On LLMs building blocks

I've been recently trying LLMs (ie. ChatGPT) both for work and personal projects. You can read my full experience here.

A fundamental issue with LLMs is that their building blocks is not language, but rules and math.

What we percieve is the high-level assembled artifact.

LLMs won't be able to Tweet. Not because it's impossible and yes you can ask ChatGPT to write tweets for you. The issue is that it won't understand Twitter because Twitter operates at a high-level of abstraction beyond language.

Twitter's building blocks is the tweet itself, the RT and Quote RT as dynamics, but also wit, humor, sarcasm, challenge of compressing ideas, context, news, and pop-culture. Yes, these are built on top of language and words, but there's a stack that happens. Sarcasm, humor, and so are results of not even words, but language.

When we build something, we usually don't work at the level of the result, but at a lower level of abstraction (closer to the realms fo the concrete) to materialize.

The challenge of human fields is that we move up and down this ladder. An expert in a field is always "eye-balling" where the manipulation of the material will land once is assembled into the final artifact.