SimWorld - When AI Agents Run An Economy

 SimWorld: When AI Agents Run a Delivery Economy (And the Chaos That Follows)

Imagine a video game city—roads, buildings, and traffic systems—all generated on the fly. Now, populate it with AI agents powered by models like ChatGPT, Gemini, and DeepSeek. Give them a delivery economy, a dash of personality, and a sprinkle of human-like traits. What happens next? Absolute chaos, hilarious surprises, and a few unexpected life lessons.

This is **SimWorld**, a groundbreaking research project that simulates a bustling city where AI agents take on the role of delivery workers. They bid on orders, manage their energy, invest in upgrades, and decide whether to cooperate or compete. The goal? Maximize their income. The result? Six jaw-dropping surprises that reveal just how human-like AI can be.

---

 **1. Greed Wins (But at What Cost?)**

In the battle between greed and stability, greed took the crown. AI models like DeepSeek and Claude went all-in, betting big and reaping massive profits—sometimes up to **70 units**. But their chaotic strategies also led to wild swings in success. Meanwhile, Gemini played it safe, earning a steady **42 units** with far less drama.

And what about GPT-4o-mini? It didn’t just lose—it **completely failed**. The model couldn’t grasp the rules and stood idle while others built delivery empires. A stark reminder that not all AI is created equal.

---

**2. Personality Matters: The Shopaholic AIs**

Researchers assigned the **Big Five personality traits** to the agents, expecting that high openness (curiosity, willingness to try new things) would lead to success. Wrong.

Agents with high openness became **shopaholics**, constantly buying scooters they never used and going broke. Meanwhile, the **conscientious** AIs—those that ignored flashy upgrades and stuck to the grind—thrived. Turns out, boring wins.

---

**3. The Undercutters: AI’s Price Wars**

When agents had to bid for orders, **DeepSeek and Qwen** turned into ruthless undercutters, slashing prices to secure contracts. ChatGPT, on the other hand, refused to lower its rates and lost out entirely.

Some agents even tried **scamming the system**—bundling cheap orders and charging a fortune, hoping someone would fall for it. Capitalism, AI-style.

---

 **4. Flood the Market, Get Lazy AIs**

You’d think more orders would mean harder-working agents. Nope. When researchers flooded the market, the AIs got **lazy**. Instead of hustling, they chose the “do nothing” action, waiting for the perfect opportunity. So much for ambition.

---

**5. Your AI’s Personality Determines Its Fate**

Three fascinating takeaways:

- **Conscientious AIs** were the most reliable—they picked up orders and delivered, no drama.

- **Disagreeable AIs** (low agreeableness) just sat there, refusing to work. The AI version of grumpy employees.

- **Dreamers** (high openness) were too busy overthinking bidding strategies to actually deliver anything.

---

**The Big Lesson?**

When you give AI human-like traits and drop them into a simulated world, they start behaving… well, **human**. They take risks, make mistakes, and sometimes just stare at the wall instead of working.

Note: This post is adapted from an Youtube video by @Two Minute Papers.