LLM use is like MCMC: reward-seeking prompts that never backtrack

rickasaurus · x · 2026-08-25

@TricklerHQ argues that current LLMs and how we use them resemble a reward-seeking stochastic process, like MCMC—each incremental prompt climbs toward a higher energy state without ever backtracking. A reply pushes back: if models were that dumb they couldn't handle long-running tasks, and argues LLMs will always need a harness providing "sense organs" and capabilities; all the nudging and rewriting just reflects that models are still bad.

Original post →

LLM use is like MCMC: reward-seeking prompts that never backtrack

More from Models