LLMs excel at generating plausible artifacts but fail in adversarial, multi-agent scenarios because they lack world models. Unlike domain experts who simulate how counterparties will react, adapt, and exploit patterns, LLMs are trained on static text and optimized for single-shot outputs that sound reasonable in isolation. The
•18m read time• From latent.space
Table of contents
A simple Slack MessageAdversarial Models in real worldPerfect Information Games: When You Don’t Need a Theory of MindWhen the Other Side Has Hidden StateScaling Test Time Compute to Multi-Agent Civilizations: Noam BrownPluribus: Adversarial RobustnessThe LLM Failure Mode: They’re Graded on Artifacts, Not on ReactionsBeing ModeledWhy “More Intelligence” Isn’t the FixThe Expert’s EdgeLanguage Data Hides the Real SkillLLMs dominate chess-like domainsThe Coming CollisionTraining for the next state predictionClosing the LoopSort: