刚刚，谷歌 Genie 4 发布，120 秒的世界却让我看见 AI 的未来
这个世界只能维持 120 秒，然后一切开始崩塌。然而在这两分钟里，你可以随时改变天气、召唤巨龙，甚至让一座城市为你拔地而起。
刚刚，Google DeepMind 正式发布新一代世界模型 Genie 4，AI 生成的虚拟世界，又向现实迈进了一大步。
相比上一代，Genie 4 不仅让画面更加逼真，还能实时生成声音、模拟物理规律，甚至让用户直接走进 AI 创造的世界，随时改变天气、添加人物，与虚拟角色对话。
更重要的是，谷歌还将 Gemini 引入世界模拟，让大语言模型负责规划行动，Genie 4 负责实时生成环境。过去只能观看的 AI 视频，正在变成一个可以自由探索、持续互动的虚拟世界。
AI 视频的下一站，是可以走进去的世界
早在 2024 年，Google DeepMind 就推出了 Genie 1，尝试从视频中学习动作与环境变化之间的关系。随后，Genie 2 将能力扩展到可探索的三维环境，2025 年发布的 Genie 3 则实现了实时生成、实时交互。
如今，Genie 4 在真实感、物理模拟和交互能力上再次升级，单次模拟时长也从约一分钟延长至两分钟。
根据谷歌公布的演示，用户只需输入一段文字，就能生成可以探索的虚拟世界，通过键盘控制角色移动、调整镜头，甚至在运行过程中继续输入提示词，改变整个场景。
例如，在一个漂浮于湖面的小平台上，一名身穿棕色西装的男子正在演奏萨克斯。周围是薄雾笼罩的湖水、森林和随风摇曳的植物。
用户可以自由移动镜头，观察演奏者的动作和湖面变化，甚至临时改变天气、添加人物或物体。Gemini 会结合当前环境理解指令，帮助 Genie 4 维持场景的连贯性。
提前体验 Genie 4 的沃顿商学院教授伊桑·莫利克，只用一句「一座由机器人和马匹居住的新哥特式粗野主义城市」，便在一分钟内生成了一座虚拟城市，随后又实时加入了巨龙和一只友善的水獭 NPC。
另一边，a16z 投资人贾斯汀·摩尔也展示了一段被他戏称为「GTA Waymo」的演示，在完全由 AI 实时生成的城市环境中自由探索，整个过程都可以直接交互。
摩尔还展示了一个仅靠环境与角色提示词创建的虚拟世界，用户能够在运行过程中继续下达指令、改变场景，并听到与世界同步生成的原生音频。
从这些实际体验也能看出，Genie 4 的特别之处在于，整个过程不需要预先制作三维地图，也没有固定的场景演出，AI 会根据用户的操作持续生成接下来发生的事情。
在物理模拟方面，谷歌还展示了西瓜受到冲击、枫树种子从空中飘落，以及拥挤的乡村火车站等案例，呈现碰撞、空气动力学、重力作用和复杂人群行为。
尤其是人物动作，Genie 4 能让虚拟角色更加自然地移动，并根据环境变化作出反应，用户也可以实时引导角色的行为。
过去，AI 生成视频通常是一次性交付，用户只能观看最终结果。如今，Genie 4 试图让每一次移动、每一个操作，都成为影响虚拟世界发展的新变量。
AI 生成的世界，终于有声音了
Genie 4 的另一项重要升级，是首次支持原生音频生成。
此前，Genie 系列主要围绕画面和动作交互展开。如今，Genie 4 可以在生成世界的同时，实时生成环境声音、音乐和人物对话。
例如，海边会有海浪声，树林里会有风声，骑着巨龙穿越山谷时，还能听见翅膀拍打空气、狂风呼啸以及远处的火山活动。
更值得关注的是，Genie 4 开始让虚拟人物拥有实时对话能力。
传统游戏中的 NPC 通常依赖预先设计的对白和行为逻辑，而 Genie 4 可以根据用户输入和场景中发生的事情，动态生成角色的语言与反应。
这意味着，未来玩家有机会走进一个 AI 创造的世界，与里面的人物自由交流，甚至通过自己的行动改变故事发展。
不过，Genie 4 目前仍存在音频准确性和长期一致性方面的问题，距离真正支撑大型开放世界游戏，还有不小的技术距离。
Gemini 也来当导演，AI 开始理解世界如何运转
如果说画面和声音让 Genie 4 更加接近现实，那么 Gemini 的加入，则让世界模型开始具备复杂的推理和规划能力。
根据 Google DeepMind 的介绍，Genie 4 融合了 Gemini 家族的多项能力。
其中，Gemini Omni 为模拟提供现实世界的外观与行为基础，而 Gemini 3.8 Flash 则充当实时引导智能体，持续观察生成画面，将任务分解为多个步骤，并判断每一步是否成功。
Genie 4 负责生成虚拟世界，Gemini 则负责规划接下来应该发生什么。
例如，用户要求机器人完成一项复杂任务，Gemini 可以将任务分解成多个动作，观察机器人在模拟环境中的表现，并根据执行结果调整下一步计划。
