Two fictional analogies for large language models

When using a large language model to gain knowledge or perform tasks, they remind me of the Library of Babel: they're capable of outputting basically all grammatical assemblages of tokens, and their probability distribution contains (a fuzzily associative, highly compressed, copy of) essentially all of human knowledge and thought. Thus contained within it is the complete catalogue of useful, insightful, correct, and wise things a human being might say, and all the wrong, dumb, or plain nonsensical things a human might say, just as the Library contains all of that and also all of the nonsensical assemblages of letters around it. Also like the Library, even if you find something within the total distribution that appears coherent, you don't know if it's correct — after all, the library contains "the faithful catalogues of the Library, thousands and thousands of false catalogues, the demonstration of the fallacy of those catalogues, the demonstration of the fallacy of the true catalogue" — so you can only gain something useful from the library if either, one, you already know the answer, two, don't need the answer to be correct, or three, have a way to verify the answer to the degree you need it to be correct.

Thus the boundless hope and optimism with which some view large language models, because useful combinations of tokens can be found within that distribution, and likewise the endless depression others find therein, because for every useful output there is a useless one as well, or ten, or a hundred:

When it was proclaimed that the Library contained all books, the first impression was one of extravagant happiness. All men felt themselves to be the masters of an intact and secret treasure. There was no personal or world problem whose eloquent solution did not exist in some hexagon […] As was natural, this inordinate hope was followed by an excessive depression. The certitude that some shelf in some hexagon held precious books and that these precious books were inaccessible, seemed almost intolerable.

The differences between LLMs and the Library are:

1. that not every output is equally probable: only generally grammatically correct and seemingly-comprensible ones, usually;
2. you can direct an LLM to different parts of the probability distribution, guiding them and increasing your chances that the Library will produce the knowledge you need;
3. you can control how likely or unlikely the outputs they generate are, to expand and contract the space within which you must search.

Thus, as long as the conditions for the usefulness of a Library hold, this one is much more useful than Borges'. Nevertheless, many of the same psychological and almost theological phenomena result: some people view this fact, that we've compressed essentially all of knowledge and human thoughts patterns into 500 gigabytes that you can direct to a space of probability within that wider probability curve with awe and curiosity, and see the possibilities; others are filled only with despair that much of what this thing will produce is bullshit.

When using LLMs as conversational mirrors, however, LLMs remind me of something stranger. In Nova, Samuel R. Delany argues that the Tarot is in fact a useful tool for cognition:

Mouse, the cards don’t actually predict anything. They simply propagate an educated commentary on present situations…The seventy-eight cards of the Tarot present symbols and mythological images that have recurred and reverberated through forty-five centuries of human history. Someone who understands these symbols can construct a dialogue about a given situation. There’s nothing superstitious about it. The Book of Changes, even Chaldean Astrology only become superstitious when they are abused, employed to direct rather than to guide and suggest.

As this article by someone who actually practices such a thing expands:

A blind faith in the tarot as a means of predicting the future, like any form of divination, is characteristic of childhood, and most of us mature to the point where we no longer take it seriously. For artists like Delany, however, there’s another stage beyond this, in which we realize that structured repositories of arbitrary symbols—often preserved in the cultural memory as oracles—can be a tool for thinking through questions that are resistant to more rational analysis. Writing about the tarot a few years ago, I called it "a portable machine for generating patterns," and I noted that it results in a temporary structure, spread out across the table, that feels like an externalization of the problem in your head. You can examine it, scrutinize it from different angles, or even walk away from it. […] It won’t tell you the future, but by forcing you to map or analogize your current situation onto a matrix of charged symbols, it can provide surprising insights into the present.

I propose that one way to view large language models, is as something like this: talking to an ungrounded large language model (e.g. having it play devil's advocate and ask you questions and try to disprove you and stuff like my dad does) functions as sort of like doing a tarot reading. It cannot predict the future, or even accurately describe the present. There is no truth or mind or meaning there inherently — it's all assigned by your mind. But through the act of trying to map whatever discourse or concern you have onto this semi-randomly generated pattern, you may get a new perspective or be jogged out of limited ways of thinking. And through the act of trying to analogize the entities operating in whatever discourse you're concerned with, to the layered meaning of the symbols on the table that it is manipulating uncomprehendingly, you may layer more interesting meaning or uncover more of your own personal unconscious thoughts. As long as you remember that you can always dismiss what the cards, or the model, have so positively said as meaningless or useless and walk away from the table, you may still glean something useful from it.

I've run a few experiments essentially using LLMs this way and it seemed profitable — taught me a few things about myself I hadn't phrased explicitly before, through asking me randomly generated questions. I think when it comes to the Tarot itself, his argument is somewhat flawed, at least as it applies to me, but with LLMs it is perhaps less so, and works better for my more literal, less mystical nature, since the patterns it creates are not purely random, but based on a probability space trained on all human reasoning and thought patterns, and since the symbols it works with are all of human language, not just a few set archetypes. It mimics the patterns it has seen of criticism, critique, questioning, and sometimes it's able to ask insightful questions that challenge me to construct new answers, whereas other times it may ask nonsense or say things that are not well reasoned, which I can ignore. And unlike another human interlocutor, I can dump entire pages from this blog, collections of writings on various topics that can sometimes reach into the tens of thousands of words, into it, and it can combine that with reading any of the texts I link to, and all of its compressed world-knowledge, to find the right patterns to question me about, thus giving it infinitely more specificity and context than any human interlocutor I'm likely to find, as well as infinitely more patience.

Is it a replacement for discussion with a human being, who can more trenchantly challenge my views? No. It tends to accept, or at least not challenge, the underlying moral values and assumptions I make, which is nice when I don't want to stand around arguing with a nazi, a statist, a Luddite, or a primitivist, but sometimes it's worth arguing with those — well, not the first one — to sharpen my wits. But you can do both/and.