look, no. The infinite monkey-typewriter thought experiment has nothing to do in any conceptual or practical way with Large Language Models. If you think that, you don't understand one or both.
LLMs are statistical models that improve on simple Markov chain text generators by not only mapping the statistical likelihood of words near-to-other-words, but extending it to in-the-presence-of-given-metadata. Given a novel arrangement of metadata — a "prompt" — the software does a walk through the statistical model to generate a sequence of words that, given a large enough training set, resemble meaningful text.
The infinite monkey-typewriter thing is about infinity, okay? It says: if you brute-force generate completely random sequences of characters, doing so at infinite scale guarantees that your set of output sequences includes a sequence containing the complete works of Shakespeare, because that's just a sequence of characters; a very low-probability sequence to simply happen to generate, but in conceptual Rock-Paper-Scissors, infinity beats unlikely.
That's the point.
(Because all of the following examples, and more, are representable as sequences of characters, generating an infinite number of character sequences is guaranteed to include all of them:
- complete plays of Shakespeare, arranged alphabetically by title
- complete plays of Shakespeare, arranged reverse alphabetically by title
- literally every other distinct order you can arrange the complete plays of Shakespeare in
- complete plays of Shakespeare with a single instance of the word "motherfucker" inserted in a gramatically correct position within a well-known speech
- this post
- the last message you, the reader, sent to anyone via a textual medium)
The problem with LLMs is capitalism. It's literally capitalism. If not for capitalism, they might have been invented, but you'd never have heard of them, because they would only ever have been put to appropriate — read near-zero — uses.