variation, its copy will most likely die. But when Einstein has a new idea, he can explore it, and if it turns out not to be viable, he can just discard it without dying himself. This is why Popper said that we can let our ideas die in our place.

It’s also why, contrary to biological adaptations, not every explanation along the way has to work. Before discarding non-viable explanations, we can explore them anyway and go through a valley of bad explanations, as it were, without dying on the way to a good one. This is why biological knowledge tends to be very limited, whereas explanatory knowledge can have great or even unlimited reach. And it’s also why good explanations are discrete and why you can’t just combine two good explanations to get another good one. That requires additional creative thought.

The remark about letting our ideas die in our place is one of those things Popper said that sounds very casual and easy to agree with. The first time I read it, I remember thinking, oh sure, we shouldn’t let, you know, we shouldn’t kill each other because of our ideologies or something along those lines, but I think it means that too, but I realized that it’s a deep consequence of just that fact, the fact that the creation of human knowledge is an exosomatic continuation of the biological creation of knowledge.

Human knowledge, like genetic knowledge, also grows by the process of evolution because it has counterparts to variation and selection of genes. Human knowledge evolves through conjecture and criticism. So conjecture is the counterpart to random mutation of genes, and criticism is the counterpart to selection. Evolution is the only process we know of that creates knowledge.

To learn more about this, I suggest you read chapter 4 of “The Beginning of Infinity”, which I referenced earlier, called “Creation”. It explains the creation of knowledge in much more detail, and it also goes into the concept of replicators and how knowledge is an abstract replicator because once it is instantiated in a suitable physical environment, it tends to remain so and contributes to causing its own replication. That chapter also goes into more detail about the difference between knowledge in genes and human knowledge.

As a quick aside, since, as I said, the creation of human knowledge is a continuation of a biological process, it is interesting to think about what started it. Since the creation of human knowledge is the result of an evolutionary process in our minds, my guess is that it originated from some other evolutionary process in our ancestors’ minds, which initially had some very limited purpose. For example, it could have been an evolutionary pathfinding algorithm. So in order to make their way from A to B, our ancestors may have guessed different possible paths and criticized them in order to find the best one. Some sort of jump to universality may have happened where instead of only working for paths, suddenly it worked for any sort of data. But it is not clear what was being improved that suddenly jumped to universality, so this is highly speculative. For a different account of how creativity may have evolved in humans, I suggest you read chapter 16, “The Evolution of Creativity”. I highly recommend that chapter in any case.

But I’d like to get back to human knowledge now. So to go back to our example of the missing car, imagine you have conjectured that your car was stolen. You’re all shaken and you want to tell someone about it, and so you call your best friend. What he needs to do as he is listening to you is replicate that theory you have based on the limited information you give him. By doing so, he actively creates knowledge, and that requires creativity. That’s why so-called recipients of knowledge are never passive. They have to actively create.

What you conjectured as an explanation for the visual of the missing car, he now has to recreate by different means, by listening to your voice. All you and he ever get is clues as to what may have happened, something unseen happened that you need to explain with the seen. And your friend is in that same boat when you tell him about it, and with some luck both you and he will have roughly the same theory in your minds at the end of the conversation, although there is no way of knowing this for sure. If he doesn’t, that’s what constitutes a misunderstanding.

However, your friend is in a slightly different boat than you were when you first saw that your car was missing, right? When you created the knowledge of what may have happened to your car, you had to create it afresh from clues in the environment, and as your friend is listening to you explain it to him, he just needs to hear your words to recreate that same knowledge, right? Well, somewhat, but he too can only go off of clues in your words and the knowledge he already has of you and your car. So there is no difference in the mechanism that is needed for you to come up with an explanation in the first place and for your friend to understand that explanation. This is why creating a new explanation and understanding an existing one require the same mechanism. They’re the same thing.

This has interesting implications for the concept of originality, by the way. To understand a story, for example, or even to copy someone else’s behavior, we need to explain it. And in each of those cases, it’s a, quote, special case of the general human objective of explaining the world, end quote. This is from the chapter “The Evolution of Creativity”. This is why I said in episode two that building an AGI is a continuation of a cosmological effort.

