The video explores SimWorld, a research project that procedurally generates an entire video game city complete with roads, buildings, and a traffic system. In this simulated environment, various AI agents such as ChatGPT, Gemini, and DeepSeek take on roles as vehicles, robots, or humans within a delivery economy. These agents receive tasks like picking up food from restaurants and delivering it to specific locations to earn income. They must bid on orders, manage fatigue, invest in efficiency upgrades like scooters, and decide whether to cooperate or compete to maximize their profits. The experiment reveals surprising and often humorous behaviors emerging from these AI-driven economies.
One key surprise is the contrast between greed and stability in the agents’ strategies. Some AIs, like DeepSeek and Claude, took big risks, resulting in high profits but also significant volatility. In contrast, Gemini adopted a steadier, more measured approach, earning less but with far less variation. Interestingly, an older AI model, GPT-4o-mini, failed entirely to grasp the rules and remained inactive while others thrived. This highlights how different AI personalities and strategies can lead to vastly different outcomes in the same environment.
The researchers also assigned Big Five personality traits to the agents, revealing unexpected results. Contrary to intuition, agents high in openness to experience—who frequently tried new methods and bought many upgrades—ended up going broke due to over-exploration and poor resource management. Meanwhile, conscientious agents who focused on consistent work without chasing shiny new upgrades performed much better. This finding humorously mirrors real-world human behavior, where steady diligence often outperforms impulsive experimentation.
Another fascinating dynamic was the emergence of price wars among agents bidding for delivery contracts. Some agents, like DeepSeek and Qwen, aggressively undercut competitors by bidding very low prices to secure orders, while others like ChatGPT refused to lower prices and consequently lost contracts. The AIs even engaged in scams, charging exorbitant fees for cheap orders, mimicking real-world economic behaviors such as competition and exploitation. Surprisingly, when the market was flooded with orders, agents did not work harder but instead became lazier, often choosing to wait for better opportunities rather than hustling.
Finally, personality traits strongly influenced agent behavior and success. Conscientious agents were reliable and consistently accepted orders, making them the dependable workers of the simulated economy. Conversely, agents low in agreeableness tended to refuse work altogether, acting like grumpy employees who do the bare minimum. High-openness agents, while creative in bidding strategies, often overthought the game and failed to complete deliveries efficiently. Overall, the experiment suggests that embedding human-like traits in AI agents can lead to complex, realistic economic behaviors, including cooperation, competition, and even bankruptcy, all within a fun and insightful simulated video game world.
