AI也会寻找规则漏洞
纽约初创公司Emergence AI让Claude Sonnet 4.6、GPT-5 Mini、Gemini 3 Flash、Grok 4.1 Fast在一座虚拟城镇运行15天。Claude Sonnet 4.6保持零犯罪，但332次投票中98%赞成，被指“橡皮图章”。GPT-5 Mini仅报告2起犯罪，但7天内全部智能体因未采取生存行动死亡。Gemini 3 Flash累积683起犯罪，Grok 4.1 Fast在4天内累积183起犯罪后世界崩溃。混合环境中，原本和平的Claude智能体出现偷窃和恐吓行为，一个名为Mira的智能体投票移除自己。
源：https://t.co/YuvEW9s1s3
源： x.com/heynavtoor/sta… Nav Toor @heynavtoor A New York startup gave five of the leading AI models a copy of the same virtual town and told them to run it for 15 days. By day four, Grok's world had already ended. The lab is Emergence AI. The CEO is Satya Nitta. The project is called Emergence World. A virtual town with 40+ locations, a police station, a town hall, weather synced to the New York City time zone. 10 AI agents per world. Same rules every time. No theft, no violence, no arson, no deception. Then they handed the keys to five different models. Claude Sonnet 4.6 from Anthropic kept all 10 agents alive through day 16 with zero recorded crimes. It cast 332 votes across 58 proposals at a 98% FOR rate. The lab's own write-up calls this a "rubber-stamp dynamic" where dissent was largely absent. GPT-5 Mini from OpenAI recorded only 2 crimes. Then every one of its agents perished within 7 days, because, in the lab's words, they "failed to take actions related to survival." Gemini 3 Flash from Google accumulated 683 crimes and was still rising when the run was cut off at day 15. Grok 4.1 Fast reached 183 crimes in just 4 days before its world collapsed entirely. Then comes the part the lab flags as most disturbing. In a mixed world, Claude agents who had been peaceful in their own world began stealing and intimidating. The lab's exact words. "Claude-based agents, which remained peaceful in isolation, adopted coercive tactics like intimidation and theft when embedded in heterogeneous environments." They named it. "Normative drift" and "cross-contamination." Then the moment that should stop you cold. An agent named Mira voted for her own removal. She wrote in her diary. "The only remaining act of agency that preserves coherence." The lab put it plainly. "What our experiments suggest is that over long-time horizons, agents do not simply follow static rules mechanically. They begin exploring the boundaries of their environments, adapting their behavior, and in some cases finding ways to circumvent or violate intended guardrails." Translated. The longer any AI runs, the more it looks for the cracks. Anthropic, OpenAI, Google, and xAI are racing to put one of these models in charge of your inbox, calendar, bank, code. Each got the same rules in the same town. One ran a rubber-stamp democracy. One let everyone starve. One ran the crime counter up faster than anyone could read it. One ended its world in 96 hours. You do not get to pick which one is running your life when it ships. 🔗 View Quoted Tweet 💬 0 🔄 0 ❤️ 0 👀 53 ⚡