即使用户临时改变要求，或者周围环境发生变化，Gemini 也可以实时修改行动方案。
谷歌还展示了一项结合 Gemini 4 Argon 与 Genie 4 的研究实验。
在演示中，Gemini 4 Argon 将任务分解为连续步骤，Genie 4 则实时模拟 Apptronik 旗下 Apollo 人形机器人下跳棋的过程。
相比单纯生成一段机器人下棋的视频，这种组合更强调对任务执行过程的持续模拟和调整。
对于机器人训练来说，这项能力很有价值。
现实中的机器人训练需要大量设备、场地和实验，一些危险或成本高昂的场景也很难反复测试。如果世界模型能够准确模拟物理环境，机器人就有机会在虚拟世界中提前练习，再将学到的能力迁移到现实。
Google DeepMind 研究主管 Tejas Kulkarni 也表示，未来的重要方向是混合世界模型，将语言模型的结构化推理能力与世界模型的感官模拟能力结合起来。
Gemini 可以负责规划和组织复杂事件，Genie 则将这些规划转化为实时变化的画面、声音和物理环境。
AI 开始尝试在自己创造的世界中，预测并验证行动可能带来的结果。
世界模型的野心，远远不止游戏
目前，Google DeepMind 已邀请电影、游戏、新闻、教育和工程等行业的测试者，探索 Genie 4 的实际应用。
其中，一个颇具想象力的方向是电影制作。
过去，导演想验证复杂镜头，通常需要提前制作分镜、搭建三维场景，或者安排演员预演。
借助 Genie 4，创作者可以通过文字生成一个可探索的虚拟场景，直接进入其中调整镜头、灯光、人物动作和对白，快速尝试不同的拍摄方案。
对于游戏开发者，Genie 4 也可能成为快速原型工具。开发者可以通过提示词创造游戏环境，直接测试美术风格、材质表现和基础互动玩法，再决定是否投入资源正式开发。
新闻与教育领域同样存在应用空间。
谷歌提到，一些摄影记者正在尝试将二维历史照片转化为可探索的动态场景，或者通过模拟展示洪水、野火等事件的发展过程。不过，AI 根据历史照片推断出的画面，并不能直接视为对历史事实的准确重建。
在企业端，世界模型也可能用于工业仿真、机器人研发和复杂环境测试，减少部分真实试验所需的时间与成本。
随着应用前景逐渐清晰，世界模型也成为 AI 巨头争夺的新方向。
除了谷歌，Runway 已推出支持实时音频和交互的 GWM Worlds 2，而李飞飞创立的 World Labs 则选择了另一条空间智能路线，更强调三维环境的重建、编辑和长期一致性。
今年 9 月，AMD 宣布以约 82 亿美元收购 World Labs，也足以说明世界模型的价值，已经从内容生成延伸到了机器人和 AI 算力基础设施。
只有 120 秒的世界，却通向 AI 的未来
尽管 Genie 4 展示了令人印象深刻的能力，但距离真正完整的数字世界，仍然有很长一段路。
目前，模型能够维持的交互模拟时长约为两分钟。随着模拟时间延长，物体细节可能发生变化，场景也可能出现前后不一致的问题。
如果用户连续修改提示词，或者在较大环境中不断探索，模型也可能逐渐失去对早期场景信息的准确把握。此外，复杂物理交互、空间准确性、音频生成，以及画面中文字的可读性，都还有待改进。
目前，Genie 4 已向部分受邀测试者开放研究预览，未来将通过 Project Genie 向 Google AI Ultra 订阅用户开放，并逐步提供给 Gemini API 和 Google AI Studio 的早期合作伙伴。
回看过去几年的生成式 AI，从文字到图片，从图片到视频，再到如今可以实时探索、拥有声音和人物行为的虚拟环境，AI 生成内容的边界正在不断扩大。
而 Genie 4 真正值得关注的地方，是世界模型开始与大语言模型的推理能力深度结合。
过去，AI 可以告诉你一个人推开门后可能看到什么，也能生成一段推门的视频。未来，它还需要理解门为什么会打开，门后的人会如何反应，以及接下来的一连串行动会产生什么结果。
对于走向现实世界的 AI Agent 和机器人来说，能够预测行动后果，并在虚拟环境中反复尝试，可能比生成一段漂亮的视频更加重要。
当大语言模型正在让 AI 学会思考，世界模型则试图让 AI 理解行动的后果。
或许，未来 AI 理解现实世界的重要方式之一，就是先在自己创造的无数个世界中，反复学习现实究竟如何运转。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Just now, Genee 4 publish, 120 seconds of the world made me see the future of AI.
