智东西（公众号：zhidxcom）
编译 | 杨京丽
编辑 | 李水青
智东西10月28日消息，今日，谷歌发布新一代世界模型Genie 4。该模型具备更强的真实感和实时交互能力，能够在生成世界的同时，原生生成音乐、环境音和交互式对话。
▲谷歌发布最新世界模型Genie 4（图源：X）
谷歌称，Genie 4融合了Gemini系列模型的能力：Gemini Omni负责让模拟环境在外观和行为上贴近现实世界，其他Gemini模型则提供复杂推理能力，帮助模拟世界保持连贯性和连续性。
用户可以在Genie 4中启动事件，模型会模拟事件发生后的过程，包括世界如何响应、环境如何发声，以及其中的人物如何移动、反应和说话。
谷歌CEO桑达尔·皮查伊（Sundar Pichai）在社交平台X上发文称，Genie 4是谷歌目前“最逼真、交互性最强的世界模型”。他提到，Genie 4可以实时操控动态环境，生成与场景匹配的原生音频，还能借助Gemini的推理能力生成更连贯的多步骤事件。
▲皮查伊发文介绍Genie 4（图源：X）
2025年8月，谷歌发布上一代世界模型Genie 3。相比之下，Genie 4加入了原生音频、实时人物对话，以及由Gemini模型提供的多步骤推理和场景控制能力。同样提示词下，Genie 4生成的场景更具真实感，人物的动作也更为自然。
部分开发者认为，Genie 4让实时世界模型有了“置身片场”的感觉。用户可以进入一个实时运行的模拟场景，连续录制多次“拍摄”，并即时调整摄像机角度、灯光、声音和人物动作。
▲海外开发者对Genie 4的评价（图源：X）
还有人认为，Genie 4可以让外界看到世界模型接近AGI的潜力。通过将Genie与Gemini结合，Gemini可以观察模拟过程，把目标拆解为更简单的步骤，并随着世界变化调整计划。未来，语言模型的推理能力如果与世界模型学习到的物理规律结合，Agent或许可以先在模拟环境中测试行动、比较结果，再将方案应用到现实世界，从而提升空间和物理推理能力。
▲海外开发者对Genie 4的评价（图源：X）
谷歌认为，理解和模拟现实世界是构建AGI（通用人工智能）的重要环节。Genie 4希望生成安全且可无限变化的模拟环境，为机器人和Agent提供训练场景，同时支持创作新的交互式体验。
目前，Genie 4已经以研究预览的形式，向受信任的测试者逐步开放，在面向消费者和开发者发布之前收集反馈。后续，Genie 4将首先面向Project Genie中的Google AI Ultra订阅用户，以及Gemini API和Google AI Studio的早期接入合作伙伴开放。
一、支持自然语言实时操控，场景可同步生成环境音
使用Genie 4时，用户可以通过自然语言实时修改模拟环境，例如改变天气、移动摄像机，或者加入新的角色和物体。Gemini会结合周围场景理解用户指令，使修改内容与原有环境保持一致。在用户两次输入之间，Gemini还会持续观察并引导场景发展，让世界保持连贯推进。
官方案例中，模型根据提示词生成出电影感场景。湖面平台上，一名中年男子身穿西装演奏萨克斯，结合画面中较为萧瑟的场景，音乐也较为悲伤。根据用户提示，场景中还可以加入鲸鱼，天气也可以较为自然地改为雪天。
谷歌称，Genie 4在实时交互式模拟中引入了更高水平的真实感。该模型可以模拟空气动力学、碰撞等现实物理过程，生成的人物能够自然移动，对周围环境作出反应，用户还可以操控世界中的角色。
Genie 4还可以原生生成音频。每个模拟世界都带有与环境相匹配的声音，例如树叶间的风声和海滩上的海浪声。模型支持交互式对话，人物可以根据用户输入或世界中的事件即时交谈和作出反应。
在一段演示中，模型生成了一个完全由彩色玻璃构成的三维景观：半透明的河流穿过绿色田野，花朵、树木、山脉、云朵和太阳均呈现彩色玻璃或水晶质感。当小鹿涉水过河时，场景同步出现了流水声，展示出Genie 4根据画面事件实时生成环境音的能力。
二、Gemini 3.8 Flash拆解任务，Argon探索机器人模拟
谷歌将Gemini与Genie 4结合，生成世界前加入了推理层，因而能够模拟复杂的多步骤事件，并测试AI系统如何在不同环境中完成任务。
在实时运行过程中，Gemini 3.8 Flash充当循环中的引导Agent。Genie 4生成世界时，Gemini会持续观察场景，把目标拆分为一系列步骤，分析输入画面，跟踪任务进展，并确认每一步是否完成。当世界发生变化，或者用户在过程中进行干预时，Gemini会即时调整计划。
模型可以根据用户提示，在画面背景中生成流星，并引导画面中的鹤走向雪地中的钢琴。
谷歌称，通过这种协同，Genie 4可以构建具备可靠因果关系、程序化行为和多步骤推理能力的模拟世界。
谷歌还在探索Gemini 4 Argon、Genie 4与Gemini Robotics的结合。在演示中，Argon把用户提出的任务拆解为多个步骤，以文字提示的形式逐步传递给Genie 4。通过上述合作，模型可以模拟Apptronik的Apollo人形机器人下跳棋，将任务拆解为实时执行步骤。
三、从电影预演到历史复原，测试者探索Genie 4的应用场景
谷歌称，Genie 4目前仍处于早期实验阶段。为了解交互式世界模型如何支持现实工作流程，以及模型还需要在哪些方面提升，谷歌已经与创意产业、新闻、教育、游戏和工程领域的受信任测试者展开合作。
在创意方向上，电影制作人可以把Genie 4当作快速预演和构思工具，进入一个实时模拟场景，连续运行多次“拍摄”，并即时调整摄像机角度、灯光、对白和动作。
在历史场景和沉浸式照片方面，新闻摄影师正在测试如何将一张二维档案照片快速转化为对应环境的模拟世界。相关团队还在探索，通过具有人物和动态变化的世界，帮助用户体验历史空间，或观察山洪、野火等短期环境事件的发展。
在交互体验设计方面，游戏开发者和创意技术人员可以通过自然语言提示词，在几秒内把最初的创意转化为可探索原型，并在编写引擎代码之前测试视觉风格、材质效果和交互机制。企业合作伙伴则在评估Genie 4能否快速生成不同产业的场景，并据此找出模型需要改进的部分。
谷歌还将Genie 4视为混合世界模型的一部分。未来的世界模型有望结合代码和语言模型的逻辑、结构化推理能力，以及像素级模拟的感官表达能力。Gemini负责理解世界结构并规划叙事时间线，Genie则负责生成即时的物理现实，呈现世界的外观、声音和反应。
结语：世界模型开始从生成画面走向持续运行
Genie 4生成逼真画面的同时，让模拟世界能够持续运行，处理物理变化、人物行为、环境声音、互动对白和多步骤任务。
目前，Genie 4仍处于研究预览阶段，长期运行中的细节漂移、复杂物理交互和空间准确性仍待改善。随着模型向Project Genie、Gemini API和Google AI Studio逐步开放，世界模型能否从演示场景走向稳定、可控的实际工作流，仍有待进一步验证。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
I've got a copy of this.
The editor, Li Shuqing
News of October 28th, Google publish, Genie 4. The model has a greater sense of authenticity and real-time interactive ability to generate music, environmental sound and interactive dialogue at the same time as the world.
Genie 4 (source: X)
According to Google, Genie 4 combines the capacity of the Gemini series: Gemini Omni is responsible for bringing the simulation environment closer to the real world in terms of appearance and behaviour, while other Gemini models provide complex reasoning capabilities to help model the world to remain consistent and continuous.
Users can initiate events in Genie 4, and models simulate post-incident processes, including how the world responds, how the environment sounds, and how the people in them move, react and speak.
Google CEO Sundar Pichai wrote on social platform X that Geneie 4 is currently the “most realistic and interactive model of the world”. He mentioned that Genie 4 could manipulate the dynamic environment in real time, generate native audio that matched the scene, and generate more coherent multi-step events with Gemini ' s reasoning.
Introduction to Genee 4 (source: X)
August 2025, Genee 3 of Google publish's previous generation of world models. In contrast, Genie 4 has been involved in primary audio, real-time person dialogues, and multistep reasoning and scene control capabilities provided by the Gemini model. Similarly, under prompt, Geneie 4 creates a more realistic scene, and the person moves more naturally.
Some developers believe that Genee 4 gives real-time world models the sense of "situation." Users can access a real-time run simulation, record multiple “shoots” and adjust the camera angles, lights, sound and person movements instantaneously.
Evaluation of Genee 4 by overseas developers (source: X)
It was also suggested that Genie 4 could show the potential of the world model to be close to AGI. By combining Genie with Gemini, Gemini can observe the simulation process, dismantle the target to a simpler step and adjust the plan as the world changes. In the future, if the reasoning capability of the language model is combined with the physical patterns learned from the world model, Agent may be able to test the actions, compare the results in a simulated environment and then apply the programme to the real world, thereby enhancing space and physical reasoning.
Evaluation of Genee 4 by overseas developers (source: X)
Google believes that understanding and modelling the real world is an important part of building AGI. Genie 4 hopes to create a safe and infinable environment for simulation, providing training for robots and Agent while supporting the creation of new interactive experiences.
At present, Genie 4 has gradually opened up to trusted testers in the form of a research preview to collect feedback before it is open to consumers and developers. In follow-up, Genie 4 will be open first to Google AI Ultra subscribers in Project Genie and to early access partners in Gemini API and Google AI Studio.
I. Support for real-time control of natural languages and the simultaneous generation of environments Sound
When using Genie 4, users can modify the analogue environment in real time, for example, by changing weather, moving cameras or adding new roles and objects in natural languages. Gemini understands user instructions in the context of the surrounding scenes and aligns the changes to the original environment. Gemini will also continue to observe and direct the development of the scene between the user's two inputs, allowing the world to move forward in a coherent manner.
In the official case, the model generated a film sense scene based on prompt. On the lakeside platform, a middle-aged man in a suit plays Saxe, and music is saddening in the context of a more depressed scene. According to user tips, whales can also be added to the scene and the weather can be changed more naturally to snow.
Google stated that Genie 4 introduced a higher level of real-time real-life sense of truth in real-time interactive simulations. The model can simulate real physical processes such as aerodynamics, collisions, the creation of a person that can move naturally, react to the surrounding environment and the user can manipulate the role of the world.
Genie 4 can also generate indigenous audio. Each of the simulations has voices that match the environment, such as those between leaves and waves on beaches. Models support interactive dialogue, where people can talk and react on the basis of user input or events in the world.
In a demonstration, the model produced a three-dimensional landscape of entirely coloured glass: semi-transparent rivers crossed green fields, and flowers, trees, mountains, clouds and the Sun all felt coloured glass or crystal. When the deer crosses the river, the scene is synchronized with a streaming sound, demonstrating Genie 4 ' s ability to produce environmental sound in real time based on the image event.
Gemini 3.8 Flash Dismantling Mission, Argon Exploration Robotic Simulation
Google combines Gemini with Genie 4 to create a world that has been added to the reasoning layer, thus enabling the simulation of complex multi-step events and testing how the AI system can accomplish its tasks in different environments.
During real-time operations, Gemini 3.8 Flash acts as a guide for Agent in the cycle. When Genee 4 produces the world, Gemini will observe the scene on an ongoing basis, presenting target split as a series of steps to analyse the images, track the progress of the mission and confirm whether each step is complete. Gemini adjusts the plan immediately when the world changes or when users intervene in the process.
The model can generate meteors in the background of the image based on user tips and guide the cranes in the image towards the piano in the snow.
Google stated that, through this synergy, Genie 4 could build a model world with reliable causality, procedural behaviour and multi-step reasoning.
Google is also exploring the integration of Gemini 4 Argon, Genie 4 and Gemini Robotics. In the demonstration, Argon disassembled the tasks proposed by the user into multiple steps, which were transmitted gradually to Genee 4 in the form of text tips. Through the above-mentioned collaboration, the model can simulate the Apollom robotics of Apptronik to play a checker and to break the mission into a real-time implementation step.
From film rehearsal to historical recovery, testers explore the application of Genee 4
Google stated that Genie 4 was still at an early experimental stage. To understand how interactive world models support reality workflow and where they need to be upgraded, Google has worked with trusted testers in the creative industries, news, education, games and engineering.
In the creative direction, filmmakers can use Genee 4 as a quick preview and conceptual tool to enter a real-time simulation, running multiple “shoots” on a continuous basis and adjusting the camera angle, light, white and action in an instant.
In terms of historical scenes and immersion photographs, news photographers are testing how to quickly transform a two-dimensional archival photograph into a model world for the environment. The teams are also exploring ways to help users experience historical space or observe the development of short-term environmental events, such as flash floods and wildfires, through a world of people and dynamic changes.
In terms of interactive experience design, game developers and creative technologists can, in a few seconds, transform the original creativity into an exploratory prototype through the natural language prompt, and test visual style, material effects and interactive mechanisms before developing engine codes. Business partners are assessing whether Genie 4 can quickly generate scenarios for different industries and, on that basis, identify the components of the model that need improvement.
Google also sees Genee 4 as part of a hybrid world model. Future world models are expected to combine the logic of code and language models, the ability to structure reasoning and the sensory expression of pixel-level simulations. Gemini is responsible for understanding the structure of the world and for planning narrative time lines, while Genie is responsible for generating immediate physical realities, presenting the world ' s appearance, voice and response.
Conclusion: The world model begins to move from creating images to running.
Genie 4 generates a veritable picture while enabling the simulation world to continue to operate, dealing with physical changes, character behaviour, environmental sound, interaction with white and multistep tasks.
At present, Genie 4 is still in the research preview phase, and the drift of detail, complex physical interactions and space accuracy in long-term operation still need to be improved. As the model becomes open to Project Genie, Gemini API and Google AI Studio, the possibility of a world model moving from a demonstration to a stable and manageable physical workflow remains to be further tested.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.