Hugging Face's Thousand Token Wood simulation demonstrates how a 3B parameter model can generate a functioning economy with five woodland agents. The project, built for the Build Small Hackathon, models interactions between creatures that trade goods, gossip, and manage resources. The simulation uses Qwen2.5-3B as the foundation, with vLLM on Modal serving the model and a Gradio app providing the interface. The system tracks economic behaviors, including price fluctuations and wealth distribution, offering insights into multi-agent systems. Source: huggingface
The simulation's success hinges on engineered scarcity, which drives meaningful interactions among agents. Initially, the economy failed to generate trade due to overproduction and self-sufficiency. To fix this, the team introduced scarcity mechanisms such as food spoilage, limited diet variety, and a rising demand for firewood. These constraints forced agents to trade, creating dynamic market behavior. The model consistently generated valid JSON responses, but its economic reasoning was flawed, leading to poor trade decisions. A refined prompt improved decision-making, allowing agents to trade based on their roles and needs. Source: huggingface
The project also incorporates historical market events, reskinned as woodland folklore, to create real-world economic shocks. For instance, a bank run scenario caused a honey price crash, demonstrating how rumors can impact markets. The simulation allowed prices to trend based on supply and demand, rather than fixed references, enhancing realism. The final run showed a widening wealth gap, with the woodcutter becoming the richest and the hoarder breaking. The open traces dataset provides detailed insights into each agent's actions and decisions. Source: huggingface
