Mechanistic Iteration
Inspired by a conversation about how it’s impossible to resist the natural language interace of an LLM without anthropomorphizing them, I wanted to illustrate a “constrained” LLM generator that still has an internal monologue.
LLMs reason through language and requiring ALL their output to be constained means they would be less effective. This is ~an approach to solving the weird socio-technical by product of conversing with agents, while hopefully still retaining their utility.
Some ways this could go is dispatching “dynamic” grammars based on the context, and coming up with some interface of question asking that’s “not human” like so that there can be a tool-oriented dialogue.
This tool itself is generated w/ Claude mostly as a way to get the thought out of my head. It may not go anywhere after this and I haven’t really put it through the paces.