刚刚 谷歌 Genie 4 发布 120 秒的世界却让我看见 AI 的未来Just now, Google unveiled Genie 4, and the 120-second world showed me the future of AI.


AI free token/AI免费积分

这个世界只能维持 120 秒,然后一切开始崩塌。不过在这两分钟里,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。

Image

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,甚至让使用者直接走进 AI 创造的世界,随时改变天气、添加人物,与虚拟角色对话。

更关键的是,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 则达成了实时生成、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。

依据谷歌公布的演示,使用者只需输入一段文字,就能生成可以探索的虚拟世界,借助键盘控制角色移动、调整镜头,甚至在运行过程中继续输入提示词,改变整个情形。

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。

Image

使用者可以自由移动镜头,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,协助 Genie 4 维持情形的连贯性。

提前感受 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。

Image

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。

Image

Image

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,使用者可以在运行过程中继续下达指令、改变情形,并听到与世界同步生成的原生音频。

Image

从这些实际感受也能看出,Genie 4 的特别之处在于,整个过程不需要预先制作三维地图,也没有固定的情形演出,AI 会依据使用者的操作持续生成接下来发生的事情。

在物理模拟维度,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,展现碰撞、空气动力学、重力作用和复杂人群行为。

Image

尤其是人物动作,Genie 4 能让虚拟角色更加自然地移动,并依据环境变化作出反应,使用者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,使用者只能观看最终结果。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 的另一项关键升级,是首次兼容原生音频生成。

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 开始让虚拟人物具备实时对话能力。

传统游戏中的 NPC 通常依赖预先外观的对白和行为逻辑,而 Genie 4 可以依据使用者输入和情形中发生的事情,动态生成角色的语言与反应。

这意味着,将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,甚至借助自己的行动改变故事发展。

不过,Genie 4 当下仍存在音频准确性和长期一致性维度的难题,距离真正支撑大型开放世界游戏,还有不小的工艺距离。

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。

依据 Google DeepMind 的介绍,Genie 4 融合了 Gemini 家族的多项能力。

其中,Gemini Omni 为模拟给浮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,使用者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,并依据执行结果调整下一步计划。

Image

即使使用者临时改变要求,或者周围环境发生变化,Gemini 也可以实时修改行动思路。

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。

Image

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情形也很难反复测试。如果世界模型可以准确模拟物理环境,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,将来的关键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

当下,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、搭建三维情形,或者安排演员预演。

借助 Genie 4,创作者可以借助文字生成一个可探索的虚拟情形,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思路。

对于游戏开发者,Genie 4 也可能成为快速原型工具。开发者可以借助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。

Image

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情形,或者借助模拟展示洪水、野火等事件的发展过程。不过,AI 依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。

随着应用前景逐步清晰,世界模型也成为 AI 巨头争夺的新方向。

除了谷歌,Runway 已推出兼容实时音频和交互的 GWM Worlds 2,而李飞飞创立的 World Labs 则挑选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。

Image

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 的将来

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。

当下,模型可以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,情形也可能浮现前后不一致的难题。

如果使用者连续修改提示词,或者在较大环境中不断探索,模型也可能逐步失去对早期情形信息的准确把握。另外,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。

当下,Genie 4 已向部分受邀测试者开放研究预览,将来将借助 Project Genie 向 Google AI Ultra 订阅使用者开放,并逐步给出给 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、具备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。

Image

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。将来,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,可以预测行动后果,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关键。

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,将来 AI 理解现实世界的关键方式之一,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。

本文来源:https://www.ifanr.com/1686234
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

The world can only last 120 seconds, then everything starts to collapse. But during these two minutes, you can change the weather at any time, summon dragons, or even have a city rise from the ground for you.

Just now, Google DeepMind officially released the new-generation world model Genie 4, bringing AI-generated virtual worlds another big step closer to reality.

Image

Compared to the previous generation, Genie 4 not only makes visuals more realistic but can also generate sounds in real time, simulate physical laws, and even let users walk directly into the AI-created world, change the weather at any time, add characters, and converse with virtual characters.

More importantly, Google has also brought Gemini into world simulation, letting the large language model handle action planning while Genie 4 handles real-time environment generation. AI video that could previously only be watched is becoming a virtual world that can be freely explored and continuously interacted with.

