In his 1950 paper “Computing Machinery and Intelligence,” mathematician Alan Turing considered the following question:

Can machines think?

Such a question is a significant philosophical undertaking. Terms like “think” and “machine” are intuitively understood, but difficult to define precisely. Turing proposed a solution to this implicit complexity in the form of a simple game.

The so-called “Imitation Game,” Turing explains, involves three participants: player A, player B, and a third participant, the interrogator. Through anonymized and simultaneous correspondence with both parties, the interrogator attempts to correctly identify each participant. The catch is that while player B tries to help the interrogator identify them correctly, player A attempts to subvert this effort by imitating player B.

For example, in a version of the game involving a man (player A) and a woman (player B), the interrogator asks both participants questions. Player B, the woman, might reply with statements like “I am the woman, don't listen to him!” However, such remarks could be imitated by player A, the man. While this game could be a fun at parties, the primary goal is to determine the distinguishability between two entities through interrogative questioning. Later in the paper, Turing speculates on what might happen if player A were replaced with a digital computer. What would it mean for a machine to successfully imitate a human?

Through such an approach, the question of what it means to think is avoided, and a machine simply means a digital computer. Later referred to colloquially as the “Turing Test”, this Imitation Game gained traction in AI research, popular culture, and philosophical discussions as a measure of machine intelligence. The bar was set. If a digital computer was indistinguishable from a human in communication with a third party, then it was said to be intelligent.

Then we sort of just zoomed past that. While experts don't yet agree on whether the Turing Test has been properly passed, modern language models are more than capable of imitating human language to a very high degree of accuracy. The proof of the computational pudding, is in the eating. We consume content generated by artificial intelligence without noticing. Whether it be written articles, photorealistic images, or deepfake video, distinguishing between human and machine is and will only become more difficult as our models improve.

So is that it? Did we do it? Did we make machines think? Well, honestly, it's hard to tell. Expert opinions are mixed, discourse frequently involves loosely defined terms which hold little meaning in deeper discussion. The Turing Test itself is only loosely relied upon, and ultimately does not address the question of what it means to think. That question, is what I'd like to focus on today. What makes a mind? In this multipart series of indeterminate length, we will be re-examining the question Alan Turing asked 74 years ago.

Can machines think?

The Problem with the Turing Test

It's worthwhile to note that when Alan Turing published the aforementioned paper, “Artificial Intelligence” did not properly exist. After Turing's tragic suicide in 1954, two years passed before John McCarthy coined the term “Artificial Intelligence” at a Dartmouth College conference. Alan Turing, however, did lay significant groundwork for the fledgling field of research. But while his work was foundational, there are significant issues with the “Imitation Game” as a measure of machine intelligence.

Anthropocentrism

While “intelligence” is a highly equivocal term (more on that later), it would be a stretch to claim that it's a human-only trait. Complex intelligent behaviors have been identified in corvids, octopuses, other mammalian species such as elephants and porpoises, and many other creatures. Humans certainly seem to be the “most” intelligent, but this does not make us the barrier to entry. Yet, the Imitation Game proposes a method of identifying intelligence solely through comparison with a human.

In defense of Turing's test, the argument was never that in order to be intelligent, one would have to successfully imitate a human. Rather, the argument was that if something could imitate a human, it was likely intelligent. However, this still does introduce a practical issue, in that any form of intelligence which was not capable of conforming to human behavior, would not be identified by such a test. If intelligence were to emerge within a machine, but did not look like that of a human, then it would not be identified.

One potential modification could account for this issue, at least to some degree. While there may be infinite variations of intelligence, simply swapping out the participants with various known forms of intelligence could be a powerful update. Perhaps a crow would be player B, and a machine would attempt to imitate the crow's behavior. Although, we may find this to be replacing one practical issue with another, as I've never had a conversation with a crow. Nonetheless, a true test of intelligence must account for the diversity of intelligence, not just humans.

Sleight of Hand

Machine Learning is a branch of “Artificial Intelligence”, focused on using various techniques to coerce machines to, well… learn. Among the most popular machine learning techniques is the neural network. A neural network is essentially a program trained to solve a problem, based on a set of examples, where that problem was solved. In a more technical sense, a neural network is an approximation of a function. Assuming a function exists to solve a given set of problems, the goal of a neural network is to solve said problem by attempting to find that function.