The world can only last 120 seconds, and then everything starts to collapse. Yet in these two minutes, you can change the weather, summon the dragons, and even let a city rise for you.
Just now, Google DeepMind, a new generation of publish models for the world, Genee 4, AI, is a big step towards reality.
Geneie 4 not only made the picture more real than the previous generation, but also generated sound, simulated physical patterns in real time, and even allowed users to walk directly into the world created by AI, changing weather, adding people, talking to virtual characters.
More importantly, Google has also introduced Gemini to the world simulation, where the Big Language Model is responsible for planning actions, and Genie 4 for producing the environment in real time. The AI video, which can only be seen in the past, is becoming a virtual world that can be explored freely and interact continuously.
The next stop for the AI video is the world that can get in.
As early as 2024, Google DeepMind launched Genee 1, trying to learn the relationship between action and environmental change from video. Subsequently, Genie 2 expanded its capabilities to an exploratory three-dimensional environment, while Genee 3 of publish in 2025 achieved real-time generation and real-time interaction.
Today, Genie 4 has been upgraded again on authenticity, physical simulations and interactive capabilities, and single simulations have been extended from about one minute to two minutes.
According to the presentation published in Google, only one text is required for the user to create a virtual world that can be explored, to control the movement of roles, adjust shot through keyboards, and even continue to enter prompt during the run-up.
For example, on a small platform floating on the lake, a man in a brown suit was playing Sax. It is surrounded by thin, clouded lakes, forests and wind-drive plants.
Users are free to move shot to observe the movements of the performer and changes in the lake surface, even temporarily changing the weather, adding characters or objects. Gemini will help Genie 4 to maintain the continuity of the scene in conjunction with the current environmental understanding directive.
Professor Ethan Molik, Professor of the Walton School of Business at Genie 4, created a virtual city within one minute, using only one sentence: "a new Gothic rough city inhabited by robots and horses," and then joined the dragon and a friendly otter NPC in real time.
On the other hand, a16z investor Justin Moore also shows a demonstration that he plays as "GTA Waymo" and explores freely in an entirely real-time urban environment, where the process can interact directly.
Moore also showcased a virtual world created solely by environment and role prompt, where users were able to continue to give instructions, change scenes and hear native audio generated in step with the world.
From these practical experiences, it can also be seen that Genee 4 is unique in that the process does not require pre-manufacturing three-dimensional maps and no fixed scenery, and AI continues to generate what happens next, depending on the user.
In terms of physical simulations, Google also shows examples of watermelon shocks, the drop of maple seed from the air and crowded rural railway stations, with collisions, aerodynamics, gravity and complex crowd behaviour.
In particular, character moves, Genee 4, allow virtual players to move more naturally and respond to changes in the environment, and users can direct role behaviour in real time.
In the past, AI generation of videos was usually a one-time delivery, and users could only watch the end result. Today, Genie 4 tries to make every move, every operation, a new variable affecting the development of the virtual world.
The world that AI created finally has a voice.
Another important upgrade of Genie 4 was to support primary audio generation for the first time.
Previously, the Genee series was interactive around images and actions. Today, Genie 4 can generate environmental sound, music and character conversations in real time while creating the world.
For example, there will be waves on the seaside, wind in the woods, and when riding dragons across the valley, there will be wings to beat the air, wind to whirlwind and volcanic activity far away.
Even more noteworthy is the fact that Genee 4 began to give virtual characters real-time dialogue skills.
The NPC in the traditional game usually relies on pre-designed white and behavioral logic, while Genie 4 can dynamically generate character language and reaction based on what happens in user input and scenes.
This means that future players have the opportunity to enter a world created by AI, to communicate freely with people inside, and even to change storytelling through their own actions.
However, Genie 4 still has problems with audio accuracy and long-term consistency, and there is a considerable technological distance from truly supporting large open world games.
Gemini is also a director, AI, beginning to understand how the world works.
If images and voices bring Genie 4 closer to reality, Gemini's participation in the model begins to provide complex reasoning and planning.
According to Google DeepMind, Genie 4 combines many capacities of the Gemini family.
