Mark Eschbach - Interpreting Inference

Pure programs make their effect on the world explicit. In logic code, the effect is a term. In functional code, it is a monad, a value the system knows how to reason about. An LLM agent has neither.

Instead, the agent becomes the monad. On the way in, it constructs the state the model can read. On the way out, it reconstructs the effect from text that was never typed.

This talk builds the interpreter that fills that role by hand. It reads reasoning bodies, decisions, and confidence, and drives real changes from them. Implemented against Ollama and OpenRouter, it generates self-modifying code: code that produces, breaks, and repairs itself. What survives the translation from stochastic text to a changed world?
💡

#PLnerd