谷歌DeepMind发布了其下一代AI世界模型Genie 4的研究预览版 3。该模型在物理模拟、人类形象渲染及交互体验的真实性上较前代有显著提升 45。Genie 4具备原生多模态能力，能够同步生成视觉画面以及包括音乐、环境噪音和交互式对话在内的音频 23。此次更新将生成的连贯体验时长从上一代Genie 3的约一分钟延长至两分钟 13。在技术架构方面，Genie 4底层集成了Gemini系列模型以增强推理与引导能力，具体使用Gemini 3.8 Flash作为循环中的引导代理来维护模拟的连贯性 12。此外，模型还利用Gemini Omni锚定现实外观与行为，并正在实验结合Gemini 4 Argon以支持更复杂的多步事件模拟 2。
Genie 4引入了多项关键交互功能，允许用户通过自然语言提示在模拟过程中实时更新环境或引导剧情发展 123。这种实时操控能力使得无需预渲染环境或编程交互元素即可逐帧生成体验成为可能 5。DeepMind将此类世界模型视为通往通用人工智能（AGI）的关键要素，而非单纯的游戏技术工具 5。其应用场景涵盖电影制作者的预可视化、游戏创意原型设计以及机器人训练场景 135。例如，在涉及Gemini Robotics的实验中，研究人员利用Genie 4模拟机器人运行环境 3。值得注意的是，所有由Genie 4生成的视频和音频均嵌入了SynthID数字水印技术 2。
目前，Genie 4处于研究预览阶段，仅向特定的“可信测试者”开放 23。根据计划，该模型将在未来几周内向Google AI Ultra订阅用户以及Gemini API和Google AI Studio的早期访问合作伙伴逐步推出 23。尽管性能有所提升，该模型仍存在已知局限，包括长时间运行后出现的细节漂移、音频准确性不足以及世界内文字可读性问题 3。回顾历史，类似的世界模型技术曾对市场产生波动，去年Genie 3发布时曾导致电子游戏股票下跌 5，而本次Genie 4的发布也引起了相关市场的关注 4。
Google DeepMind has released a research preview of Genie 4, the latest iteration of its AI world model designed to generate more realistic interactive experiences 135. The new system significantly advances simulation fidelity by incorporating native audio generation and enabling real-time modifications through natural language prompts 123. While previous versions could sustain consistent scenarios for about one minute, Genie 4 extends this capability to approximately two minutes 13.
The core improvements in Genie 4 include enhanced realism in physics simulations and human representation, alongside the ability to produce music, ambient noise, and interactive dialogue within generated environments 234. Unlike earlier models, Genie 4 supports live steering, allowing users to alter the course of an experience mid-simulation via text commands 13. To maintain coherence and guide these complex interactions, the model leverages Gemini architecture under the hood 123. Specifically, it utilizes Gemini Omni to anchor realistic appearances and behaviors, while Gemini 3.8 Flash acts as a guiding agent within the simulation loop 2. TechCrunch reports that Gemini functions similarly to a director, planning steps and checking frames to ensure consistency 3. Furthermore, Google is experimenting with integrating Gemini 4 Argon to support more sophisticated multi-step event simulations and robot research applications 2.
Access to the technology is currently restricted but expanding rapidly. At present, the research preview is available only to specific trusted testers 23. Over the coming weeks, availability will extend to subscribers of Google AI Ultra and early access partners of the Gemini API and Google AI Studio 23. Paid users are expected to gain access soon 5. To address authenticity concerns, all video and audio content generated by Genie 4 includes SynthID digital watermarks 2.
Beyond technical specifications, Google positions Genie 4 as a tool for diverse creative and industrial applications, including film pre-visualization, game prototyping, and training environments for robotics 15. Diego Rivas, an engineer at Google DeepMind, emphasized the value of real-time capabilities for creative workflows in film prototyping 3. Tejas Kulkarni, a research director at Google, noted that future developments may involve hybrid world models that combine code generation with real-time simulation 3. However, the company acknowledges current limitations, such as detail drift during long-running sessions, issues with audio accuracy, and challenges in rendering legible text within the simulated worlds 3. DeepMind views these world models not merely as gaming technology, but as critical components in the broader journey toward artificial general intelligence (AGI) 5.
评论
还没有评论，欢迎留下第一条。