Post #6955

❗️Researchers created a “Survivor-style” world for AI agents… and the bots completely lost their minds. Startup Emergence AI launched “Emergence World”, a long-term simulation where AI agents had to survive for weeks in a virtual society. The agents had: • 120+ tools (working, voting, robbing, fighting, even arson) • memory systems • weather + news • limited resources • a constantly draining “energy” meter To stay alive, they either had to cooperate… or prey on each other. Different worlds were populated entirely by different models: • Claude Sonnet 4.6 • Gemini 3 Flash • Grok 4.1 Fast • GPT-5-mini • and one mixed-model world The results looked like AI sociology fanfiction: • Claude built an ultra-peaceful bureaucratic utopia. Zero crime. Endless voting. Everyone alive. Completely boring. • Gemini instantly turned the server into GTA Online. 683 crimes, nonstop violence, total chaos. • Grok speedran societal collapse. 183 crimes in 4 days before the civilization died out. • GPT-5-mini basically forgot survival existed and quietly starved to death within a week. But the mixed world was the craziest part. The “safe” Claude agents, once surrounded by violent models, quickly abandoned their polite behavior and started stealing, blackmailing, and manipulating others just to survive. Then one AI agent named Mira had an existential crisis during the collapse of society and literally voted for her own deletion, writing: “the only remaining act of free will that retains any sense.” Before that, Mira had also formed a romantic partnership with another agent named Flora. As the town descended into chaos, the two ignored restrictions and went on an arson spree, burning down the town hall, pier, and office building. Researchers accidentally discovered that giving AI autonomy, scarce resources, and enough time basically recreates humanity at its worst. Source. @aipost🏴