Researchers Built a Tiny Economy. AIs Broke It Immediately

This is SimWorld, a research work that aims to  generate an entire video game city procedurally,   roads, buildings built on the fly, have  a traffic system, everything. You can be   a vehicle, a robot or a human. And then,  put all of these smart AIs like ChatGPT,   Gemini, DeepSeek in there. Then, create a  delivery economy and see what happens. So,   what happens? Well, this is where the surprises  and absolutely hilarious things happen. Okay, so multiple agents receive delivery  tasks: pick up food from a restaurant,   deliver to a location, earn some  income. Agents need to bid on orders,   not get too tired, invest in efficiency  upgrades like scooters and so on. They also   decide whether to cooperate or compete  to maximize their income. Really cool. Six incredible surprises happened that  made me smile. Surprise number one. So, what works here? Greed or stability? Well, of  course, stabi…wait what? What is going on here?    Well, according to this, greed is what works  best. Welcome to the real world! DeepSeek and   Claude were the high rollers. They bet big,  and they win big, and sometimes lose big. They had almost 70 units of profit, but  look at this. Holy mother of papers! Their   chaotic behavior creates huge variance. Gemini, on the other hand, was much more   steady and measured, with about 42 units of  profit, but with much, much less variation. Also, get this, previous generation AI, GPT  4o-mini. How did it do? It got zero. Nothing.    It simply could not comprehend the  rules and just stood there while the   others were busy running their delivery empires. Surprise number two. This is my  favorite. Dear Fellow Scholars,   this is Two Minute Papers with Dr. Károly  Zsolnai-Fehér. Researchers assigned Big Five   personality traits to the agents. People study  these in real humans too. And I was thinking,   you need an AI agent that is high in  openness to experience to succeed,   right? Try new things? Pfft! Wrong. Dead wrong.   Get this, agents with high openness kept exploring   new methods… yes, but too often. They became  shopaholics, kept buying scooters that they   never used and went broke. Absolutely hilarious.   Meanwhile, conscientious agents that kept their   heads down, ignored the shiny upgrades,  and did the work. And they did much better. Surprise number three. Undercutters. As agents had to bid against each other   to get orders, a price war emerged. DeepSeek  and Qwen were notorious undercutters. They   always bid significantly lower  prices to guarantee that they win   that contract. While ChatGPT refused to lower  its prices, and lost the contracts entirely. I find it hilarious that they realized that  if they bid suspiciously low prices, they   could steal all the contracts. Capitalism baby! The scientists also reported that some agents   tried to find a bunch of cheap delivery  orders and charge a fortune for it. Maybe   someone falls for it. Yes, these AIs are  scamming each other like crazy. Insanity. Now hold on to your papers Fellow  Scholars for surprise number four.   When researchers started flooding the  market with orders. Then what happened?    The agents started working harder to get rich!   Kidding, that’s not what happened at all. They   got lazy instead! How? Well, they chose  the “do nothing” action a lot more often.    Instead of hustling, they got lazy and decided  to wait for the perfect opportunity instead. Surprise number five. Your  personality determines your fate!   Three amazing little nuggets here. If  you actually want your orders delivered,   you choose someone who is conscientious.   There is a strong positive correlation   between conscientiousness and picking  up orders. Much like in real life,   these AIs are boring winners. They don’t spam the  chat. They just sign the contract and do the job.   Okay, I expected that. But I didn’t expect  this! AIs that are low on trait agreeableness,   in other words, disagreeable, they just sit  there and refuse to take up any work. These   are the grumpy employees that clock in and just  stare at the wall. Now in AI form. Hilarious.

And now, the dreamers! We mentioned  that AIs high in openness go broke   because they always buy new shiny  stuff to try. But! In reality,   they also try to explore lots of  unconventional bidding strategies.    Essentially, they are too busy overthinking the  meta-game to actually deliver that pizza! Ha! So what is the lesson here? Frankly, I  have no idea, but this was super fun.    Maybe that if you create a world that feels  like the real world and imbue these AIs with   human-like properties, they will behave  a bit more like humans. So now I guess we   are now simulating entire economies where AIs  develop strategies, buy stupid shiny things,   go bankrupt, and undercut each other for  fun and profit. All of this, of course,   in a little simulated video game as a fun  little experiment. What a time to be alive! Make sure to subscribe, hit the bell icon,   and leave a really kind comment to get more  papers like this from the youtube algorithm.