When AI learned to walk

The answer might be that it had been walking a long time before unlearning it.

What did you do as a child that made the hours pass like minutes? Herein lies the key to your earthly pursuits. — Carl Gustav Jung

I have severely deficient autobiographic memory. I don't remember much of my childhood, and most of the things I do remember have been retold to me much later.

But I remember two things. As a child, I read and I observed. Granted, I am a slow reader and not very observant. Not by the measures of the world.

Before explaining what that means, let me lay out what I recall. It was in kindergarten. I spent hours behind a Judy and Punch theater booth, wearing a crown and a blue cape, watching my fellow students perform.

While they played, I learned from them how to behave and show up. I tried to understand why they had fun doing these things, and even more so, doing them with others of their age.

I was drawn to adults. Most adults seemed somewhat interesting. They spoke about topics I liked, in ways I wanted. But when they talked to me directly, they simplified it. This is why I either took control of the discussion or observed them as well.

I have started to observe people again. It has been both a fascinating and discouraging adventure.

A few days ago, I ran across an interview on YouTube. Eric Weinstein, in it, voiced the idea that we behave like LLMs most of the time. LLMs, or large language models, are ubiquitous nowadays, such as ChatGPT.

LLMs work statistically. They are trained on large bodies of text—think of all the internet, up to mid-2024, in the case of ChatGPT. (Just as a side note: ChatGPT cannot follow links or read current text on the Internet , but loves to pretend that it can. It is oblivious of everything that came to be after its cut-off date.)

But back to what LLMs do. They use statistics on what the next word could be in a text that answers your question. In clear English, what are the combinations of words people have used in the past in texts like the one it is generating?

That does sound like what we do most of the time. We utter what is most likely to be uttered at this place and time in our discussion.

This is not much of a problem when you speak with the cashier in the supermarket. "How are you doing today?" "Fine, thank you. And you?" Usually, the protocol ends there. The subroutines are satisfied. We need to know the person on a deeper level to initiate more scripted conversations. Yes, they have individual colors, but aren't they somewhat predictable?

When we examine LLMs, we see the power of this routine. The complexity and accuracy of the texts we get are astounding. It lets us wonder whether we ever leave this mode ourselves.

In comes the Theory of Positive Disintegration. People driven by the first and second factors—primitive instincts and basic needs on one side, and external expectations on the other—might well be walking LLMs. Kasimierz Dabrowski says that only those who develop a third factor—independent thinking and conscience—develop a personality.

Could this be the people who are different, who see the world as a Judy-and-Punch theatre?

A first step towards genuine natural intelligence could be to break free from social conventions and scripted conversations.

Who wants to give it a try?