DeepMind 发布 Genie 4 世界模型，改进物理模拟并加入音频和更真实的人物

Last year, Google DeepMind sent gaming stocks tumbling with the reveal of its Genie 3 world model. Since then, world models have become a major focus of AI labs around the world, but Google isn’t resting on its laurels. Researchers have been hard at work on an updated model that improves on Genie 3 in almost every way. And if you pay Google enough money, you might actually get to use Genie 4 before too long.

It’s easy to look at the output of Genie 4 and see a video game, which is why Genie 3 caused so much consternation in the gaming industry. DeepMind doesn’t think about it that way, though. According to Google, world models are a key element in the march toward artificial general intelligence (AGI). A world model creates a virtual world, but it doesn’t require pre-rendering environments or programming interactive elements. You simply tell the model what you want, and it generates the experience frame by frame in real time as you explore it.

DeepMind says this technology could create infinite training environments for physical robots or unlock new avenues of rapid prototyping for creative fields. Of course, that requires the model to accurately simulate what you want. Genie 4 has several improvements that make its simulations more useful, and that’s thanks in part to Google’s updated Gemini models.

When you prompt a world into existence in Genie 4, it leverages Gemini 3.8 Flash as a guiding agent to break down goals and evaluate each output frame. This helps keep the simulation on track, letting you move through varying environments without losing the original context. However, DeepMind notes that Genie 4 worlds still only remain consistent for a couple of minutes before they forget how things started.

Like Genie 3, the new version supports changing the parameters midstream. You can change the weather, add elements, or move the camera, and Gemini will help steer the simulation with each update. This is similar to the way reasoning models generate additional internal prompts to zero in on the right output tokens.