The next destination of AI video: a world you can walk into

As early as 2024, Google DeepMind launched Genie 1, attempting to learn the relationship between actions and environmental changes from video. Subsequently, Genie 2 expanded capabilities to explorable three-dimensional environments, and Genie 3, released in 2025, achieved real-time generation and real-time interaction.

Now, Genie 4 has been upgraded again in realism, physics simulation, and interactive capabilities, and the duration of a single simulation has also been extended from about one minute to two minutes.

According to the demonstration published by Google, users only need to input a piece of text to generate an explorable virtual world, using the keyboard to control character movement, adjust the camera, and even continue to input prompts during operation to change the entire situation.

For example, on a small platform floating on a lake surface, a man in a brown suit is playing the saxophone. Surrounded by misty lake water, forests, and plants swaying in the wind.

Image

Users can freely move the camera to observe the performer's movements and changes on the lake surface, and even temporarily change the weather, add characters or objects. Gemini will combine the current environment to understand instructions and assist Genie 4 in maintaining the coherence of the situation.

Wharton School professor Ethan Mollick, who experienced Genie 4 ahead of time, used just one sentence—"a new Gothic brutalist city inhabited by robots and horses"—to generate a virtual city within one minute, and then added a dragon and a friendly otter NPC in real time.

Image

On the other hand, a16z investor Justin Moore also demonstrated what he jokingly called a "GTA Waymo" demo, freely exploring a completely AI-generated real-time city environment, with the entire process being directly interactive.

Image

Image

Moore also demonstrated a virtual world created solely through environment and character prompts, where users can continue to issue instructions during operation, change the situation, and hear native audio generated in sync with the world.

Image

From these actual experiences, it can be seen that the specialness of Genie 4 lies in that the entire process does not require a pre-made three-dimensional map, nor is there a fixed scenario performance. The AI continuously generates what happens next based on the user's operations.

In the dimension of physics simulation, Google also showcased cases such as watermelons being impacted, maple seeds drifting from the air, and a crowded rural train station, demonstrating collisions, aerodynamics, gravitational effects, and complex crowd behavior.

Image

Especially for character movements, Genie 4 can make virtual characters move more naturally and react according to environmental changes, and users can also guide character behavior in real time.

In the past, AI-generated video was usually delivered as a one-time output, with users only able to watch the final result. Now, Genie 4 tries to make every movement and every operation a new variable affecting the development of the virtual world.

AI-generated worlds finally have sound

Another key upgrade of Genie 4 is the first-time compatibility with native audio generation.

Previously, the Genie series mainly focused on visuals and motion interaction. Now, Genie 4 can generate environment sounds, music, and character dialogue in real time while generating the world.

For example, there will be the sound of waves by the seaside, wind in the forest, and when riding a dragon through a valley, you can also hear wings flapping in the air, howling winds, and distant volcanic activity.

What is even more noteworthy is that Genie 4 begins to equip virtual characters with real-time dialogue capabilities.

NPCs in traditional games usually rely on pre-scripted dialogue and behavioral logic, but Genie 4 can dynamically generate character language and reactions based on user input and what happens in the situation.

This means that in the future, players will have the chance to step into a world created by AI, converse freely with its characters, and even alter the story’s development through their own actions.

However, Genie 4 still faces challenges in audio accuracy and long-term consistency, and remains a considerable distance from truly supporting large-scale open-world games.

Gemini also steps in as a director, as AI begins to understand how the world works

If visuals and sound bring Genie 4 closer to reality, then the addition of Gemini enables world models to begin gaining complex reasoning and planning capabilities.

According to Google DeepMind’s introduction, Genie 4 incorporates multiple capabilities from the Gemini family.

Among them, Gemini Omni provides the foundation for the appearance and behavior of the simulated real world, while Gemini 3.8 Flash serves as a real-time guidance agent that continuously observes the generated footage, breaks tasks down into multiple steps, and determines whether each step succeeds.

Genie 4 is responsible for generating the virtual world, while Gemini is responsible for planning what should happen next.

For example, if a user asks a robot to complete a complex task, Gemini can decompose the task into multiple actions, observe the robot’s performance in a simulated environment, and adjust the next-step plan based on the execution results.

Image

Even if the user changes requirements on the fly, or the surrounding environment shifts, Gemini can revise its action strategy in real time.

Google also showcased a research experiment that combines Gemini 4 Argon with Genie 4.

