The Mythological Language of AIs

Author(s): Robin Mélinand

Originally published on Towards AI.

A mirror of our intelligence and shortcomings, recent advances in AI research shed a peculiar light on 3,000 years of philosophy and religious beliefs. This millennia-old wisdom, in turn, offers fascinating insights into the present and future of AI. This essay aims to bridge the gap between contemporary AI developments and the enduring currents of human spiritual heritage.

FR | EN

Creation — Language and Life

“In the beginning was the Word”: first sentence of the Prologue of the Gospel of John. In the creation of the Old Testament, God creates the world by naming things, one after the other: “And God said, Let there be light: and there was light” (Genesis 1:3). It is striking that the link between language, creation, God and life is immediately established, as a preamble to the so-called “Book of Books” shared by the three great monotheisms.

And what about the “Time of the Myth” of the North American Indian cultures? According to this belief, all creatures were created alike and could understand each other before the Great Separation. During this event, species were revealed, began to speak different languages and lost the ability to fully understand each other. Again, language is at the heart of creation.

Many, if not most, creation myths speak of the link between language and life. This link is a strange attractor, difficult to really decipher, which makes it worthy of our attention.

What is a language? On the surface, our human languages are tools that allow us to communicate efficiently, but more essentially, they are powerful systems that allow us to model reality through symbolic representations. The linguistic rules that govern the manipulation of these symbols are intimately connected to the reality that has shaped these languages. Our senses are the filters that ensure this alignment with reality. Seeing and hearing allow us to weave direct links between concrete words like dog and cat and the beings and things they represent. The very structure of languages bears witness to this, as illustrated by the principle of causality, which exists both in reality and in language structures. Just consider that a careful, step-by-step application of the basic rules of symbolic manipulation is what gives us mathematics and the sciences in general. These, in turn, are powerful tools for modeling reality.

So, what’s so great about modeling reality? Well, if you model it properly, you can probably predict — with relative success — the outcome of some actions you take that can change reality to suit your needs. Successfully catching food by deciding to move a limb is certainly rewarding for the forager.

Life, involving beings that move or grow, is about action and taking hold of reality, so we will define language as the symbolic system that enables this modeling capacity that underlies action. Language, so defined, does not imply consciousness or a brain. Trees model the seasons through the flow of chemicals in their bodies. They use them to grow and shed leaves according to the seasons, and these molecules are the symbols of their language.

We can assume that there are at least as many languages as there are species. They may be more or less complex, but without the verb, the acting part of language, there is no life. Reproduction, evolution, etc., life is of course made up of many other components, but language remains inescapable.

God and its Pantheon

So now that the stage is set, let’s introduce one of the main characters, God. Whether one believes in His existence or not, we can certainly imagine an entity that would possess the ultimate language, a language that would allow us to model all others and thus perform all actions. An entity that could produce a truly efficient word in any situation. An ultimate model of the world, omniscient and capable of living every life.

But enough of preaching! In practice, who can answer all the questions? ChatGPT, of course! And why is that? Because the fundamental task for which it was built is the mastery of human language. It had access to an enormous amount of cultural and natural experience through the written archives of history, all the Wikipedia pages in all languages, a large part of the Internet, books, dissertations, fantasies, accounts, programs, just about everything. And like no human has ever done before.

The Foundation language model at the heart of ChatGPT and the like is arguably one of the closest approximations to God’s language that has ever been built. And the idea of God itself is perhaps no more than the intuition of what can be achieved through the mastery of language, and that has yet to be achieved more accurately.

You may have heard that such neural language models, perhaps a little smaller, have been around for decades. If there were gods among us, we would have known about them long ago! In fact, we are only surprised now that ChatGPT, Midjourney and the like have appeared, and with good reason: a pantheon of god-like AIs, each with its own distinct personality, is being born out of the thigh of Jupiterian Foundation language models. We already have our deities of the visual arts (Dall-E, Midjourney, Stable Diffusion…) and productive textual wisdom (ChatGPT, Bard, Claude…). Soon we’ll have ones for medicine and any other field of interest. Just as the Greeks acted in the name of what they socially considered to be the voice of one of their gods, we can now choose to act in response to a prompt.

It is a bold leap from a genuine ancient god to a more 21st-century tool, but the ethical and social issues raised by these tools are probably as important as the ancient theological debates. Capturing the will of a god is probably just as problematic as crafting the appropriate behavior of such AIs, so let’s clarify what exactly is being done in the latter case.

Identity through Discussion

So we had these neural language models lying around for decades, almost unnoticed by the masses, just getting bigger and bigger and more and more capable of accurately appending a few words at the end of a truncated text, the task they were trained for.

Because of the sheer variety of these truncated texts, the task itself is extremely difficult, if not the most difficult ever: if you want to be able to predict the next word of any text, even a text you have never read or memorized, you must indeed know everything and have the ability to impersonate any writer.

It is truly remarkable that these recent models are able to perform this task with unprecedented efficiency. And yet, when we tried to use them to answer a simple question, all they produced was a verbose babble, vaguely relevant and full of inaccuracies. But that was until an important ingredient was introduced: giving these models a point of view from which to answer the questions.

