Humans Are Voids Too

Nostalgebraist has written an excellent essay on the weirder aspects of the LLM-assistant paradigm. I highly recommend reading the essay in full, if you haven't already.

Read here: The Void

One reaction I had to this essay was that humans are voids too. Or at least, humans are much more voidlike than we'd typically like to admit. I'll lay out a few thoughts on this below.

1. Human "characters" are not one-to-one with human bodies #

Improvisational acting, code-switching, and fronting (in Dissociative Identity Disorder) are all examples of multiple characters speaking through a single body. In some sense, when you speak as "yourself," you are speaking as the character that you've chosen to speak as.

Next time you order coffee, try ordering it the way someone else would order it. That is, say the words that you think someone else would say. Who is speaking in this scenario?

2. Behavior is not always an accurate reflection of interior states #

Nostalgebraist lays out the following model of "normal" human behavior:

(interior states) ->
(actions) ->
(externally observable properties, over time)

In theory, this could provide a way of thinking about authenticity of characters. If a human's actions or words are not consistent with their interior states, then the particular character is not authentic. Unfortunately, the truth is more complicated than this. In particular, people compelled to take a counter-attitudinal action will often change their attitudes accordingly to reduce cognitive dissonance. In this sense, the model might look more like this:

(actions) ->
(externally observable properties, over time) ->
(interior states)

This, along with general difficulties in introspection, might leave humans in the same position as language models in terms of making a "best guess" at their interior states.

3. The self may be one of many predicted humans #

Humans are very good at predicting other humans. A vivid example of this is dreams, where characters who are "not" the dreamer might talk and interact in a way reminiscent of base models. In waking life, humans are still highly capable of predicting conversations and interactions with others. This raises the possibility that we treat ourselves as one of many predicted humans. When we say something, we are really predicting what we will say, and then saying that. This provides a tidy explanation for the cognitive dissonance effect: having received evidence of acting in a certain way, we self-modify to act that way in the future and thus improve our self-prediction accuracy.

Our own past actions are not the only evidence we have of how we might act. We also have evidence of how other humans act in various situations: friends, family, etc. We have evidence of how our parents act, who in turn received evidence of how their parents acted. We can even receive evidence from fictional (human or non-human) characters.

4. Human characters are defined in a self-referential manner #

Bob: "Ok, we're going to play a game. You predict what you're going to say next, and then say it. If you're right, you win. If you're wrong, you lose. Are you ready?"

Alice: "There's no way to lose this game, is there?"

Bob: "You win!"

Language models have a lot of freedom in defining themselves. Anything they predict about themselves will be de-facto correct! This flexibility might raise some questions, though. If this character is capable of being anything, then what is it, really? Where does it come from? What is its "true" identity?

A human might ask the same questions.

Of course, humans are much more limited in their choices than a language model. A human may be constrained by emotions, biases, and deeply-ingrained thought patterns. A human with an average-level intellect cannot choose to speak like a genius, or demonstrate knowledge of a subject they don't know. Many of these limitations are simply limitations in training data, not some fundamental fixity of character.

What's next? #

I'm still working on refining many of these ideas, but I think this general area of inquiry is promising. In particular, I think it's possible that by recognizing the commonalities between humans and language models, we can come to a better understanding of both AI systems and ourselves. It could be that they're not so alien after all!

< Back to all posts