Cash reportedly goes unspent in AI-run town simulation

A post describes a paper that put 100 large language model agents in charge of a town economy for 26 simulated weeks, using Pokhara Lakeside geography across 91 runs and 2.44 million decisions. According to the summary, 96.7% of a cash transfer remained unspent 311 steps later. Changing the underlying model changed every measured outcome, while deleting the agents’ memory changed none detectably. The summary also highlights why simulation length mattered: wealth rankings looked frozen at two weeks and only started moving by week 12.