Google DeepMind releases the world model Genie 4, supporting real-time text-based scene changes and native ambient sound

I finally understand why the tech giants are all moving toward physical reality at any cost! 🤯 Google DeepMind has just officially released the world model Genie 4, which supports real-time text-based scene changes, comes with native ambient sound and interactive dialogue, and has Gemini 3.8 Flash serving as the behind-the-scenes director. The threshold for machines to simulate the real physical world has been completely smashed today, and the simulation hypothesis has been confirmed! Setting aside the blogger's emotional foam of memes, if we take this technology apart and look at it, we will find that Google is not competing in ordinary video generation at all this time: it is not generating a cold MP4; it is using pure compute to run an entire physical world, in real time and out of thin air, that you can walk around in and interact with. Previous models like Sora or the various video large models were essentially high-precision dream recorders. They calculated and rendered the pixels in advance and fed them to you; you press pause and it stops, and if you tell it to change the weather, it has to overturn the entire video and recompute it. Genie 4, however, completely cuts off the old logic of "pre-rendering": you are not the audience, you are the player inside it; as you walk forward in the scene, the picture simulates aerodynamics and physical collisions in real time with your footsteps, and when you type a line of text on the side, a rainstorm and new characters are immediately dynamically loaded into the same space. Even more aggressive is the fusion of native audio and video. Past world models were all a bunch of silent pantomimes running in a vacuum, but this time Genie 4, for the first time, makes the world model natively emit ambient sound, background music, and character dialogue that responds dynamically in real time to the player's actions. It is also paired with Gemini 3.8 Flash as a "real-time director," which decomposes causal logic and calibrates picture coherence at the millisecond level in the background, forcibly maintaining a continuous, stable simulation for two minutes. But if you think Google went to all this effort just to make a plug-in for game companies, then you are underestimating DeepMind's ambition. Building a realistic virtual world has never been the end goal; it is an endless training ground prepared for the next generation of AGI. Real-world data is too expensive, too slow, and too costly for trial and error; Tesla has to drive billions of kilometers to train a single autonomous-driving system, and a robot breaking a single mechanical arm costs hundreds of thousands. But when Google can generate interactive worlds that conform to Newtonian mechanics in real time at extremely low cost, hundreds of millions of embodied agents can run, collide, and learn inside them 24 hours a day at ten-thousandfold speed. Of course, as usual, let us keep a cool head: the current limit for a single simulation is still around two minutes, details still drift over time, and this is only a research preview that a few testers can get their hands on. But this represents an extremely dangerous generational watershed: from today on, AI competition has officially leapt from "imitating human text and images" to "simulating the physical laws of the universe."

View original

我终于明白为什么科技巨头都在不惜代价地向物理现实靠拢了！🤯 Google DeepMind 刚刚正式发布世界模型 Genie 4， 支持实时文字改动场景、自带原生环境音与互动对话，Gemini 3.8 Flash 担任后台导演。 机器模拟真实物理世界的门槛，今天被彻底踩碎了，模拟假说被证实了！ 抛开博主玩梗的情绪泡沫，如果我们把这套技术拆开看，就会发现 Google 这次根本不是在卷普通的视频生成： 它不是在生成一段冷冰冰的 MP4，而是在用纯算力，实时凭空跑出一整个可以走动、可以交互的物理世界。 以前的 Sora 或各种视频大模型，本质上是一台高精度的梦境录像机。 它提前把像素算好、渲染出来喂给你看；你按暂停它就停，你喊它改个天气，它必须把整个视频推翻重算一遍。 而 Genie 4 彻底切断了“预渲染”的旧逻辑： 你不是观众，你是里面的玩家；你在场景里往前走，画面随着你的脚步实时推演空气动力学和物理碰撞，你在旁边敲一行字，暴雨和新角色立刻动态加载进同一个空间里。 更狠的是原生音视频的融合。 过去的世界模型全是一帮在真空中运转的无声哑剧，而这次 Genie 4 第一次让世界模型原生吐出了环境音、背景音乐，以及随着玩家操作实时动态响应的角色对话。 旁边还配了 Gemini 3.8 Flash 当“实时导演”，在后台毫秒级拆解因果逻辑、校准画面连贯性，硬生生把时间维持在了两分钟的连续稳定模拟。 但如果你以为 Google 费尽心力搞这个只是为了给游戏公司做外挂，那就太小看 DeepMind 的野心了。 造一个逼真的虚拟世界，从来不是终点，而是给下一代 AGI 准备的无尽练兵场。 现实世界的数据太贵、太慢、试错成本太高，特斯拉要跑几十亿公里才能训好一辆自动驾驶，机器人摔断一根机械臂就是几十万成本。 而当 Google 能够以极低成本实时生成符合牛顿力学的交互世界，数以亿计的具身智能体就可以在里面 24 小时以万倍速狂奔、碰撞、学习。 当然照例保持一份冷静： 目前单次模拟极限依然在两分钟左右，时间一长细节依然会漂移，而且这只是少数测试者能摸到的研究预览。 但这代表了一个极其危险的代际分水岭： 从今天起，AI 的竞争正式从“模仿人类的文字与画面”，彻底跃迁到了“模拟宇宙的物理法则”。