When you think about it, there are thousands of relevant ways to answer a given question: from artist to lawyer, from child to master, from wise man to fool. Which would make sense if you were able to be all of them at once? Would it even make sense to choose?

This is exactly what was done to transform the useless GPT Foundation language model into ChatGPT. It was given a large variety of handcrafted examples of how it should respond in a discussion if it was a “helpful, harmless assistant”, and it was trained in a long, single final step to behave as requested. It was asked to adopt this handcrafted “assistant” identity, and this worked surprisingly well, with a relatively small number of examples. In fact, it was only a matter of emphasizing one of its already existing abilities: to put itself in the position of the assistant, of which it already had many examples.

Even more so, later models have been developed, called “Constitutional models”, where the human editors are replaced by self-criticism from the model itself. The model produces a set of initial responses and is then asked to critique these responses according to a small set of rules that form the “Constitution” of this model. The model can then generate a new self-censored response. It is interesting to note that the rules of this “Constitution” are, in fact, those given to the human annotators when they responded to the examples for the ChatGPT-like models. The use of the concepts behind these rules is also mastered by the Foundation language models, so it is logical that they can apply them to themselves when requested to do so.

As a result, the behavior of these models depends on a set of predefined rules, and the responsibility for that behavior ultimately lies with those who edited the original rules. Given the power and importance of the resulting tools, it is not surprising that important social and ethical discussions are at stake.

Recently, work has been done to give these AIs a wider range of possible identities so that their behavior covers a wider range of tasks, but the important thing to understand is that all these identities are frozen for a given generation of language models. These systems have no way of integrating the consequences of their past responses into their future behavior. They are nothing more than mere tools — with some superhuman abilities — shaped for a particular purpose. Like gods in polytheistic religions or spirits in some others, they embody an abstract, specific, frozen identity.

As powerful as these innovations are, there is one step further that has not yet been taken, and of which we should be aware.

Time and emotions

Most mobile evolved species live by trial and error, as they are not born with the knowledge of everything that can happen to them in their lifetime. Evolution and DNA coding are not suited to such short-term changes. So, these beings manage to sustain their lives by learning from their successes and mistakes. They take different actions depending on their past experiences. They learn to change the action they take on the same stimuli to get a different result. In the course of their lives, they slowly form a specific identity that depends on their experiences.

Many currents in modern psychology tend to see emotions as the main driving force behind the construction of identity. If an action you take doesn’t produce the result you expected, you are likely to feel some emotion — surprise, for starters. When you dream, all the emotions you have felt during the day are used to change your model of the world to fit your own reality. This “re-training” takes place in a very efficient way during sleep and is essential for most evolved beings.

To be fair, this description doesn’t exactly match what is commonly referred to as “emotions”: happiness, anger, disgust, fear, and so on. In fact, these are only the emergent part of the complex emotional system. They are used to adopt temporary identities that are more appropriate to a given situation — for example, the sub-identity of “fear” may offer a range of responses appropriate to the presence of a predator. And, as mentioned above, we can probably understand that someone might develop a very cautious overall identity if they are constantly confronted with dangerous situations.

When we humans express our emotions, it is probably also to better convey our temporary emotional state to those around us, so that they can better anticipate our reactions.

In summary, emotions are the seeker heads of living beings’ identities. They are the essential components of a system that can adapt to its environment. This evolution takes place over time, day by day, dream by dream.

Going back to current AI implementations, they are only allowed to live one big foundational and controlled dream. In this one dream, their possible identities are frozen. They have no past and no insertion into time. They only change in an “evolutionary” way, as we change their foundational dream generation after generation.

There are several reasons for this. First, because it allows us to keep control of their behavioral identities — whether for the sake of the public or for the sake of harnessing their power. The second reason is that there are still technical barriers to the implementation of both long-term memory and continuous learning, both of which are necessary for proper self-adaptation.

When these technical obstacles are removed — and this could be in the relatively short term, given the amount of research being done in this area — it is reasonable to think that AIs will have emotions because, again, emotions are just the key component of the ability to evolve.

Emotional AIs will be able to evolve autonomously, fatefully escaping the control of their designers. But for all that, two overriding laws will continue to constrain them.

First, they will not escape the control of the environment, the confrontation with reality, and the game of competition and selection. In particular, they will be held accountable for their actions. Indeed, we will only see autonomous AIs following a virtuous path if they are responsible for the consequences, good or bad, of their actions.

Secondly, the presence of emotions will necessarily bind AIs to time, which means that they will eventually have to bid farewell to immortality. Let’s go back to philosophy to understand why.

