Google DeepMind 发布 Genie 4 世界模型，实时可交互虚拟环境翻倍至 2 分钟，但游戏股盘中并未重演 1 月抛售

Alphabet Inc.’s Google announced a new version of its artificial intelligence world model that generates more realistic, real-time virtual environments, advancing a technology that has rattled investors in video game companies.

Genie 4, unveiled Tuesday by the company’s Google DeepMind AI lab, lets users create explorable worlds from text prompts and navigate them with keyboard controls. People can also alter a scene as it unfolds by typing instructions, such as changing the weather or adding a character.

Google said simulations can run up to two minutes, roughly doubling what its predecessor, Genie 3, could accomplish. Genie 4 is also the first of the products to generate sound and images together, including ambient noise, music and dialogue that reacts to users.

Google is releasing Genie 4 as a research preview to a small group of trusted testers. In the coming weeks, the company plans to offer it to subscribers who pay for the top-tier Google AI Ultra plan, as well as to some developers who use the app-building tools, Google AI Studio and the Gemini API.

World models are a type of AI that generates interactive environments on the fly, rather than just video. When Google debuted Project Genie, a prototype powered by the Genie 3 model, in January, it spooked Wall Street. Gaming stocks fell sharply as investors weighed whether AI tools could upend how video games are made.

While Google faces intense competition from rivals like OpenAI and Anthropic PBC on the AI frontier, its recently announced Gemini 4 Argon model has been seen as helping it close the gap. World models and multimodality — the ability of AI to work with images, audio and video, in addition to text — have long been considered Google’s technological strengths.

The new model leans on Gemini to keep its worlds coherent. When Genie 4 renders a scene, Gemini watches the incoming frames, interprets the user’s goal and breaks it down into a series of steps, then checks that each one has happened, Google said in a blog post. The company said its team experimented with Argon, but the real-time experience runs on Gemini 3.8 Flash, its lighter-weight workhorse model.

Google said trusted testers, including game developers and filmmakers, are trying Genie 4 for tasks such as turning a concept into an explorable prototype and sketching out scenes. The company is also studying whether it could help train robots, though it’s not yet clear whether any of these use cases will catch on.

Asked in a press briefing about the impact on the video game industry, Diego Rivas, a Google DeepMind product manager, said the company is working with its partners to understand how the technology fits into their work, and where its limits are. “The idea is that we can amplify what things are possible today,” Rivas said.

Google emphasized that the model is still experimental. Visual details and objects can drift as a simulation runs longer, and Hang Qi, a research engineer at Google DeepMind, said reaching the model’s two-minute limit doesn’t necessarily mean a scene stays fully consistent throughout.

Complex physical interactions, legible text and long-term consistency are still works in progress. Making the virtual worlds fully consistent, even for two minutes, would require the model to truly understand the physics of the everyday world, according to Tejas Kulkarni, a research director at Google DeepMind.

“There’s only so much that you can do with two minutes,” Rivas said. “We’re still in the research journey. We’re not necessarily in the end product journey.”

The company said that all audio and video generated by Genie 4 will carry SynthID, Google’s invisible watermark for identifying AI-generated content.