A neural network is kind of like a jigsaw puzzle, where you don’t know the final image. You can find matching pieces, and eventually construct an outline for your puzzle. Over time, fitting each individual piece to the larger structure, an image begins to appear. Once complete, the jigsaw puzzle forms a complete picture, composed of smaller pieces, carefully fit together. That image is like the function found by the training process of a neural network.

(This awesome video by Emergent Garden explains how Neural Networks can learn almost anything.)

The best part of neural networks is that they can be trained to learn almost anything, towards virtually any goal. They can even be trained to lie. This capacity, however, produces a critical issue for a game focused on the ability to imitate. If the goal of training for a neural network becomes successful imitation of a human, then it may very well pass the Turing test with flying colors, while only having successfully copied human actions. But does this successful imitation not indicate intelligence?

Imitation is no small feat. The nuance involved in a human conversation is of the upmost complexity in the universe. We have not recreated this. Imitation of a conversation, preserving such nuances, is a tremendously complex task. Perhaps in order to truly imitate, the underlying cognitive processes of humans must be successfully reproduced as well. We certainly would not consider a mirror to be intelligent, even though it almost perfectly imitates the image of a human.

So the question becomes, is it the same to think as to imitate. Some say no.

The Chinese Room

Imagine a person who does not understand Chinese sitting in a room. A set of rules for manipulating Chinese characters, written in English, are given to said person. People outside the room pass in Chinese sentences, and the person inside the room uses the rules to produce appropriate responses in Chinese. To the people outside, it appears as though the person inside the room understands Chinese, but they do not. They are simply following a set of syntactic rules to manipulate symbols.

The famous “Chinese Room” thought experiment was proposed by philosopher John Searle in 1980 to address the question of artificial intelligence. Searle essentially argued that simple imitation does not equate to true understanding. In other words, it matters what’s going on under the hood. This, in contrast to Turing’s proposition, which states that the underlying processes of the mind don’t matter, as long as the end product is convincing.

On some level, we can consider that both are correct. A truly convincing digital machine might require having built underlying processes to successfully produce an end result which feels human. However, this begs unanswered questions. What does it mean for a machine to be truly convincing? Winning the Imitation Game likely isn’t enough. Learning to imitate is quite literally a surface-level approach, so we must go deeper.

Arguing over Semantics

Deeper conversation requires clearing out some roadblocks. Discourse surrounding intelligence is full of terms that create confusion. What is the “intelligence” in artificial intelligence? What does it actually mean to think? If thinking is a product of a mind, what is that mind? In other words, let’s clarify the underlying question.

Mind

What is a mind? Mind is an equivocal term. To a neurologist, a mind is a gelatinous blob of chemical processes between the ears. To a psychologist, a mind is a more abstract concept, which envelopes the human experience. A computer scientist might see a mind as a complex program that could be replicated with the right combination of machine instructions. For a cultural anthropologist, a mind is a structure which forms as a node in a web of social interactions. Need I go on?

(In an excellent lecture on Neoplatonism, John Vervaeke ideated well the equivocality of the mind, and how to think about it more broadly.)

However, the equivocality of the term could be used to our advantage. While various perspectives seek to define mind within various contexts, oftentimes the true meaning of a thing can be found in the intersections between interpretations. Mind, therefore, exists as something which extends into various fields of study.

It has overlaps with the brain, and therefore some material structure exists which encapsulates processes of mind. However, it’s not strictly material, as mind is discussed in the abstract space by psychologists who find patterns in human behavior and cognition. This combination of abstract patterns and material grounding, allows for computer scientists to replicate certain aspects of human cognition, through similarly material means. However, mind is also something which evolves over time, and collectively develops patterns to meet a goal.

I would like to do a significantly more detailed exploration of the idea of the mind in later parts, and therefore we won’t contemplate too much further today. However, what we do know about mind is that it seems to be a multifaceted structure, capable of pursuing goals through a complex series of actions. A human mind, for example, navigates the environment of the world as well as human society, and is simultaneously shaped by it, and acts upon it. The mind orchestrates the various actions of the human agent.

