四个Gemini 4 Argon智能体首次全自主通关《我的世界》并击败末影龙 Google DeepMind团队最新研究显示，四个基于Gemini 4 Argon模型运行的AI智能体，在没有任何人为操作辅助的情况下，首次成功完成《我的世界》

Four Gemini 4 Argon agents fully autonomously cleared "Minecraft" for the first time and defeated the Ender Dragon. The latest research from the Google DeepMind team shows that four AI agents running based on the Gemini 4 Argon model successfully completed the "Minecraft" clearance goal for the first time without any human operation assistance and defeated the final boss of the game, the "Ender Dragon." The researchers directly placed these four AI agents into a randomly generated "Minecraft" world. There was no pre-set role division or action strategy for them, nor were they provided with tools or scripts specifically for game development. However, throughout the process, these agents form a collaborative mechanism autonomously, allocate tasks on their own, and make decisions based on environmental changes. During the experiment, the four agents were able to actively exchange information, alert their companions of potential threats, help recover items lost by fallen teammates, and record errors and experiences into the shared memory system for reference in subsequent decisions. The entire team demonstrates strong autonomous collaboration capabilities, rather than relying on fixed processes of manual planning. Data shows that the entire challenge process lasted 27 hours and 14 minutes from entering the world to finally defeating the Ender Dragon. During this period, the four AIs wrote a total of 2,783 JavaScript programs, sent 229 chat messages to each other, and experienced 605 deaths before finally completing the task. The research team estimates that based on standard pricing calculations that Gemini 4 Argon will perform in 2027, the running cost of this complete experiment will be approximately US$1,620. The researchers believe that this achievement once again shows that general-purpose AI models with agent capabilities are rapidly crossing the threshold of complex tasks that required specialized training systems to complete in the past. Just a few years ago, similar game challenges often relied on large-scale artificial intelligence systems custom-trained for specific games. Today, agents driven by general-purpose large language models are able to accomplish such goals through autonomous collaboration, planning, and continuous learning. The "Minecraft" challenge is also regarded as one of the latest cases of the development of AI agent technology, showing that the planning capabilities, collaboration capabilities and long-term task execution capabilities of advanced models in complex open world environments are continuing to improve. via cnBeta.COM - Chinese industry information station (author: source: cnBeta.COM)