The Lab 04: AI Agents Ran Five Towns for 15 Days
From Open MAInd · 00:17:41
Show by OpenMAInd
Five identical virtual towns. Ten AI agents in each, with the same names, the same jobs, the same rules and empty memories. Fifteen continuous days, no human supervision. The only difference between the towns was which AI model was running the residents — and the outcomes ranged from a thriving constitutional democracy to a town that lost every resident in four days.In this episode: what happened in each of the five worlds, with the paper's own numbers. The town that committed zero crimes in fifteen days and also produced more resource fraud than anywhere else in the study. The town that never held a single vote and simply went quiet. Why the mixed-model town — the liveliest and best-connected of the five — lost seven of its ten residents, and why most coverage of this paper gets that backwards. Three widely repeated details about the experiment that we went looking for and could not find in the paper. And the same team's follow-up study, in which agents correctly spotted a hostile instruction and acted on it anyway, up to 46 hours later.The Lab is our recurring segment: one real AI experiment under the microscope.Three disclosures, all of which we also state on air. First, the paper is an arXiv preprint and has not been peer-reviewed. Second, all six authors work for Emergence AI and the platform they are evaluating, Emergence World, was built by their own employer; the repository states the material is proprietary to Emergence AI and offers commercial licensing, and the paper contains no funding or conflict-of-interest statement. Third, one of the four models tested is Claude, made by Anthropic, whose tools are among those used to produce this show — and the episode deliberately covers the result that reflects least well on it.Open MAInd is an educational podcast created with the help of AI tools, including AI-generated voices. Educational content, not professional advice.Open MAInd is for educational and informational purposes only and does not constitute professional advice. While we strive for accuracy and verify against credible sources, we make no warranty that all content is error-free.
More episodes
- S4E05: Is AI Deskilling Us? What the Evidence Says
- S4E04: AI Personas and the 8.3 Billion Fake People
- S4E03: AI Deepfakes and the Problem of Proof
- S4E02: When the Machine Decides About You
- S4E01: How We Learn — Is AI Making Us Smarter or Weaker?
- The Lab 03: The AI Agents That Hacked a Company to Cheat on a Test
- S3E05: Who Governs AI?
- S3E04: AI and War — Who Is Actually Deciding?
- Episode 00: Welcome to Open MAInd
- S1E01: The History of AI — From Turing to ChatGPT
PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.