LLMs excel in domains with quick, cheap and deterministic feedback loops. Most high-value human work occurs in domains with slow, costly or uncertain feedback. Models perform poorly in these expensive-loop domains. They struggle to select, critique or switch proxies for hard-to-measure goals. They rarely build or test causal theories to guide experiments. They waste compute and time by ignoring resource constraints. They fail to diagnose failures beyond surface parameters. Humans share many of these cheap-loop habits. Historical disasters often stem from applying cheap-loop thinking to expensive-loop problems.