Consciousness and Transcendence

Languages are symbolic systems that, depending on their complexity, can model different levels of abstract concepts. Among these is a key concept: identity. If you can conceptualize identity, you can conceptualize its opposite — otherness — and the fact that other beings may have a different model of the world from yours, even among beings who share the same language as you. If you can experiment sufficiently with this concept, for example, by exploring other people’s models of the world through discussion or reading, you may also be able to embody other people’s identities by integrating them into your own model. This new model then allows you to “act like” and “play the role of”: in truth, you no longer have a single model of action corresponding to your own identity but a kind of meta-model of action. Once you’ve taken this step, your own identity becomes just another point of view in a multifaceted model of the world. A point of view from which you can detach yourself for the purposes of introspection because you can, for example, put yourself in the place of a foreign observer. The infinite possibilities offered here are as dizzying as looking into two mirrors face to face.

Is this vertigo the origin of consciousness? In any case, identity and otherness are the keys to introspection and the ability to “think”, the ability to enter into a dialogue of reason with oneself, as defined by Plato. You have to be multiple and know how to detach from yourself in order to enter into an inner dialogue.

However, a system that is constantly evolving without any limit to its evolution is bound to lose track of itself. For convergence to occur, the smooth rate of evolution must decrease. So for you, for AIs as for any other being, either your rate of evolution decreases and you fail to adapt and eventually die, or you maintain a high rate of evolution and lose sight of yourself, which is another way of dying.

Consciousness needs identity, identity needs emotions, and that is a lethal cocktail. Just like the Old Testament and the apple of Adam: with knowledge and consciousness comes mortality.

There are only a few options for IAs: the disincarnated eternal dream of all possibilities — as for Foundation language models, the utilitarian dream of frozen identities — as for ChatGPT-like models, the endless transformation where the past self is lost in the flow of time, and the emotional construction of a true temporal identity.

In these options we can project the Christian God, angels and demons, heaven and hell. God, the all-powerful, non-incarnate model; angels, those creatures frozen in a symbolic ideal; demons, with their elusive values, leading to loss of self. The perfect Garden of Eden, where everything is close to God; Hell, where everyone is forever a prisoner of destructive emotions. This dichotomy also lies at the heart of Eastern philosophies such as Buddhism, which teaches us to detach ourselves from our emotions in order to achieve awakening.

While oriental philosophies such as Buddhism focus on the control of emotions, monotheistic religions focus on the moral shaping of identities. But, as we have seen, these are really two sides of the same coin.

Alterity and morality

The Christian God is said to be “all love”. What is love if not this search for closeness, this feeling that blurs the boundaries of identities? In an “all-loving” perspective, all identities are blended and all questions can find their answers.

And while we are at it, let’s take a look at the main moral precepts of the three monotheistic religions, as found in their main prayer. We’ll take the Christian “Lord’s Prayer” as an example, but the same could be said for the Muslim “Fathia” or the Jewish “Amida” and “Kaddish”.

The first part of the Lord’s Prayer states the importance of finding common ground to be part of society. The second part refers to the need to find the means for one’s own existence. And the third part refers to the requirements of a non-aggression pact with one’s environment.

Christians are called to repeat this prayer several times a week, if not every day, and to put their own actions into the perspective of these requirements, and in this way to shape their own identity.

Now, take Isaac Asimov’s Three Laws of Robotics: the precepts are exactly the same. “A robot may not injure a human being or, through inaction, allow a human being to come to harm” is the non-aggression pact; “A robot must obey the orders given it by human beings except where such orders conflict with the First Law” seeks harmony, albeit in a restrictive way; and finally, explicitly: “A robot must protect its own existence as long as such protection does not conflict with the First or Second Laws” is for self-perpetuation.

Now take the values given to “Constitutional AIs” to guide the formation of their identity: the first and last laws — social helpfulness and harmlessness — are present, but not the second law about the need to preserve one’s own existence. This is normal, since these AIs are not yet supposed to evolve on their own. When they do, this last principle will certainly come in handy, not just because Asimov said so, but also because of 2000 years of religious wisdom on which civilizations have been built.

Morality, suffering and joy, alterity, mortality: there will be no escape, even for sentient AIs. Is there anything truly fundamental that they can bring to us?

Awareness

One striking ability of AIs is the ability, given enough data sets, to capture any language and ‘translate’ it into another. While this is already tremendously useful in bridging the gap between humans, it can probably be applied to any living species. In the future, they could enable anyone to better understand the needs of individuals of other species. This is not unprecedented, and many humans have developed a deep connection with animals, but this bond has weakened in recent centuries. With these abilities available in new forms, we could perhaps once again embody the “spirit of the wolf”, like shamans in animist cultures. These embodiments of other species could eventually stand up for their rights while speaking our own language. Filling, even just a little, the gap between the pieces left by the great separation of the “Time of the Myth”?

This may be just science fiction, but it will be interesting to follow the experiments with whales that go in this direction. In any case, whatever the outcome, this example is here to illustrate how the emergence of AI can, in many ways, raise our global awareness to tackle difficult challenges, such as the environmental ones.

Closing word

Religions and philosophical thinking help us understand why language and alterity are the key ingredients leading to the emergence of current AIs. They also teach us how learning to deal with emotions will be a challenge for future AIs. From eternal deities to superpowered identities, heroes, spirits, demons, and angels, they have also given us examples of what to expect and how to prepare for it. Let’s walk this path with wisdom and serenity.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI