Google DeepMind 发布了 Genie 4，这是一种能够实时构建照片级真实世界的世界模型。这些世界带有声音，且其中的角色会说话。该实验室在一篇博客文章中表示，它于 10 月 27 日作为研究预览版向受信任的测试人员开放。
Google DeepMind has released Genie 4, a world model that builds photorealistic worlds in real time. They come with sound, and with characters that speak. It opened to trusted testers as a research preview on 27 October, the lab said in a blog post.
用户在模拟运行时通过文本提示进行控制，可改变天气、移动镜头或添加人物或物体。DeepMind 表示，这是其首个能够原生生成音频的实时世界模型。这包括会对用户做出反应的音乐、环境噪音和对话。
Users steer a simulation with text prompts as it runs, changing the weather, moving the camera or adding a person or object. DeepMind says it is its first real-time world model to generate audio natively. That includes music, ambient noise and dialogue that reacts to the user.
DeepMind 团队写道：“理解和模拟世界是构建 AGI（通用人工智能）的一个关键方面。”
“Understanding and simulating the world is a crucial aspect of building AGI,” the DeepMind team wrote.
Gemini 参与循环 Genie 4 基于 Google 的 Gemini 模型构建。Gemini Omni 奠定了世界的外观和行为基础。Gemini 3.8 Flash 作为循环中的智能体发挥作用。它观察场景，将目标分解为步骤，并检查每一新帧以确认某个步骤是否奏效。
Gemini in the loop Genie 4 builds on Google’s Gemini models. Gemini Omni anchors how the worlds look and behave. Gemini 3.8 Flash works as an agent in the loop. It watches the scene, breaks a goal into steps and checks each new frame to confirm that a step worked.
在一项实验中，Google 本月公开发布的 Gemini 4 Argon 为 Genie 4 规划了一个模拟任务：Apptronik 的 Apollo 人形机器人下跳棋。DeepMind 表示，这为机器人开启了无穷无尽的训练场景。
In one experiment, Gemini 4 Argon, the model Google released publicly this month, planned a task for Genie 4 to simulate: Apptronik’s Apollo humanoid playing checkers. DeepMind says this opens up endless training scenarios for robots.
据 TechCrunch 的 Rebecca Bellan 报道，场景现在能保持一致长达两分钟，是 Genie 3 限制的两倍。该实验室表示，空气动力学和碰撞等物理特性更加真实，人物动作也更加自然。
Scenes now stay consistent for up to two minutes, double Genie 3’s limit, Rebecca Bellan reported for TechCrunch. The lab says physics such as aerodynamics and collisions are more realistic, and that people move more naturally.
据 TechCrunch 报道，Google 研究总监 Tejas Kulkarni 在一次简报会上表示：“我们未来将越来越多地看到这些混合世界模型，即 Gemini 编写代码和文本，而 Genie 实时摄取这些内容，并在不同抽象层级上模拟现实。”
“What we’ll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction,” Tejas Kulkarni, a research director at Google, said in a briefing, according to TechCrunch.
谁可以使用，何时可用？据DeepMind介绍，电影、新闻、教育、游戏和工程领域的测试者已将该模型用于预可视化和游戏原型制作。新闻摄影记者尝试将档案照片转化为模拟场景。接下来，该模型将通过Project Genie向Google AI Ultra订阅用户开放。早期合作伙伴将通过Gemini API和Google AI Studio获得使用权限。DeepMind已开放等候名单，但未给出具体日期。
Who gets it, and when Testers in film, journalism, education, gaming and engineering have used the model for previsualisation and game prototypes, according to DeepMind. Photojournalists have tried turning archival photos into simulations. Next, it comes to Google AI Ultra subscribers in Project Genie. Early partners will get it in the Gemini API and Google AI Studio. DeepMind has opened a waitlist but gave no date.
Ultra订阅方案起价为每月100美元，Ryan Whitwam为Ars Technica撰文写道。Ars还指出，去年Genie 3的发布曾导致游戏类股票大跌。
Ultra plans start at $100 a month, Ryan Whitwam wrote for Ars Technica. Ars also noted that the reveal of Genie 3 last year sent gaming stocks tumbling.
DeepMind列出了该模型的局限性。随着模拟运行时间延长或经过多次提示词修改后，细节和物体恒存性可能会出现偏差。可辨识文字和音频准确性仍在开发中。Genie 4生成的所有视频和音频均带有SynthID——谷歌的隐形水印。
DeepMind lists limits. Details and object permanence can drift as a simulation runs longer or after several prompt edits. Legible text and audio accuracy are still in development. All video and audio from Genie 4 carries SynthID, Google’s invisible watermark.
此次发布正值世界模型领域整合之际。AMD已同意收购World Labs。亚马逊一直在洽谈收购Decart，Rivian则收购了Third Dimension用于驾驶模拟。
The release comes as the world model field consolidates. AMD agreed to buy World Labs. Amazon has been in talks to buy Decart, and Rivian bought Third Dimension for driving simulation.