A fundamental aspect of human civilizations is culture, which I would describe as the preservation of complex thoughts and societal behavior in abstract or physical artifacts that can be passed on to future civilizations, e.g., literature, pottery, and music.

To understand the creation of culture, we could try to understand the very thing that generates it—the human brain. But the brain is so complex that “understanding” it is ill-posed. In fact, we have already created something that can speak and be understood like a human being, namely a large language model (LLM), and, yes, we don’t fully understand how it works. Some people are even dramatically questioning whether these models are sentient. Perhaps understanding sentience is as ill-posed as understanding the brain. What is there to understand? The brain makes its body walk and talk, so go ahead, ask the brain. Ask it about culture.

Wouldn’t it be interesting if we could query anything and get its response back? For example, we could ask a stone how long it’s been there, what it saw, or what its properties are. We tend to anthropomorphize things and give them attributes and functions that allow them to interact with the world (that is what object-oriented programming does too, in some sense). But there is something deeper about speaking. It’s free and open-ended.

Through this lens, we can look at culture as an abstract object that describes a group of people—a society. When we query a society, we query its members directly and receive a variety of responses that we do not have the capacity to process individually. Since we humans have limited memory and time, we preserve our collective thoughts and behavior, our answers to our own queries, in concise, idiosyncratic forms to pass on to the next generations. We aggregate these thoughts, compress them, or make them concise by summarizing them in our own styles. Thus, we reduce an inquiry into a society to a single summary—an artifact of that society. Culture, in that sense, is a society’s preserved self-description.

Human civilizations are constantly compressing thoughts in their endless journey of inquiry about life. Our judgments about life become centralized into a few artifacts, and a complete understanding of them would require a brain with a capacity as big as life itself. Imagine if life had its own brain. We could ask it anything and it would gladly respond with the truth, whatever that is and in whatever form it can. I think the brain is the ultimate compression machine; it compresses the world into a view of it and reacts to external stimuli based on its own memories and experiences of life. Ideally, everything would have its own brain, and instead of seeking or inferring the truth about something, we could simply ask it!

The ability to ask an abstract object is marvelous. Imagine if we could ask a society, heck, even a virus(!), about its beliefs. When would this make sense, and when would it not? Perhaps a virus is all there is about it: DNA with some shell—a protected piece of code. However, humans and societies are much more complicated when taken together with their constituents, and so a query about such an object would necessarily involve a “summarization” of the raw facts and memories behind the response. Cultural artifacts are such summaries of the deep inquiries humans make into various aspects of life. Nowadays, we might attach an LLM to a company and fine-tune it on all of its documents, giving the company its own brain. We could ask the company about itself, and it would happily reply with whatever mundane stuff HR writes about it.

I’m hinting at something deeper. LLMs are the best summarization mechanism we have so far; they model the tokens they consume very well. Everything we humans have preserved online has become the Internet, and LLMs can learn to model the content we preserved on the Internet as long as it can be tokenized (such as language, images, video, audio, and more). In short, LLMs summarize the culture we have digitally preserved online. For example, they can speak in different accents, draw art in many famous styles, and know a lot of history. However, and more importantly, we should note that an LLM does not represent a single constituent of a culture. It is more like a archive of culture than a creator of culture. Perhaps an LLM by itself may never create “culture,” at least not in the way we know it. But what if we create a society of LLMs, each with its own individuality and personality? Then, we may or may not achieve a new level of civilization that has never been seen before. The answer to that question will unfold sooner than we realize.

By the way, have you ever wondered where our personalities come from? Genes, environment, or a chaotic accumulation of differences stemming from a perturbation of both? Well, does it even matter to us where they come from? Because, for all we know, if a tiny tweak to a copy of an LLM plus a tiny tweak to its training algorithm can suffice to give it its own personality, then maybe that’s exactly what we are: tiny tweaks in a universe of predeterminate homogeneity.