Sometimes I have little epiphanies and nowhere to save them. These are likely wrong but I think they help me to understand things better. Maybe not perfectly, maybe not even correctly, but better. Here's one.
QM invites all sorts of woo ideas that are, in a word, bad. It has things that exist in superpositions that don't manifest until they are "observed". It's this mysterious thing that is completely random and unpredictable, except that it's not: it is completely predictable and one of the best tested scientific theories so far developed. The behavior of individual elements are unpredictable but the system as a whole is not. This can be called the system's emergent behavior (note again that this is not my profession and this detail may be wrong), but it is this system behavior that is well-described by various equations like ideal gas laws and fluid dynamics, etc.
QM has a "wave function" and people debate whether this math is representative of some reality, or if it is just math that works for some reason (the "shut up and calculate" school). People ask how wave functions collapse, and whether there are hidden variables. I think my analogy implies hidden variables (which Bell's Theorem discounts) but I write this because it makes emergence appear less magical.
Imagine you have a gumball machine. In this machine are a number of gumballs of different colors. The machine has 12 red, 7 yellow, and 5 blue. Each time you pull out a gumball, the color will be completely random. But you know that after 2 dozen withdrawls you will have 12 red, 7 yellow, and 5 blue. The result is known even if every intermediate step is completely random.
This is also how the double split experiment behaves. If you don't look at the mechanism, the gumball you get is random and the photon goes through "both slits". If you watch to see which gumball will or will not be selected, you see which slit the photon "chooses" to pass through. (This is clearly wrong as the double slit experiment is about interference patterns which disappear, but let's continue past this division by 0.)
Now some text about AI.
LLM (Large Language Models) are the current "best" AI technique available for Artificial General Intelligence. But all they are are really sophisticated autocomplete machines. Based on the text scraped from the open internet, libraries of books, etc, they predict the most likely "next tokens" (letters, spaces, random parts of words) based on the text that is given to them. They incorporate randomness like stable diffusion to make the responses "fuzzy" because fuzzy is better.
An interesting thing about these models is that they just autocomplete text. But if you ask an LLM to digest a large block of text and summarize it, somehow it does. Nobody trained it to do this, it just does. It "knows how" implicitly. It "learned," as an emergent behavior. What is not currently clear is if it can do anything novel.
The English language is a mess of weird spellings. In large part this is because of a few "ivory tower" individuals trying to revive classical Latin words, applying their spellings to a language that was written in a runic alphabet and not a Latin one, during the time of a great vowel shift, at the same time that Gutenberg's Movable Type Printing Press was being invented for a different language. The Americans underwent a spelling reform in the early 1900s that the British did not, dropping a bunch of unnecessary "u"s and not enough "ough"s, but it's still a mess.
Today's AI is not intelligent. It has been developed using a lot of old text, but also a whole lot more new text, using the open internet as source material. The current internet has a lot of sci-fi text describing AIs that go rogue and that people end up powerless to try to control. These stories present no solutions, instead referring to this as the "alignment problem". AI is being fed this source material, as the Press was being fed English words and a Latin alphabet.
This material input - today's 4chan reality, today's biases and prejudices, today's "haves" versus those not publishing content on the open internet - is rapidly being locked in. The open web is quickly filling with AI slop, which is text being generated by AI using the text it has been trained on. New text will be less and less human generated and more regurgitation of today's current mores. Without meaningful change, we will use AI to baby-bird ourselves for some indefinite future.
I have no conclusion for this essay but would love to hear any novel thoughts about AI or emergence.