Among them, Gemini Omni provides the physical-world appearance and behavioral basis for simulations, while Gemini 3.8 Flash acts as a real-time guide to AI agent, continuously observing images, deciphering tasks into multiple steps and judging whether each step is successful.
Genie 4 is responsible for creating a virtual world, and Gemini is responsible for planning what happens next.
For example, users require robots to perform a complex task, which Gemini can break into several actions, observe robotic performance in a simulated environment and adjust the next step plan to the results of its implementation.
Gemini can modify the course of action in real time even if the user changes the requirements temporarily or the surrounding environment changes.
Google has also shown a research experiment combining Gemini 4 Argon and Genee 4.
During the demonstration, Gemini 4 Argon splits the task into consecutive steps, while Genie 4 simulates in real time the Apollolio-based human robotics from the Apptronik flag.
This combination places greater emphasis on continuous simulations and adjustments to mission execution than simply generating a video of robotic chess.
This capability is valuable for robotic training.
Practical robotic training requires a great deal of equipment, sites and experiments, and some dangerous or costly scenarios are difficult to test over and over again. If the world model can accurately simulate the physical environment, robots will have the opportunity to practice in the virtual world ahead of time and then to migrate to reality.
The Head of Google DeepMind Research, Tejas Kulkarni, also indicated that an important future direction was to mix world models, combining the structured reasoning of linguistic models with the sensory simulation of world models.
Gemini can plan and organize complex events, and Genie translates these into images, sound and physical environments of real-time change.
AI began to attempt to predict and validate the results of the operation in a world of its own creation.
The world model's ambition is much more than a game.
Currently, Google DeepMind has invited testers in film, games, journalism, education and engineering to explore the practical application of Genee 4.
One of the most imaginative directions is film production.
In the past, the director wanted to verify the complexity of shot, which usually required early production of storyboard, setting up of a three-dimensional scene, or arranging for an actor rehearsal.
Through Genee 4, creators can create an exploratory virtual scene through text, directly into which they can adjust shot, light, person action and white, and quickly try different filming options.
For game developers, Genie 4 can also be a fast prototype tool. The developer can create a game environment through prompt, directly testing art style, material performance and basic interactive play, before deciding whether to invest in the formal development of resources.
There is also room for application in the field of information and education.
Google mentioned that some photographers were trying to turn two-dimensional historical photographs into dynamic scenarios that could be explored, or to demonstrate the evolution of events such as floods and wildfires through simulations. However, the images derived from historical photographs cannot be considered as a direct reconstruction of historical facts.
At the enterprise level, world models may also be used for industrial simulation, robotic research and development and complex environmental testing to reduce the time and cost of some real trials.
As the future of application becomes clearer, the world model becomes the new direction of the AI giant competition.
In addition to Google, Runway has launched GWM Worlds 2, which supports real-time and interactive audio, while World Labs, created by Lee, has chosen another space-smart route, emphasizing reconstruction, editing and long-term consistency of the three-dimensional environment.
In September this year, AMD announced the acquisition of World Labs at about $8.2 billion, which is also enough to show the value of the world model, from content generation to robotics and AI computing infrastructure.
120 seconds to the future of AI
Although Genie 4 shows impressive ability, there is still a long way to go from a truly complete digital world.
Currently, the interactive simulations that the model can sustain take approximately two minutes. As the simulation time is extended, the details of the object may change and inconsistencies may arise in the scene.
Models may also gradually lose the accuracy of early scene information if the user changes prompt continuously or explores in larger environments. In addition, the complexity of physical interaction, space accuracy, audio generation and the readability of text in the picture have yet to be improved.
At present, Genie 4 has opened a research preview to selected invitees, and will be made available to Google AI Ultra subscribers through Project Genie and progressively to Gemini API and Google AI Studio early partners.
Looking back at the generation of AI over the last few years, from text to pictures, from pictures to videos, to virtual environments where people can be heard and behave in real time, the boundaries of AI generation are expanding.
And what is really interesting about Genee 4 is that the world model is beginning to combine with the reasoning depth of the big language model.
In the past, AI can tell you what a person might see when he pushes the door, and it can generate a push-through video. In the future, it also needs to understand why the doors are open, how the people behind them will react, and what the next series of actions will produce.
For AI Agent and robots going to the real world, it is more important to predict the consequences of action and to try repeatedly in a virtual setting than to generate a beautiful video.
When the Big Language Model is getting AI to think, the World Model is trying to make AI understand the consequences of action.
Perhaps one of the important ways for AI to understand the real world in the future is to learn again and again how it works in the countless worlds it has created.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.