In the demo, Gemini 4 Argon breaks the task down into sequential steps, while Genie 4 simulates in real time the process of Apptronik’s Apollo humanoid robot playing checkers.

Compared with simply generating a video of a robot playing chess, this combination places greater emphasis on continuously simulating and adjusting the task execution process.

Image

For robot training, this capability is highly valuable.

In the real world, robot training requires a large amount of equipment, space, and experiments, and some dangerous or costly scenarios are difficult to test repeatedly. If world models can accurately simulate physical environments, robots can practice in advance in virtual worlds and then transfer the skills they learn to reality.

Google DeepMind Research Director Tejas Kulkarni also stated that a key future direction is hybrid world models, combining the structured reasoning capabilities of language models with the sensory simulation capabilities of world models.

Gemini can be responsible for planning and organizing complex events, while Genie turns those plans into real-time evolving visuals, sound, and physical environments.

AI is beginning to try predicting and validating the outcomes that actions may bring in worlds it has created itself.

The ambitions of world models extend far beyond games

At present, Google DeepMind has invited testers from industries such as film, games, news, education, and engineering to explore the practical applications of Genie 4.

Among these, a particularly imaginative direction is filmmaking.

In the past, if a director wanted to validate complex shots, they typically needed to create storyboards in advance, build 3D scenes, or arrange for actors to rehearse.

With Genie 4, creators can generate an explorable virtual scene from text, enter it directly, and adjust camera angles, lighting, character movements, and dialogue, quickly trying out different shooting approaches.

For game developers, Genie 4 may also become a rapid prototyping tool. Developers can use prompts to create game environments, directly test art styles, material rendering, and basic interactive gameplay, and then decide whether to invest resources into full development.

Image

There are also application opportunities in the news and education sectors.

Google mentioned that some photojournalists are exploring ways to turn two-dimensional historical photographs into explorable dynamic scenes, or to use simulations to demonstrate how events such as floods and wildfires unfold. However, images inferred by AI from historical photos cannot be directly regarded as accurate reconstructions of historical facts.

On the enterprise side, world models may also be used for industrial simulation, robotics R&D, and testing in complex environments, reducing the time and cost required for some real-world trials.

As the application outlook becomes clearer, world models have also become a new direction in the competition among AI giants.

Beyond Google, Runway has released GWM Worlds 2 with support for real-time audio and interaction, while World Labs, founded by Fei-Fei Li, has chosen a different path in spatial intelligence, placing greater emphasis on reconstructing, editing, and maintaining long-term consistency in three-dimensional environments.

Image

In September this year, AMD announced it would acquire World Labs for about $8.2 billion, which also underscores that the value of world models has extended beyond content generation into robotics and AI compute infrastructure.

A world of only 120 seconds, yet it leads to AI’s future.

Although Genie 4 demonstrates impressive capabilities, there is still a long way to go toward a truly complete digital world.

Currently, the duration of interactive simulation a model can sustain is about two minutes. As the simulation runs longer, object details may drift, and inconsistencies may arise between earlier and later moments.

If users repeatedly revise prompts or continuously explore in a larger environment, the model may gradually lose a precise grasp of information from earlier situations. In addition, complex physical interactions, spatial accuracy, audio generation, and the readability of text in the imagery still need improvement.

Genie 4 is currently available as a research preview to some invited testers. In the future, it will be opened to Google AI Ultra subscribers via Project Genie, and will gradually be provided to early partners on Gemini API and Google AI Studio.

Looking back at generative AI over the past few years—from text to images, from images to video, and now to virtual environments that can be explored in real time with sound and character behavior—the boundaries of AI-generated content are continually expanding.

Image

What makes Genie 4 truly worth attention is that world models are beginning to combine deeply with the reasoning capabilities of large language models.

In the past, AI could tell you what someone might see after pushing open a door, and it could also generate a video of pushing a door. In the future, it will also need to understand why the door opens, how the person behind it will react, and what results a cascade of subsequent actions will produce.

For AI Agents and robots moving toward the real world, being able to predict the consequences of actions and repeatedly test them in virtual environments may be more critical than generating a beautiful video.

As large language models are helping AI learn to think, world models are attempting to help AI understand the consequences of actions.

Perhaps one of the key ways AI will understand the real world in the future is to first repeatedly learn how reality actually works across countless worlds it creates itself.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry