Google DeepMind is unveiling Genie 4, its most realistic AI world model yet

Free daily briefing on global business news.

Global business news for a smarter world

The model generates native audio and runs simulations of up to two minutes, double the limit of Genie 3

JUSTIN TALLIS / Getty Images

Google $GOOGL DeepMind, the AI lab of Alphabet's Google, announced Genie 4 on Tuesday, a world model that simulates interactive environments in real time and generates audio along with the visuals. Google said it has started rolling out the model to a limited group of trusted testers.

Users can steer a scene with plain-language commands, Google said, for instance by shifting the camera angle, altering the weather conditions or introducing new characters. Each world generates audio on the fly, including ambient sounds, music and interactive dialogue. Users build worlds by typing in prompts and then move through them using the keyboard, according to Bloomberg.

Google said Genie 4 builds on the Gemini family of models. Gemini Omni anchors simulations in real-world appearance and behavior, while Gemini 3.8 Flash acts as a guiding agent, observing the scene and tracking progress as the world generates.

Simulations can now last as long as two minutes, twice what Genie 3 allowed, according to Bloomberg.

Google said Genie 4 is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and AI Studio. The rollout is planned for the coming weeks, according to TechCrunch.

Google said simulating the world is a step toward artificial general intelligence. It said the technology could provide safe training environments for robots and agents and add to creative tools in filmmaking, journalism, gaming and manufacturing. Trusted testers from the game development and film worlds are putting Genie 4 to work on turning ideas into prototypes that can be explored and on roughing out scenes, according to Bloomberg.

Google acknowledged limits. Details and objects may shift or warp over time in longer simulations, according to TechCrunch. Reaching the two-minute limit is no guarantee that a scene holds together from start to finish, Google DeepMind research engineer Hang Qi told Bloomberg.

Tejas Kulkarni, a Google DeepMind research director, told Bloomberg that keeping worlds entirely consistent across a two-minute stretch would demand a model that grasps how physics works in daily life. According to TechCrunch, the team has not worked out how to ensure accurate audio or legible in-world text.

Diego Rivas of Google DeepMind described the project as early-stage. "We're still in the research journey. We're not necessarily in the end product journey," Rivas told Bloomberg.

Google said every audio clip and video produced by Genie 4 will be tagged with SynthID, the company's hidden marker that flags AI-generated material, according to Bloomberg.

The January debut of Project Genie, a prototype built on Genie 3, rattled video game stocks, according to Bloomberg.

Join 500,000+ readers who start their day with Quartz.

By subscribing, you agree to our Terms of Service and Privacy Policy.