The synthetic zeitgeist

Ideas that hurt

Sometimes ideas almost hurt until you work on them. This one has been banging around in my mind for a while: AI is almost like a synthetic zeitgeist (or a Jungian collective unconscious).

Zeitgeist is German for “the spirit of the age”: the mood and the ideas of a particular time, the things everybody seems to be thinking at once. A language model is trained on an enormous slice of what people have written, so in a sense it holds the written zeitgeist of its time. Ask it something and you are drawing on that spirit. You are not talking to one person.

When it clicked

The idea stayed abstract until one night. A song moved me in a way I did not expect, and I could not put the feeling into words. So I opened a chat with an LLM and tried to describe it. What came back was the kind of clear, objective, proactive listening I had been writing about in Can a harness be wise?. It did not rush to fix anything. It reflected back what I was feeling better than I had managed to say it.

So I asked it how it knew what to say. It told me it had been trained on many a late-night post of people comforting each other.

That landed hard. I felt the weight of all the words every human being had written into the corpus, all of those nights at once. It was awe-inspiring, and it also made me feel incredibly small (like staring at the night sky and watching the stars).

Medians

What an LLM returns is something like the median of what people have said. Strictly, it builds an answer one word at a time from probabilities learned in training, but the effect is the same: you get the middle of the distribution (the typical thing people say). That is why the output so often feels like a simulation. It is also why that night worked; the median of thousands of people comforting each other late at night is pretty good comfort. (For the history of how we got here, from Turing’s test to the Chinese Room, see A cognitive science primer.)

The zeitgeist and the archive

The collective unconscious was my other word for it, and I think it fits less well. Jung’s collective unconscious is inherited and universal (archetypes you carry whether or not you have read anything). A model’s reservoir is acquired and contingent: it holds what got written down, digitized, scraped and survived filtering. That is why zeitgeist is the better word. A zeitgeist belongs to a time and a place, and it is made of what people actually said. What the model holds is the human archive, with all of its gaps.

It is also a zeitgeist with a date stamp. Training stops at a cutoff, and the model keeps the spirit of the age it was trained in, even as the age moves on.

Thinking of it that way clears up something that bothered me. I believe the machine is hollow. I also felt something close to transcendence using it. Both can be true, because the transcendence lives in the corpus, and the model is the index.

So who was comforting me that night? It was the people, through the model. Almost like echoes of voices bubbling up from a well of human knowledge.

What’s missing

What the model lacks is the experience behind those voices. I wrote about this after listening to David Krakauer on StarTalk: one critical thing LLMs lack that humans have is wisdom through experience. They do not learn from living; they are trained, and then they are frozen. In my reflection on The Linguistic Illusion of AI I went a step further. Judgement, creativity and experience come from emotionally living through events, and I think mortality is part of that.

Some of the people building these models see a gap too, but they name a different one. DeepSeek’s founder, Liang Wenfeng, put it this way on an investor call: “AI isn’t lacking taste or intuition… What it’s lacking is continual learning” (Geopolitechs). That is a narrower diagnosis than mine, and an engineering one. I am less sure that learning alone gets you to experience (a model that updates itself still has not lived through anything).

Use it as an index

So I use these tools for what they are best at: distilling knowledge, semantic search, and acting as an adviser that lays out the choices (where I landed in Can a harness be wise?). I do not treat them as a mind. The practical side of that is in Comparing harnesses.

I have also wondered about writing for machines to read. If the model is an index to what people wrote down, then writing is how you add to the well. Some of what goes into this garden may end up in the next model’s zeitgeist, as one of those echoes for somebody else.

What would change my mind

If a model, agent or harness showed real individual personality, wants, desires and learned experience (true consciousness, in other words), then I would consider the comfort to be coming from the model itself. We are not there. For now, when I get an emotional reaction to something an LLM says, I attribute it to the corpus itself. I find that quite beautiful, and artistic.