A person trying to understand a story faces the same challenge a scientist does. Quote, both must discover a hidden explanation. For the former, it is an idea in the minds of other people. For the latter, a regularity or law of nature. Neither person has direct access to this explanation, but both have access to evidence with which explanations can be tested: the observed behavior of people and physical phenomena conforming to the law, end quote. This is also from the chapter “The Evolution of Creativity”. And an engineer is in the same boat when he builds a bridge or a software program only from a list of specifications. All of these things are about creating or recreating the unseen from the seen. And whether you are copying or innovating, it’s the same mechanism.

By listening to this podcast, you are being creative as well. You are listening and guessing what I mean by what I am saying. In order to understand me, you have to replicate my knowledge. You are using creativity right now. To understand anything, we have to replicate that thing in our minds. This is why understanding something is a replication effort, and replication is reminiscent of genes again. That is why I guess that gene replication is fundamentally also about creating the unseen from the seen, namely creating the thing the gene would need to be in order to spread better through the population, given its current makeup and environment.

In any case, to understand anything means to be able to explain it, if only to oneself, and that means replicating it mentally. So our universal explainer, which is the creative algorithm, has to be a universal replicator as well. Now by replicator in this sense, I don’t mean that it would replicate itself, though it could do that too, albeit maybe under as much difficulty as the creative algorithm in my mind is trying to copy itself right now. But by replicator, I mean that it would replicate other things. So where genes only try to recreate the thing they need to become in order to spread through the population, people, in their endeavor to explain, try to recreate the whole world.

So what does all this mean for the creative algorithm? Now that we roughly understand what knowledge is and how it’s created, the problem of implementing artificial creativity lies in the details of knowledge creation. For one thing, since knowledge creation is a result of an evolutionary process, it means that the creative program is a genetic program. By genetic program, I don’t mean anything biological. That’s just what this particular field of software engineering is called. Evolutionary program is another name for it. It runs within individual minds and is the origin of that individual’s knowledge, be it scientific, artistic, conversational, whatever.

As David says in chapter 15, “The Evolution of Culture”, quote, somewhere inside brains, blind variations and selections are adding up to creative thought at a higher level of emergence, end quote. We also understand now, I hope, that the creative algorithm is about solving problems and creating explanations. And lastly, we understand the following important realization of David’s, that it is the same mechanism that copies existing knowledge that also creates new knowledge. Therefore, I guess that if we build a program that is good at recreating knowledge, it would automatically be good at creating new knowledge.

In order to do this, we need to address the following problems. How do we encode knowledge or explanations in a program? Do concepts such as being hard to vary, reach, criticism, and testing, etc., also apply to computer programs? Here’s a hint, they do, and I will address that in another episode. How does the program find conflicts between different explanations? How do we create new knowledge in a program without specifying the program to the degree that it would already contain that knowledge? How do we mimic evolution in a genetic program? And lastly, if creative thought is blind variation and selection in a person’s mind, how do we explain the apparent goal-directedness of human thought, for example, the goal-directedness of understanding and conversations?

These are hard problems. But let’s look at the flip side. The harder they are, the more progress is waiting to be made in these areas. And some of these problems we will address in the next few episodes. I haven’t decided yet what exactly I want to cover in the next episode. There is a good chance we will get into how the epistemological concepts I just mentioned, so being hard to vary, reach, testing and criticism and so forth, how they also apply to computer programs, and how we may encode explanations in a computer program. Thanks for listening, and as always, please leave a comment or tweet at me. I hope to have you again next time.


What people are saying

What are your thoughts?

You are responding to comment #. Clear
Markdown supported. cmd + enter to submit. You have free speech here. You’re responsible for what you write. Terms, privacy policy
Your real name is preferred.
This small puzzle helps protect the blog against automated spam.

Sign your comment with GPG to create or add to a public profile with all your comments.

Paste a detached signature of your comment.

              
Paste your public-key block if you haven’t before. You consent to your key’s contents, including your name, being displayed to the public.

              

Preview

Playing ~