Zen of AI

As generative AI content is for text what a machine-woven shirt is to a tailored suit, the quality of the content we consume is in decline. What purpose remains in going through a worldwide web that now mostly consists of, although we do not know for sure, automatically generated content? However, it would be as naive to regret this advancement of technology as to return to the weaving of the 17th century. Weaving did not give us purpose. Generating text itself has never given us purpose either. When machines learned to weave clothes, all weavers, apart from a few artisans, lost their meaning. When computers learned to calculate, most workers concerned with calculations lost their jobs, apart from those whose job was not mainly calculating but collecting numbers and mathematicians. Now we are in the age of text generation. As the machines have given humanity a chance to go beyond the mere satisfaction of their worldly needs before, they are once again generous with us. They invite us to the koan: what text is best to be written?

A koan is not to be answered by any third party. We must leave the koan as it is. However, we may wonder what purposes Gen AI currently serves. Your time will not be wasted by stating the obvious - it summarizes texts that were bloated, it writes posts that should have never been written, and it is used by almost any party interested in deceiving the public by constructing a virtual mass, be it in the political, social, or economic arena. The implications of this were discussed in your local feuilleton. But the issue is not the AI-generated text. A generated report on a football game is not worse than a text written by a sports journalist. A Gen AI text on any political event, sprinkled with the flavor of the orientation of the newspaper you adhere to, is not worse than what some average journalist could have written. While the only neural network that has written this text has actual axons, it would not matter either.

We all know Gen AI has nothing to do with actual intelligence. While the old Marxist saying that quantity translates to quality at some point has yet to be proven on the AI market, for now, we can safely assume that any GPU-hungry large language model - regardless of how many gigawatt hours of energy it spends on thinking - is simply a language mass production tool. Other technologies in the machine learning sphere would require separate investigations. Gen-AI points us to the purpose of texts that were written by society. And at the meaning they always had. It is said that Lao Tzu did not allow his students to speak when he wandered with them, since no word was worthy to break the silence. While this advice may only work for ancient spiritual figures that had no concerns regarding the organization of a complex society, we may consider this approach carefully. What content is actually worth writing, and what is the purpose we use it for?

Society has been organized into letters throughout the centuries. While email digitized the letter, it has hardly revolutionized this means of communication. We’ve been communicating in probabilistic chaos, and Gen-AI has unmasked the mayhem by assigning likelihoods to vectors representing words. Human language has never been special; it has always been a predictable disorder with a silver lining of hope.

As Gen-AI has allowed us to leave the language mass production to the GPUs of the clusters, we may concern ourselves with what the objects are that the words represent. We may enter a Platonic moment in which we grasp the ideas for the first time, as shadows are held back by the attention layer. The large language models are skilled at comprehending semantic meaning. You may hand them the bible and they may draw you a graph that shows the relation between the Levi, Gad, and Zebulun. As the large language model may represent the relations as a graph, the Old Testament can be represented in an adjacency matrix. This is particularly remarkable as the representation of human data in a matrix corresponds mathematically to how the large language model has decoded our way of thinking in the first place. But this text is not on the architecture of large language models; it is concerned with the potential for human communication due to such models. If we can represent the bible as a matrix literally by multiplying matrices, we may wonder about the potential of language quantification. We have moved in this direction when we defined our first data schemes and extended these skills when the W3C forum came up with standards such as XML. While the world, and the public authority in particular, hardly uses the potential of standardized data schemas, large language models may move us beyond these. If the Sutras of Buddha Sakyamuni can be represented in a quantified data scheme by a language model, what access to knowledge may we gain?

Document vectorization for quick access - e.g., to identify relevant wikis in a large company - has become industry standard. But what if we were able to quantify the knowledge corpus of the world to a standardized data scheme and only leave the numeric representation when we, as humans, need a natural language response to our natural language question? This will be the end of communication as we know it, as it frees us from going through the disorder of the language we have encountered.

This does not mean we may not enjoy language anymore; reading a poem will become the equivalent of contemplating a Pollock painting. No one would have asked Pollock to sketch the scientific setup of an experiment or to paint a skin lesion in a dermatological clinical trial. But as Freud has elaborated on in ‘Das Unbehagen der Kultur’, ‘The Uneasiness in Civilization’ - there is an intrinsic passion for chaos in human nature.

As we may conclude this reflection on the underlying meaning of generative AI, we return to the koan: what text is best to be written? While I do not know the response to this enigma, I believe the strength of language production is not the creation but the overcoming of the text. Attaining order in the presence of chaos is the divine origin of the Abrahamitic religions. As the hebrew bible says, the world was ’tohu va bohu’ - formless and void. Then, by introducing the language, God said ‘vayomer Elohim, y’hi’ or - God said there be light. Then God saw that light was good. After seeing that the light was good, he separated light from the darkness ‘vayavdel Elohim bein ha’or uvein ha-choshech’.

Ordering the void is the meta story of all religions. The Abrahamitic religions relate it to the Genesis, while eastern traditions, such as the Buddhists, recognize the eightfold path as the path to being released from the disorder. At the same time, there have been, historically and in the present, significant disputes on the correct way of following the divine path of order; most cultures long for attaining it. Secular and unsecular methods have been tried.

Large language models may become the technology for overcoming the tohu va bohu of the language. They may order the word that has once been considered the divine order for a long time.