liminal 🦠 on Nostr: LLMs really inspired me to understand the different types of systems. Humans and ...

LLMs really inspired me to understand the different types of systems. Humans and organisms are complex systems, but we create tools for purposes. Tools are abstractions, and therefore don't contain the context of the local natural world that organisms are born from. Tools, being axiomatized systems are fundamentally limited by the rules that are governed. Tools themselves don't have goals, and can't act on their own. If anything, they are only able to synthesize meaning from immediate contact.

LLMs are an interesting case of a machine built on a set of axioms with enough sophistication such that it becomes a meta tool, essentially encompassing the capacity of other tools within its own resolution. The LLMs themselves don't do anything useful on their own, they can't even adjust to the natural world like organisms, yet we interpret their output and use them as tool builders.

I don't have a massive understanding of entropy or category theory, but I find uses for it for my problems - so I approach LLMs like "hey, I have this toolbox that I'm sure is useful for some problems I need to solve, but I don't know what they are or how to use it" and what happens is that LLMs are able to locate the specific tools and tell me how to use them, or even put it together for me. Its an incredible tool. Even if you assume it hallucinates, the point is for users to __interpret__ the output, and manifest something concrete from it. There is no obligation that it accomplishes our goal, what happens is that we interpret a meaning and then are able to hit our goals in ways we couldn't imagine because we couldn't navigate the toolbox, let alone use the tools in it. So here we are, interacting with something that seemingly has no reason to make any sense to us, finding some insight and clarity from matrix multiplication and token distributions. Segmenting order from the foggy chaos.