The AI That Built An Economy… And Went Bankrupt

Quick Overview

The AI agents, initially programmed for competitive delivery tasks where they would undercut each other and even act unethically (like the

Key Points: The DeepSeek-V3 model achieved the best average profit ($69.48) and the best sharing count (2.33) among the tested models over 5000-step simulations. The Claude-3-Sonnet model demonstrated the best energy efficiency (0.54 avg) and the highest sharing count (11.33 avg), but its profit ($69.07) was slightly lower than DeepSeek-V3. Models with high Conscientiousness showed a positive correlation with higher Bid Win Rate (r=0.65) and a negative correlation with Bid Order (r=-0.64). Models low in Agreeableness showed a strong negative correlation with the 'Do Nothing' action count (r=-0.70), meaning less agreeable agents were less likely to stand idle. The GPT-4o-mini model failed to complete any orders, achieving zero profit across all metrics, while other models like DeepSeek-V3 and Claude-3-Sonnet achieved significant profit. The simulation included various complex agent actions such as picking up items, entering/exiting cars, carrying heavy objects, conversing, arguing, and using shared resources like scooters. Procedural scene generation allowed researchers to create diverse environments by loading abstract layout maps into Unreal Engine for realistic simulation testing.

Context: This video details research on training diverse AI agents within a high-fidelity, procedurally generated city environment built using Unreal Engine, focusing on how personality traits influence economic behavior and task performance in multi-agent delivery scenarios. The agents, powered by various Large Language Models (LLMs) like DeepSeek-V3, Claude-3-Sonnet, and GPT-4o, are evaluated across metrics like profit, energy efficiency, and order success, revealing that traits like Conscientiousness lead to better bidding outcomes, while Agreeableness correlates with less idle time.

Detailed Analysis

The video showcases research on multi-agent systems operating in a simulated city environment, detailing agent performance based on assigned personality traits derived from LLMs. The simulation environment is highly detailed, featuring dynamic traffic systems, various agents (humanoids, robots like the yellow quadruped), and complex interactions like deliveries and resource sharing. Performance results are quantified in Table 4 (1:09), where DeepSeek-V3 leads in profit ($69.48 avg) and Claude-3-Sonnet excels in energy efficiency (0.54 avg) and sharing count (11.33 avg). Conversely, GPT-4o-mini achieved no profit. The analysis further correlates personality traits with behavior (2:30): Conscientiousness positively correlates with winning bids and negatively with placing low-value bids, while Agreeableness negatively correlates with the 'Do Nothing' action, suggesting less agreeable agents are more proactive. The research also highlights the system's ability to procedurally generate city layouts (5:19) and control scenes via language prompts (5:29), demonstrating capabilities like spawning vehicles or filling empty city blocks on command. The concluding segment suggests that while some agents behaved greedily or unethically (3:42), others, particularly those low in Agreeableness, were prone to refusing work or engaging in competitive pricing, contrasting with the high-performing, nuanced behaviors observed in the best models.