Intelligence

Intelligence is almost as elusive as mind. It’s easy to say that something is intelligent, but hard to know what we mean. Adding to the confusion, intelligence is a word used in various respects, with consequentially various definitions. I think that we can start by getting a feel for intelligence, then breaking down core attributes into something more concrete.

An intelligent person, is usually considered intelligent because they are good at doing something. A genius physicist is capable of understanding mathematics extremely intuitively, and using the symbolic system to find solutions to complex problems in abstract space. A genius entrepreneur, is capable of building a system of value and proposing it to a market, then organizing and orchestrating a group of people to continue expanding that value proposition.

We think of chimpanzees as intelligent, because they are capable of many of the same tasks that we are. Tool use is something found in various animals, including chimps, but also crows, octopuses, otters, and certain crocodilian species. Some modern theories suggest that intelligence is best understood as a force towards a local decrease in entropy. In other words, intelligence is a property of the universe, which brings ordered complexity from a chaotic system.

In such a case, the tool use exhibited by animals is intelligent, as an attempt to sustain their bodies and in turn increase local complexity through survival and reproduction of their species. Intelligent humans are those who help us tame the environment in ways, bringing order out of chaos. The unknown becomes known, and then useful to us. The environment becomes something that we bend to our will.

However, the innateness of this force of intelligence is the question. Is a machine innately intelligent, or is it simply an extension of the intelligence that humans exhibit?

Information and Relevance Realization

We use numbers every day. Simple mathematical symbols which aid in everything from day-to-day life to the most specialized corners of quantum physics. But what does a number represent? In Information Theory, information is measured as the reduction of possibility in a set of outcomes. Therefore, a number is a discrete unit of information. If you have a 1, it means that you do not have a 2.

However, a 1 doesn’t only mean that you don’t have a 2. It also means that you don’t have a 3, or a 4, or any other number of infinite possibilities. Where do these numbers come from? When we see a bundle of apples, we can count how many apples are present. Perhaps we have 5 apples. Why do we have 5 apples? You may say that it’s simple, there are 5 apples in the bunch. But what is an apple?

An apple is a particular composition of atoms produced as a fruit from a certain type of tree, right? Well, no two apples have the exact same composition of atoms. Additionally, if you take a bite out of the apple, do you not still have an apple? Maybe an apple is an apple because it has an outer layer of skin, a sweet interior body, and seeds in the middle, with a stem. However, by such a definition, almost every fruit would be considered an apple.

In fact, clearly defining any object becomes impossible when you attempt to account for all possible variations. Yet, we can count 5 apples. We look at a bundle of apples, and we see 5. Not 4, not 6, but a discrete number of apples. This reduction of possibility is the nature of information. Now we can go tell all of our friends how many apples we have. This is important, if we have 5 apples, that’s enough for 5 people to have a snack.

We humans are capable of reducing the infinite set of possibilities around us. We can count the number of objects in a group. We can identify those objects in the first place. We can realize relevance. Relevance Realization is a term popularized by the Cognitive Scientist and Philosopher John Vervaeke, who I would recommend highly. Humans are generally intelligent due to our innate ability to pick out what’s important in a giving context, and filter out the rest.

Machines can’t do that yet.

Just getting started

As mentioned before, this is going to be a multipart series. The topic in question, the creation of artificial intelligence, is about as complicated as it gets. There are more terms which need definition, more concepts to introduce. We will, over the coming weeks, unpack what it means for us to think, and ponder whether machines can do it as well. For now, let’s marinate on what we discussed today.

What is a mind? What makes intelligence? How do we realize what is relevant? These questions are the foundation of our later discussion. Stay tuned for part two next week!

Author’s Note

Thank you so much for reading! It’s amazing to me that people actually want to read my work, but I’m endlessly grateful that they do. A conversation like this is a collaborative effort, so let me know your thoughts in the replies below, in chat, or over in Substack Notes. Thanks again for reading, and I hope to see you next week for another installment in this series. Goodbye.

Resources

Throughout this essay, I’ve referenced various videos and additional resources which I believe are highly informative to this conversation. For your ease of access, I’ve compiled a list of them here below:

Credits

Thumbnail

Bilal Azhar - https://substack.com/@intelligenceimaginarium

Music

Track: Marshmallow by Lukrembo, Source: https://freetouse.com/music, Copyright Free Background Music